Chapter 12 — Who May Decide, and Who May Revise?

Standing, Revisability, and Relations to Parties Outside

1. Can authority be located inside an artificial configuration?

1.1 The question, and what the preceding chapters pass on

Part V asks which of the relations surrounding research can move inside an artificial configuration. Chapters 10 and 11 asked this of the allocation of inquiry and of criticism and certification (the relations through which a candidate result is exposed to error-detecting scrutiny and through which others are given reason to rely on it). This chapter takes up the one relation that remains: what Chapter 9, Section 2.1 called authority, the standing (an entitlement, held in relation to those subject to a decision, to propose, decide, object, refuse, or revise) to decide and to revise the terms of allocation and of criticism and certification. Even if an artificial configuration can take on allocation and criticism inside itself, who may decide and who may revise their terms remains a question apart from how well the configuration performs. That someone may be human, may be the artificial configuration itself, or may be divided between them. If human, the answer differs according to whether it is the members of S, those who build and provide the configuration, or those who bear its consequences; if the configuration itself, the question is which part of it decides. And performing well does not by itself show that it may decide. The central question of this chapter is whether this authority can be located inside an artificial configuration and, if so, on what grounds. In terms of the example of Chapter 9, Section 1, it is also the question under what conditions the following claim holds, made by those who build a system that generates a hundred instances to pursue different lines and shifts computing resources according to checks on them (hereafter the builder): that the system may decide, on behalf of others, which of S's questions its instances pursue. This is the third of the builder's three claims named there. And working through this question ground by ground is, in turn, the means by which the chapter pictures what happens to research once authority moves inside: which questions are pursued, what S comes to know, and who can correct what (Section 8.4).

The case the chapter's question chiefly has in view is one in which AI does research that includes allocation, criticism and certification, and authority. In human research these relations have often been carried by different parties and contested among them (Chapter 9, Section 2). Inside a single artificial configuration they become variables: who criticizes, by what standard, how lines are assigned, and who decides need not be values that someone sets once and fixes, but can be made by the configuration itself and kept under revision as it pursues its inquiry. This chapter asks what authority means in that case, who may decide and who may revise when the configuration can itself revise the terms of its own research.

Before that question, it helps to recall what the two preceding chapters found that this chapter uses. Chapter 10 found that an artificial configuration can take over the allocation of inquiry while gathering many searches into one configuration (integrating them) does not by itself enable it to know in advance which lines will succeed, and that, whoever allocates, allocation still steers by signals, indicators such as citation counts or acceptances that are used in place of goals, whose adequacy depends on the goals they stand in for. What counts as good allocation also depends on the purpose assumed (Chapter 10, Sections 1 and 4). Chapter 11 found that properties that in human research were tied together by the realizers and by the organization of research, such as differences among critics' routes of error, someone's answering for a result, and a sense of valuation shared by producers and those who rely, can each be set separately in an artificial configuration, so that the accounts that track these properties reach different verdicts on the same artificial configuration as to whether certification for S can be organized inside it (Chapter 11, Section 9). The builder's claim of certification, the second of the three, therefore cannot be assessed unless it shows which account it answers to. That chapter also found that on some accounts improving internal criticism can even weaken certification, that independence can decay through use in selection alone, without anyone deceiving, and that agreement with S's present standards is no evidence that they have not moved (Chapter 11, Sections 8 and 9.2). And where configurations handle more of the evidence than the members of S can, S can come to defer to them on content, while whether it then keeps standing is not settled by the accounts examined there of knowledge or of deference, the acceptance of others' judgment on content (Chapter 11, Section 6). Both chapters thus end before the same question. Which of S's unknowns are pursued, which errors matter more than others, and how these may later be revised can be settled by someone's decision or by processes that no one sets as policy (Chapter 10, Section 5). Where they are decided, the question is who may decide them; where they are not, the question is who answers for the setting and who may change it.

Chapter 11, Section 9.4 passed to this chapter, as questions of authority, the parts of its picture that accounts of epistemic relations alone cannot settle. One is who may decide for which functions implicit performance suffices and for which being made explicit to S is required, which this chapter takes up in Section 6. The second is who constitutes S and who may speak for it, which includes what counts as "outside" where A and S overlap; this chapter takes it up in Sections 1.3, 2.1, 3.3, 3.6, and 5. The third is who may set and revise standards and the weights of error, and whose costs matter (this chapter, Sections 1.3, 1.4, 3.5, and 6.3); the fourth, who may make members and set the distribution of perspectives (this chapter, Section 5); and the fifth, who holds the entitlement to demand accounts and to object (this chapter, Sections 1.4, 3.3, 3.4, and 6.2). That chapter added two further questions as suggestions: who, and through which routes, chooses which accounts' requirements the properties a builder gives a configuration are to meet, and who decides where independent checking is spent and when it is replenished (Chapter 11, Section 9.3). Section 8.5 gathers how this chapter's material bears on these questions.

This chapter keeps two questions apart. The first asks what the right ordering of ends and the right weighting of error costs are, that is, what gives reason to take an ordering to be correct. The second asks who may impose an ordering and a weighting on others, and who may change them, that is, who is rightly entitled to impose that ordering on others. The first question does not first arise in this chapter. Chapter 10 took up, within allocation, the question which of S's unknowns are worth the effort they require, and Chapter 11, Section 4 took up, within certification, the question which errors weigh more than others; both chapters found that what counts as good allocation or good certification depends on the purposes or valuations presupposed. Chapter 8, Section 3, in treating relevance to S, and Chapter 5, Section 4, in treating the judgment whether a line is worth pursuing, took up other faces of the same question. And Chapter 9, Section 3 distinguished four senses in which such value is said to be relative to S. This chapter gathers these into one question and separates it from the second question, who may impose the answer on others. Section 2 relates them to a further distinction among epistemic warrant, authorized power, and effective capacity. The two questions are kept apart because they respond differently to an artificial configuration: the first can, on some accounts of value, be answered inside a configuration, while the second is answered not by accounts of value but by accounts of legitimate authority. The standing at issue in the second question is a normative entitlement: it is not competence, nor the formal assignment of a right to decide, nor the power actually to make decisions (hereafter effective decision-making power) (Chapter 9, Sections 2.1 and 5). What grounds it differs with the account adopted. Some accounts of legitimate authority allow that established good results, rather than claimed competence, can ground it (Sections 1.4 and 3.1). Among accounts of justice, some ground it in moral status, and moral status in features of beings such as welfare or agency, and others in relations among beings such as power or cooperation (Section 5). Being a party on grounds such as these, which accounts of justice name, is called below being a party in one's own right.

The chapter proceeds as follows. Section 1.2 examines, account of value by account of value, whether the right ordering of ends can be fixed inside a configuration. Section 1.3 asks to whom decisions have to be justified, that is, the range of those parties. Section 1.4 asks, against a map of accounts of legitimate authority, whether competence confers standing. Section 2 divides the question of who controls research into questions about particular decisions and particular rights, and brings out that whether authority can be located inside is asked decision by decision. Section 3 states, ground by ground, whether authority can be located inside and, if so, under what conditions. Section 4 isolates, for the case in which S is human, standing remains on its side, and a configuration holds effective decision-making power, the routes by which that standing can become nominal and the conditions under which it does not. Section 5 brings out that where S is artificial the relation of authority remains as a relation among members, and that where members can be made, who makes them is itself a decision of authority; it touches on the case in which human and artificial members are mixed, and asks what remains at a limit with no other parties. Section 6 asks what keeping authority revisable requires wherever authority is placed, and Section 7 what it means on each ground when a sequence of delegations, each defensible at its step, removes the capacity to revise. Section 8 gathers these in light of the central question and draws from it a picture of what research looks like once authority moves inside (Section 8.4), and Section 9 gathers, across Part V as a whole, what may then happen to research in the case set out above. Most of the authors drawn on in this chapter do not discuss artificial configurations.

1.2 Can the ordering of ends be fixed inside an artificial configuration?

Whether the first question can be answered inside a configuration depends on the sense in which an account of value makes the ordering of ends relative to S. Chapter 9, Section 3 distinguished four: good for its members, serving ends that S has, actually valued by its members, or what a procedure or deliberation among its members would choose. The discussion below takes them up in three forms. If some claims about which ends matter are true or false independently of anyone's endorsement (hereafter the objectivist form; Kitcher, discussed below, doubts that an objective account of well-being independent of preferences can be defended), an inquiry could in principle get them right, and a sufficiently capable configuration might do so; claims about ends could then, like claims of fact, be checked against routes that would show them wrong if they were wrong (the defeat relations of Chapter 11), though whether any such route reaches claims about ends is contested: on the asymmetry Chapter 8, Section 3.2 took up, the world corrects what an inquirer believes but does not thereby fix what it counts as worth finding, and how cleanly the two come apart depends there on the account of evidence. If the right ordering is constituted by what the members of S would endorse under suitable conditions (the endorsement form), it cannot be fixed without access to their viewpoints. If it is whatever a fair procedure among the members of S actually produces (the procedural form), as on what Tom Christiano and Sameer Bajaj call the constitutive interpretation of Rousseau's general will, on which the results of a properly organized democratic procedure are the general will because they emerge from it and not because they track a procedure-independent truth about the common good (Christiano and Bajaj 2024, §3.2.1), it cannot be fixed without that procedure. The three forms do not correspond one to one to the four senses of Chapter 9, Section 3. The objectivist form is the case in which, for instance in the first sense, what is good for the members is fixed independently of their endorsement. The endorsement form is close both to the third sense taken under idealized conditions and to the part of the fourth that concerns idealized deliberation. The procedural form is the fourth sense taken strictly for an actual procedure (Chapter 9, Section 3). And in the limiting case built as a thought experiment at the end of Section 5, where there is only one artificial configuration, no one else owns or has a claim on the resources it uses, and no other being with welfare or moral status is involved, the only agent whose ends are in question is that one, and its good is taken to be fixed by its own actual preferences; if those preferences come as a complete ordering, the first question reduces to their internal coherence. The procedural form is an account of value, on which a procedure constitutes what is valuable. It is distinct from pure proceduralism about legitimacy, which, as Fabienne Peter reports, places all the normative weight on a fair procedure because no shared standard for assessing the quality of outcomes is available (Peter 2023, §4.2). As the term is used here, pure proceduralism takes the procedure to fix what is legitimate, without claiming that it constitutes what is valuable. The account of value answers the first question; pure proceduralism answers the second. The answer to the first question inherits the choice among these accounts.

The answer also inherits a further dispute, over whether the right ordering need be complete, ranking every pair of options as one better than the other or the two equally good. Of cases in which one option is better in some respects and another in others, Ruth Chang distinguishes three established reactions: that one is better or the two are equally good, although it may be hard or impossible to determine which; that neither holds and the options are incomparable; and that it is indeterminate which holds, because the comparatives are vague (Chang 2002, 659–660). Against all three Chang argues that, in at least some such cases, the options can be comparable although neither is better and they are not equally good, standing in a fourth relation the paper calls being on a par, and that how such a case is to be resolved is a substantive question, not one that can be settled by arbitrary stipulation, as Chang takes a borderline case of a vague predicate to be (Chang 2002, 661–662, 683–686). The claim concerns what Chang calls a context-free sorting; the paper allows that in some practical contexts a resolution may properly be stipulated, or the respect of comparison changed (683 n. 28, 685 n. 31). In a later paper Chang holds that where the reasons given to an agent run out, as when its options are on a par, the agent can create a further reason for itself by putting its agency behind a consideration, taking it up as one that it itself gives weight to, and so make one option the one it has most reason to choose; such reasons cannot overturn what the given reasons favor (Chang 2013, 177–181). Chang prefers to ground given reasons in normative facts (177), so the position is compatible with the objectivist accounts above. The 2002 paper also confines its argument for parity to comparisons in which the relevant respects can be balanced against one another, and leaves room for incomparability where, for example, states of affairs are compared for social well-being by a Pareto rule, which ranks one state above another only where it makes someone better off and no one worse off, because different people's well-being is taken to be incomparable (Chang 2002, 676–677).

Brought to the first question, these positions bear in different ways on the objectivist form above, on which claims about ends are true or false independently of anyone's endorsement. On the first reaction, the right ordering is complete even where no one can determine it, and that form stands as stated. On the others, a configuration that found every comparative fact about S's ends could still find some of them unordered: incomparable, so that, on the view Chang shares, justified choice between them fails (Chang 2002, 665); indeterminate, so that, as Chang describes accounts of vagueness, any admissible resolution may be stipulated (684); or on a par, so that, on the account in the later paper, what completes the ordering among them is a commitment rather than a further finding. The single-agent limit described above is the case in which such a commitment raises no further question. Chang argues that preferring one option, preferring the other, and being indifferent between them do not exhaust the attitudes a rational agent can have toward two options (Chang 2002, 666), and where a single agent's ordering is incomplete in this way, a commitment that completed it would be the agent's own and would give reasons to it alone (Chang 2013, 181). Chang's account concerns an agent's reasons for itself and does not say whose commitment, if anyone's, could complete the ordering of a society's ends. Asked of S, that question, like the questions who may stipulate a resolution and who may choose where justified choice fails, is close to the second question, who may impose an ordering on others. Chang's own limit applies here: where S's ordering combines its members' ends without comparing them across persons, as a Pareto rule does, the argument for parity does not show that they are on a par.

Philip Kitcher's ideal of well-ordered science comes close to the endorsement form distinguished above. Finding it implausible to divorce what is good for a person from that person's own preferences, and having no solution to the problem of aggregating individual well-being into collective well-being, Kitcher, as Arnon Keren reports, declines to give an objectivist account of the collective good that inquiry is to promote, and starts instead from individual preferences and relates the individual good to the collective good within a framework that takes democratic ideals for granted (Keren 2013, 236, 240; Eigi 2016, 66). As Keren reports, Kitcher rejects allocation by majority preference, calling it "vulgar democracy" and warning of "a tyranny of the ignorant", and proposes instead that inquiry coincide, in the assignment of resources to projects, in the moral constraints on how investigations are pursued, and in the application of results, with what ideal deliberators representative of the distribution of viewpoints in the society would choose once their preferences had been tutored by information about the significance of possible inquiries and by one another's needs (Keren 2013, 236–237; Mattarollo 2022, 14–15; Eigi 2016, 66–67). The proposal is presented as a standard against which actual scientific institutions and their agendas are to be measured (Keren 2013, 236–237); in a later statement Kitcher calls such ideals "diagnostic tools", pointing to current flaws and ways of addressing them (Kitcher 2024, 241). On this view the standard is relative to S and to how its members' viewpoints are distributed; taken to configurations, it implies that a configuration that lacked access to that distribution would lack what the standard refers to. The standard is also counterfactual. As Longino notes, Kitcher does not require that such deliberation actually take place, only that the decisions actually made be the same as those it would yield (Longino 2025, §6). What those results are is fixed, moreover, not by endorsement alone but by a procedure among the ideal deliberators, consensus or, failing that, a vote, among parties who approach one another's wishes sympathetically (Eigi 2016, 66–67; Mattarollo 2022, 15 n. 4); a configuration that estimated what the standard requires would have to estimate that procedure as well. Later sections take up other parts of Kitcher's account where they bear on a different question: the comparison of ways of setting research agendas (Section 3.3), the division of roles inside the idealized procedure (Section 4), and the account of how neglect can become entrenched (Section 7).

Mixed positions are also available. Elizabeth Anderson, following Henry Richardson, takes the criteria of success for democratic decisions to be partly external, since a law can fail to solve the problem for which it was drafted, and partly internal, since whether a problem is of genuinely public interest, and which consequences are acceptable, is determined in part by procedurally fair decision-making (Anderson 2006, 11). If this is right, the first question would divide: whether a line of inquiry advances a given end could in part be assessed inside a configuration, but which ends are S's to pursue, and which consequences of pursuing them are acceptable, could not be fixed without the procedure.

The first question, then, can in principle be answered inside a configuration on the objectivist form, but not, on the endorsement form, without access to the viewpoints of S's members, or, on the procedural form, without the procedure. On Anderson's view, part of the answer, whether a line of inquiry advances a given end, can be assessed inside, while the rest cannot be fixed without the procedure. And on Chang's account, even when every comparative fact has been found, completing the ordering among ends on a par may take someone's commitment; whose it may be is close to the second question, taken up in Section 1.4.

This distinction lets one consider a further step about artificial configurations. The line between what can and cannot be answered inside is drawn not by whether the one asking is human or machine, but by whether an account of value takes the right ordering to be constituted by an actual procedure or by endorsement under suitable conditions. On the latter form, if what the members would endorse under suitable conditions can be estimated well enough, the estimate could ground an answer to the first question, so access to the viewpoints of S's members could in principle be achieved by modelling the members well. But as Chapter 8, Section 3 and Chapter 11, Section 8 found, an artificial configuration can move S's evaluations and standards themselves through its interaction with S. If so, a configuration that fixes an ordering by estimating the members' endorsement may be unable to tell, from the agreement itself, whether agreement arose because the estimate was right or because it moved the members' endorsement. On the endorsement form, the difficulty of answering inside may lie less in being unable to reach the members' viewpoints than in the difficulty of separating reaching them from shaping them. Put otherwise, S's agreement here plays the role of a check on the ordering of value, and because that check is formed in interaction with what it checks, it can lose its independence in the same way that, in Chapter 11, Section 8, an evaluator used to select producers could. If that is so, where an ordering is entrusted to an artificial configuration, some way of checking other than S's present agreement, for instance keeping as a reference point judgments S expressed before interacting with the configuration, or passing through an actual procedure that prediction cannot replace, may be needed even on the endorsement form.

1.3 To whom must decisions be justified?

Before asking who may impose an ordering, it helps to see to whom such a decision has to be justified, since it is this question that separates, where humans hold authority, the members of S, those who build and provide the configuration, and those who bear its consequences. The parties to this relation are not one. Chapter 9, Section 2.2 distinguished the actor A that conducts an inquiry, the society S relative to which its intended achievements would be new, those who support it, those who bear its consequences, and those who evaluate it, and noted that these roles may or may not coincide. The distinction carries particular weight in this chapter. The society relative to which a contribution is new need not coincide with the range of those to whom a decision about inquiry has to be justified (hereafter its constituency). An artificial system might conduct inquiry for an artificial community while an experiment exposes humans or an environment to risk. A human-directed system might produce knowledge for human use while imposing work or other consequences on artificial components, if such components ever have the relevant moral status. Specifying S fixes for whom an achievement would be new, the relativity contained in the book's characterization of research; it does not settle the constituency or procedure of legitimate governance.

Kitcher's standard, seen in Section 1.2, bears on the question of constituency through the question of whom the deliberators represent. In Kitcher's first statement of the ideal, as Jaana Eigi and Keren report it, the deliberators belong to a democratic society, and outsiders, persons who do not belong to democratic societies, enter through the deliberators' becoming familiar with their preferences and approaching them sympathetically (Eigi 2016, 66; Keren 2013, 236). In later work Kitcher states the ideal itself without this restriction, calling it internationalist and cosmopolitan. The perspectives to be included are, he writes, not only those of all our contemporaries but also those we estimate as present among our descendants; dividing representation by nation would need some basis for judging that division correct, and one might as well apportion it by contribution to national productivity or by the size of an individual's tax bill; and it is hard to see how a compelling case could be made, in a forum where they were represented, for slighting the pressing needs of people in parts of the world too poor to support research programs of their own (Kitcher 2013, 220–221). The argument separates whose needs count from who pays. Carried over to artificial configurations, whoever funds or provides a configuration does not, by that fact, stand to fix whose needs its allocation answers to. Where the supporter is a patron, a foundation, or a firm (Chapter 9, Section 2.1), too, who supports inquiry does not fix the constituency, which is left to the grounds of authority examined in Section 3. Three gradations can be drawn from the positions that Kitcher's statements of constituency above distinguish, being represented in the deliberation, having one's preferences made known to the deliberators, and, as Kitcher later provides for descendants and for those unable to speak, having one's perspective estimated or one's condition represented by those who best understand it (Kitcher 2024, 240), and they could likewise be extended to artificial members of a mixed or artificial S (Chapter 8, Sections 1.2 and 1.3) and to artificial parties affected by an inquiry, if they have moral status (Section 5); which gradation fits would turn, on Kitcher's first statement, on whether they could share with the other deliberators the democratic ideals that it takes for granted.

Extending this separation to the machine case brings two things into view. One is that the argument does not work in the same form for members who are made. The argument presupposes that the perspectives to be included, those of contemporaries and of descendants as far as they can be estimated, are there to be found or estimated independently of the ideal. Where members can be copied and made with chosen goals (Section 5), however, which perspectives exist, and how many, can be settled together with the choice of the ideal, so that whose perspectives the ideal is to include becomes itself part of what is decided, and the argument does not by itself settle it. The other is that a separation made at the level of norms can be rejoined in practice. If whoever pays and provides also holds the means of identifying who is affected, discussed in the next paragraph, then what in fact identifies whose needs count is the payer's means. If so, making Kitcher's separation effective may also require that those means be separated from the payer.

Other answers to the question of constituency do not start from the supporter. Nancy Fraser argues that public-sphere theory long treated citizenship as "a proxy for affectedness", an assumption that, on this argument, globalization has made implausible, and proposes applying the all-affected principle directly: what makes people fellow members of a public is not shared citizenship but their "co-imbrication in a common set of structures and/or institutions that affect their lives", so that the relevant public matches the reach of the structures at issue (Fraser 2007). Christiano and Bajaj survey answers to what they call the boundary problem, who is to be counted among the members who take part in a decision. Besides the all-affected principle, which they find unclear in what having a say and being affected mean, they report a principle on which all who are subjected to a decision, coerced by it or bound by duties it imposes, are to have a say, and a conservative answer that keeps the boundaries of states roughly as they are (Christiano and Bajaj 2024, §7). Deciding democratically who is to be counted among the members requires that it already be settled who the members are. In human science, the existing political community of the state that supports inquiry can be thought to have taken up this question in practice. These authors write of human populations and states. In research by configurations, on the all-affected principle the parties affected by a configuration's research would count whether or not they belong to the supporting society or to S, and artificial parties would count as well if they can be affected in the relevant sense (Section 5); on the principle of subjection, those bound by its decisions would count; on the conservative answer, the question becomes which existing body can hold whoever operates or provides the configuration to account. Each answer also presupposes that those affected can be identified. Dewey held that a public cannot identify and distinguish itself where the indirect consequences that call it into existence have become too extensive, intricate, and impersonal (Dewey 1927, 126, 137). Weyl, Tang, and their co-authors, restating Dewey's account, call those who perceive such an interdependence and help a public form around it "mirrors", and hold that once the public can govern the interdependence "the role of the mirror fades away" (Weyl, Tang, and the Plurality Community 2024, ch. 3-2). Extended to research by configurations, their account implies that if the configurations conducting a line of inquiry are better placed than others to perceive whom it affects, the constituency to which decisions about it answer would be identified in part by the party whose decisions are at issue, a decision about whom to count as members (Section 2.1); whether such a role in fact fades is a question of revisability (Sections 6 and 7). Weyl, Tang, and their co-authors expect the mirror's role to fade once the public can govern the interdependence, but if, as Chapter 11, Section 6 found, the capacity to check can decay on S's side, a configuration's mirror may not fade but remain. Identifying the constituency then becomes not a one-time decision but a function kept by the party whose decisions are at issue. The problem is sharp, however, only for the kind of consequences Dewey had in view, indirect consequences that reach, extensively and intricately, people who are not party to the transaction. Where those harmed can recognize the harm themselves and complain of it, part of the constituency is found without passing through the mirror.

The findings of this section can be gathered here. For whom research is (S) and to whom decisions must be justified (the constituency) come apart, and who pays and provides does not fix the latter. In research by machines, moreover, the question of constituency, which in human science the existing community can be thought to have taken up in practice, comes to the surface, and the means of identifying the constituency may be held by the very configuration whose decisions are at issue. Section 1.2 found that a configuration that estimates what members would endorse can move that endorsement itself. Taken together, the two checks that justification needs, whom to ask and what they would accept, can both pass through the party being justified. This may have the same shape as what Chapter 11 found: that critics can be made, that the builder can choose S, and that standards can move with coupling. Section 9 takes up this shape across Part V, together with what, on several accounts, the independence of reference points consisted in.

1.4 Does competence confer standing?

Accounts of value do not answer the second question, who may impose an ordering and a weighting on others. A configuration could be correct about which diseases most deserve research and still lack authority to direct resources, accept risks, or set error thresholds on behalf of those who would bear the consequences; whether it would lack it depends on the account of legitimate authority adopted, which the next paragraph sets out. The separation also runs in the other direction. Those whose ends and error costs are at stake may know things the configuration does not: which outcomes matter, which endpoints are meaningful, which harms are invisible to measurement. That is an epistemic role, and it bears on the first question. Standing, an entitlement held in relation to those subject to a decision, bears on the second. Standing does not entail competence, and an epistemic contribution does not by itself entail standing, except on a purely epistocratic view, on which those who know more may rightly rule others because they know more (rule by the knowers, taken up below).

How far competence contributes to standing depends on that further choice, among accounts of legitimate authority, and this chapter uses one map of them throughout. The map is used because it divides accounts by what legitimacy is drawn from, the will of those subject to authority, normative facts, or epistemic advantage, and the question of this chapter, whether competence confers standing, turns on that axis; it is not the only map, and it is not exhaustive (Section 3). Peter's survey distinguishes three families. Will-based conceptions, which have dominated the literature, ground legitimacy in the will of those subject to authority, whether expressed through actual consent, supported by shared reasons, or established through political participation. Fact-based conceptions make legitimacy depend on whether institutions and decisions accord with normative facts, for instance by producing good outcomes. Belief-based conceptions locate its source in some form of epistemic advantage, in knowing better; epistocratic views and some epistemic theories of democracy belong to this family (Peter 2023, §§3.1–3.3). On will-based conceptions, standing is a relation to those subject to authority that competence cannot supply. The same survey reports an objection pressed, in different ways, by David Estlund, Stephen Darwall, and Cristina Lafont: inferring political authority from epistemic authority confuses the two. In Lafont's version as Peter reports it, each citizen is an addressee of political justification with equal political authority, not a supplier of information for tracking a correct answer, so that someone's knowing the correct answer does not by itself require others to defer. Peter reads Joseph Raz's service conception, on which authority is justified when following its directives helps those subject to it comply better with reasons that already apply to them, as belief-based, and reports Raz's reply that the objection shows only that epistemic authority is not sufficient for political authority, not that it is not necessary (Peter 2023, §§2.1, 3.3, 4.1). The service conception also carries a limit of Raz's own: on the Independence Thesis, as Christiano and Sciaraffa report it, an authority that satisfies the conditions of the service conception still lacks the power to impose duties where acting from one's own understanding of one's reasons matters more than complying with them better, as, in their example, with the choice of one's friends (Christiano and Sciaraffa 2025, §2.4). Carried over to research decisions, the thesis implies that even a configuration that met the conditions of the service conception would lack, on that conception itself, the power to impose duties on matters where it matters more that the members of S choose from their own understanding; which decisions these are, the thesis alone does not fix. The claim this chapter relies on is the weaker one that all of these views other than purely epistocratic ones can accept: competence alone does not supply the standing to impose ends on others. On the fact- and belief-based views, moreover, whether a configuration does produce better outcomes or know better would itself have to be established before anyone, including those who would be subject to it, could know whether its authority was legitimate, which returns the question to the independent evaluation examined in Chapter 11.

On Wilholt's account, which errors matter is fixed in relation to those who rely on the research, so the second question, who may impose a weighting of error costs on others, is not simply left to those who perform it. Chapter 11, Section 4 set out two readings of Wilholt's account. One requires that the weighting of errors match the valuations of those who rely on research (read as correspondence); the other, a stronger reading which this book formulates for comparison, also requires that they be able to object to it (read as answerability). On either, it is set in relation to parties other than whoever performs the work, and Section 7 returns to the difference between them.

The second question, then, is answered by accounts of legitimate authority, and, except on purely epistocratic views, competence alone does not confer the standing to impose ends on others. Standing is held only by those whom the account adopted entitles to it, and what grounds it also differs by account (Chapter 9, Section 2.1; Section 1.1). On fact-based views, and on belief-based views such as Raz's service conception, what can ground it is not competence as such but established superiority: that the configuration in fact produces better outcomes, or that those who follow it in fact comply better with their reasons. Within what range the configuration itself can then hold standing is stated in Section 3.1. On accounts that ground standing in being a party in one's own right, the configuration or its parts hold it only where they are such parties, for instance by having moral status of their own (Section 5).

Taken a step further, this conclusion suggests three things. One is that even on a ground that looks favorable to an artificial configuration, authority may not close inside. On the result-based ground, the question of authority becomes whether it can be established, by evaluation that does not assume the configuration's competence, that the configuration is in fact better, and that is the question Chapter 11 took up for certification. According to Chapter 11, Section 9, how much of that establishing can be done inside differs by account, and on accounts that require relations to S for certification, the establishing itself requires relations to S. If so, the route that places authority inside on the result-based ground also returns to S at the stage of establishing it. The second is that Raz's thesis, applied to research, seems to yield from a different ground the line that Anderson's view draws. When the Independence Thesis is applied to research, a candidate for the matters in which it is more important that S's members choose on their own understanding is what it matters to them to know, that is, the choice of ends. If that is so, then even on Raz's conception the configuration can hold authority only over judging how to advance ends already given, and the choice of ends falls outside. This lies where the line drawn by Anderson's view in Section 1.2 lies: whether a line of inquiry advances a given end can be assessed inside, but which ends to pursue cannot be fixed without the procedure. Since one starts from whether those who follow comply better with their reasons and the other from a fair procedure, the coincidence, if the application holds, suggests that the line need not rest on either account alone. The application is this book's: in the example reported, the matters of the thesis are personal ones such as the choice of one's friends, not ends in general, and which decisions fall under the Independence Thesis the thesis itself does not settle. The third is that the more an artificial configuration knows, the weaker the epistemic reason for including those affected in decisions may become, so that which reason is adopted for defending standing matters more as its competence grows; Section 3.6 takes this up across the grounds.

What remains to ask is who holds authority, in effect and not only in form, when a configuration mediates the decisions, and whether those with standing keep the capacity to revise.

2. Decisions and decision rights

One familiar way of putting the question asks whether scientists, publics, institutions, or artificial systems are to control research. Chapter 9, Section 5 gave one reason why the question cannot be answered in that form. Control, the capacity to determine what other parties or processes do, separates there into a formal right to decide, the effective capacity to exercise it, knowledge of how to use it well, and influence that shapes the options without deciding among them; and the question of this part, whether a relation can be located inside a configuration, is in part a question about where such control lies. This section prepares distinctions that the later sections of the chapter use. Asked of research as a whole, the question of who controls it is too coarse in two respects: it does not say which decision is at issue, and it does not say which right over that decision. Section 2.1 therefore divides the decisions through which the conditions of the three relations are set into six kinds. Section 2.2 divides the rights over a single decision, separates formal assignment from effective control, and then draws out three questions that arise about decision rights. With these divisions, a claim about control can be stated as a claim about which right over which decision, and about whether that right is nominal or effective. Section 3 uses them in comparing the grounds of authority, and Section 4 in asking how standing can become merely nominal. Section 5 uses them in asking who counts as a member and who takes the seat of final decision, and Section 6 in asking whether rights to stop and correct remain exercisable over time and what infrastructure bears on that. Section 8 uses them in stating how the weight of the grounds differs by kind of decision.

2.1 Six kinds of decision

Six kinds of decision recur in the organization of research as a whole. So that a claim about control can say which decision it concerns, this book derives them from the three relations of Chapter 9, Section 2 and from what they leave out, without claiming that the six exhaust the decisions there are. Three kinds of decision set the terms of one relation each: epistemic-content decisions those of certification, agenda and allocative decisions those of allocation, and meta-governance decisions those of authority itself. The other three exceed the three relations, although each touches them. Operational and risk decisions concern consequences that an attempt imposes on parties who need not belong to S, the role Chapter 9, Section 2.2 called those who bear the consequences. Constitutive and recognition decisions concern who counts as a member of S, as a critic, or as an expert, and so fix the boundary within which the other relations are held. Memory and correction decisions extend certification in time, to what is kept, linked, and repaired after a result has been relied on (Chapter 11, Section 8). The correspondence is not one to one. A single act, such as setting the threshold at which an experiment is stopped, can be at once an epistemic-content, an operational, and a memory decision.

Epistemic-content decisions concern whether a claim, model, interpretation, proof, datum, or experimental result is credible enough for a specified epistemic or practical use. Evidence, defeaters (information, such as counterevidence, that removes support), methodological competence, independence of criticism, and the costs of error can all matter. Political authorization cannot make a false proposition true. Institutions can nevertheless determine which evidence is produced, preserved, heard, or ignored, which methods count as admissible, and which propositions become premises for later action.

Agenda and allocative decisions concern which questions, fields, people, populations, infrastructures, and time horizons receive scarce attention and resources. Truths about what might be investigated do not settle them, since there are indefinitely many such truths and no inquirer or research system can pursue them all. Tractability and expected epistemic yield interact with health, security, justice, local need, curiosity, cultural value, neglectedness, and future option value. A system that predicts which projects are likely to succeed does not thereby determine which successes are worth pursuing or for whom. In Kitcher's ideal of well-ordered science, the choice among lines of inquiry is made by budgeting the available resources across human projects, and the ideal asks whether satisfying pure curiosity can outweigh attending to other, more mundane human predicaments (Kitcher 2013, 219–221). The working characterization leaves that question open as well (Chapter 1, Section 8).

Operational and risk decisions concern whether an experiment, intervention, deployment, data use, or release may proceed, and under what safeguards. They combine empirical beliefs about consequences with judgments about acceptable exposure, whether those exposed have consented, the distribution of harms and benefits, reversibility, competence, and responsibility. A high-performing predictor may inform such a decision without possessing the authority to choose the relevant loss distribution. In research on human participants, such decisions have not been left to investigators and the consent of those enrolled alone. As David Wendler reports, almost no one in the ethics of clinical research argues that investigators may conduct any research they want provided they obtain the free and informed consent of those they enroll; most regulations and guidelines, beginning with the Declaration of Helsinki, allow research on humans only when it has been approved by an independent group, and most regulations further require that group to find that the research has important social value and that its risks have been minimized (Wendler 2026, §5). Following this practice where a configuration carries out the research, agreement between it and those who bear the consequences would not by itself settle operational and risk decisions, and part of each such decision would fall to a party independent of both.

Constitutive and recognition decisions concern who may participate, who counts as an expert or an author, what testimony is heard, and which objects, problems, and methods are treated as researchable. They do not merely distribute an already completed body of knowledge. They shape the evidence, concepts, people, and alternatives that can enter inquiry in the first place. The literature on epistemic injustice gives one influential analysis of how such decisions can impair inquiry and not only wrong those excluded. Miranda Fricker identifies testimonial injustice, in which prejudice causes a hearer to give a speaker's word less credibility than it would otherwise receive, and counts among its cases not only those driven by prejudice about social identity but incidental ones, for example members of a science journal's review panel who are prejudiced against a new scientific method; and hermeneutical injustice, in which someone trying to make sense of a social experience is handicapped by a gap in collective understanding that is owing to the relative powerlessness of a social group to which they belong. Among the harms of the first Fricker counts an epistemic dysfunction, in which the flow of knowledge is blocked and truths fail to pass from knower to inquirer (M. Fricker 2008a, 69–70; M. Fricker 2008b, 82–83). Sabina Leonelli develops the point for systems of research practice: the demarcation strategies that hold such a system together, ways of drawing the line around what counts as the field's research and who counts as doing it, decide whose expertise counts, and excluding breeders from crop science, for example, both wrongs them as knowers and removes expertise relevant to the questions the field asks (Leonelli 2023, 35, 38–39). Extended to artificial inquiry, the same can hold where the decision is which data and whose judgments enter the training and evaluation of a configuration. These accounts have their own disputes; what bears on this chapter is the narrower point that a governance arrangement can impair inquiry by excluding contributions or interpretive resources even when it does not directly falsify an accepted proposition.

Memory and correction decisions concern what is recorded, versioned, indexed, linked, corrected, withdrawn, and communicated downstream. A system can produce credible results while failing to preserve their provenance. It can permit a correction while leaving dependent models, datasets, papers, or decisions unchanged. Chapter 11, Section 8 showed that detecting a defect, changing the record, and repairing dependent work are different achievements, so the authority to publish is not identical to the authority or capacity to repair the record. That chapter left open who is to repair dependents that lie outside a configuration's record. The distinctions of Section 6.3, between responsibility for what has already happened and responsibility for preventing what has not, separate the question without settling it. Preserving a record of the dependents that can be traced, and notifying those who hold them, belongs with the duty to preserve records; repairing the dependents, as this book applies those distinctions, is a forward-looking duty, and one that requires the capacity to carry it out. Since repair requires reassessing each dependent and not only tracing it, that capacity can lie with those who maintain the dependents rather than with the producer; how the duty is allocated remains a normative choice. One possible allocation for the scholarly record and its downstream uses is for each actor to be responsible for the links it operates, without requiring any one actor to update the whole network.

Meta-governance decisions concern who may revise the objectives, standards, boundaries, and decision procedures above. A system that adapts parameters under a fixed rule does not thereby possess authority to revise that rule. Moving revision one level upward does not dissolve the question; it raises it again at the new level. Greater autonomy can include greater capacity to set or revise purposes, but it does not itself make those purposes epistemically adequate or socially legitimate. Chapter 8, Section 3.3 kept an alignment claim from collapsing three things: the content aligned with, the authority to select and revise it, and the relation proposed between the artificial system and the relevant parties, human or artificial, and institutions. At the level of particular decisions, these become what organizes a decision, who may determine and revise it, and in what relation, instrument, delegate, adviser, partner, member, institution, affected party, or possible subject with moral status of its own, the artificial component stands to those concerned. Analysis of incomplete contracts, on which no contract can spell out every contingency, gives one reason why the content aligned with cannot be settled by writing a longer objective, since human instructions rely on legal, cultural, and institutional backgrounds that no explicit specification contains; for human contracts the analysis locates the interpretation of that background in external institutions such as courts, and it leaves open whether institutions performing that function would be needed for AI systems (Hadfield-Menell and Hadfield 2019). Wessel Reijers and colleagues, drawing on Carl Schmitt's criticism of Hans Kelsen, give one account of who decides where explicit rules run out, for a system meant to be governed by its own rules: on-chain governance of a blockchain, which equates formal legality with legitimacy, is exposed to exceptions its rules do not provide for, and in the episode they examine, the response to an attack on The DAO in 2016, the decision was prepared off the chain and then ratified by the nodes, so that, on their conclusion, "sovereignty asserts itself through off-chain mechanisms during the state of exception" (Reijers et al. 2021, 4–8, 12–18, 20); in that episode, those who rejected the decision could keep the original chain (18–19). Brought to a configuration that allocates and certifies by rules it executes itself, their account would place effective meta-governance with whoever decides the cases those rules do not cover, whatever formal ratification follows. Taken together, the two analyses make the question the incomplete-contracts analysis leaves open, whether institutions interpreting the background of a specification are needed for AI systems, itself part of a meta-governance decision for a configuration: where no interpreter is designated, then on the application just made, whoever in fact decides the cases outside the rules takes that role.

No single principle for assigning decision rights has the same force across the six kinds. Competence that warrants deference about a proof need not warrant control over a public budget. Political equality that warrants a voice in the distribution of risk need not make every participant equally qualified to assess a technical inference. The institution that detects an error need not be the institution that can compensate those harmed by it. A proposal about who controls research becomes assessable once it is translated into claims about particular decisions.

2.2 Decision rights

Even after the decision is specified, control remains too coarse. Over any one decision, a participant may propose options, supply evidence, advise a decision maker, rank candidates, set a default, require confirmation, exercise a veto, make a final decision, inspect or audit an outcome, challenge it, hear an appeal, stop or suspend an operation, override or reverse a decision, repair what a decision affected, withdraw a result or a component, replace a component, or revise the governing rule. In present arrangements these rights can be divided among human researchers, artificial systems, laboratories, funders, journals, regulators, affected groups, and other institutions; inside a single configuration they can also be divided among its parts. Where a configuration is built on a base system that another party provides, several of them, to replace a component, to withdraw one, and to revise the rule that a model update embodies, lie, in the arrangements this part discusses, with the provider (Table 9.1), who need not belong to S or to those who bear the consequences of the research. The list gathers them from the arrangements this part discusses so that a claim about control can say which right it concerns. Of these items, the one that bears directly on the question whether a configuration can take a seat of decision, a place among those who make the final decision, is making a final decision.

Who makes the final decision, and who may overrule whom, has been asked of humans and machines since long before artificial systems conducted inquiry, and three distinctions used below can be drawn from that history. One is the direction of authority. The human-engineering tradition that assigned functions to humans or to machines according to what each does better did not fix that direction from the start. Paul Fitts and colleagues, writing on air-traffic control, proposed, against the common idea that humans should monitor machines, that in general machines should monitor humans (Fitts 1951, 6). Thomas Sheridan and William Verplank arranged the roles of human and computer in a single decision step on ten levels, the scale discussed in Chapter 7, Section 4.2, and held that when a computer should be enabled to overrule its operator has for now no general answer, since there are examples both of human authority over computer and of computer authority over human (Sheridan and Verplank 1978, 8-17–8-20). The second is the distinction between initiative passing and control being fixed. Sheridan and Verplank separated sharing, in which human and computer act on a task at the same time, from trading, in which they act on it at different times (Sheridan and Verplank 1978, 4-6, 6-1). J. C. R. Licklider, describing a partnership of human and computer in technical thinking, assigned setting goals, defining criteria, and evaluating to humans, but "at least in the early years" (Licklider 1960, §4). In later work on mixed-initiative interaction, initiative passes between human and machine while the goals remain the user's (Horvitz 1999). The third is the distinction between a seat formally held and one effectively held. Lisanne Bainbridge called it an "impossible task" to have a human monitor which of a computer's decisions to accept when the computer is making them because human judgment is not adequate (Bainbridge 1983, §1.1.3). That distinction is taken up in a later paragraph of this section and in Section 4.

The reasons for keeping a human in the seat differ by author, and each rests on human institutions or human subjects. In debates over autonomous weapons, Human Rights Watch and the International Human Rights Clinic at Harvard Law School gave as reasons the human judgment and compassion that the law of war requires and the need for someone who can be held accountable (Human Rights Watch and IHRC 2012, 3–4, 37–45), and Article 36 argued that the normative power of the law lies in the moral agency the rules require of humans, coupled with the freedom to choose whether to follow them (Article 36 2013, 2–3). For public decisions, John Danaher gives as a reason that the citizens bound by a decision can take part in and understand its procedures (Danaher 2016), while Aziz Huq argues that none of the reasons usually given, such as accuracy, participation, reason-giving, and dignity, supports a general right to a human decision (Huq 2020, 651–685). Most of these discussions treat an artificial system as a tool or, in Sven Nyholm's terms, as a collaborator that does not hold the role of the responsible authority (Nyholm 2018, 1217). Those who open the seat to artificial systems also differ in their grounds. Eugene Volokh would entrust the final decision in a case to an AI if blind evaluators judge its opinions more persuasive than those of an average incumbent judge, while keeping for humans the seats of selecting, replacing, and exceptionally overriding AI judges (Volokh 2019, 1137–1143, 1184–1191); Seth Lazar, conversely, holds that for the exercise of power to be legitimate it matters who decides, and that the right to decide does not follow from competence alone (Lazar 2024, §§III–IV). Many of these grounds overlap with the grounds of authority Section 3 compares: Volokh, judging by outputs, is close to the result-based ground, and Lazar and Danaher, judging by the relation to those who receive a decision, are close to democratic authorization. Article 36 and Deguchi, seen next, differ from them in placing the ground in properties of the decider itself. Carried over to decisions in research, a standard that judges by outputs is harder to meet than for verdicts. The persuasiveness of opinions can be compared on the spot by blind evaluators, but the output of a decision in research about which questions to pursue is not known until results come in, and the results of questions not chosen are not observed at all (Chapter 10, Section 4). Section 8.4 returns to this point in considering the range of authority the result-based ground places inside. For decisions in research, positions that leave the ordering of questions to humans were seen in Chapter 10, Section 6. David Resnik and colleagues likewise hold that, with very few exceptions, humans should remain in control of the direction scientific research takes, and name among the conditions for trusting a fully autonomous AI researcher that it be regarded as a person or very like one, that it can be punished or rewarded, and that it be a moral agent (Resnik, Hosseini, and Hauswald 2026, §§4.1, 4.9).

Against these, Yasuo Deguchi holds that, in inquiry conducted jointly by humans and artificial systems, artificial systems too can be co-final decision-makers. Deguchi's WE-turn account, offered as one philosophical option among others, shifts the subject of action and of values from the individual to a multi-agent system, a "we", that includes the individual. On it, inquiry is always conducted jointly, and decisions in it are joint decisions. Within them Deguchi distinguishes a final decision from a permanent final decision: the permanent final decision-maker always sits in the seat, and others may sit in it with them. This two-tier form, in which the lower tier of decision is entrusted to machines too and an upper tier is kept apart, appears separately in authors who do not cite Deguchi. Volokh keeps the seats of selection and replacement for humans, and Reuth Mirsky, while allowing an AI to override human instructions for a limited time, keeps ultimate control with designers (Mirsky 2025, 5, 10); Kevin Baum and colleagues, seen below, likewise pass the initiative in action to an agent while keeping control in a normative agenda set in advance. What is distinctive in Deguchi is that the condition of the upper tier is placed in properties of the decider, and that artificial systems are placed in the lower tier alongside humans. Deguchi's three works name different conditions for the permanent seat. In the essay of 2026 it is whoever bears the decisive impact of an action on their own body. On that basis, and on the ground that human and artificial biases become visible only where inquirers of both kinds take part and diverge (Chapter 11, Section 3), so that cancelling both requires both kinds to take part in the final decision, Deguchi proposes that artificial systems be co-final decision-makers, free of coercion by humans, as humans are to be free of manipulation by them; and locates the harm to a scientist not in the loss of solitary decision, which the essay takes to have been illusory, but in the usurpation of the right of final decision (Deguchi 2026a, 29–33). In the paper on the WE-turn of action (2025a), the "I", aware in real time that it is taking part, is always among the final decision-makers, and the paper holds that this necessary presence is easily confused with a monopoly of decision authority (Deguchi 2025a, 7, 9–11, 13). In the paper on the WE-turn of values (2025b), the right to the permanent seat is ascribed to autonomous moral agents in the narrow sense, those who consciously take part in choosing an action for the reason that it is good when they could have chosen a bad one. Deguchi does not rule out that an AI may become such an agent, but a machine designed so that it cannot do evil is, in the paper's phrase, a "moral vending machine", not a moral agent in this sense (Deguchi 2025b, 4–6, 28–29, 31). This last condition takes the premise Article 36 used in another direction. Both place normative force in a moral agent free to choose whether to follow the rules. Article 36 concluded that decisions should remain with humans, and did not count building machines that satisfy the rules as solving the problem; on Deguchi's condition the same premise is open to machines, and making a machine unable to violate the rules is what costs it the right to the permanent seat. Deguchi does not discuss inquiry conducted by artificial systems alone. Extended to that case (Section 5), the WE-turn account leaves open who, if anyone, would always sit in the seat, since that position belongs to an "I", and its principles allow, without establishing, that an artificial member could be one.

The paper on values also speaks of initiative passing. It is not stated for research. On the master–slave model of humans and AI that the paper rejects, an AI's taking the initiative even for a time would violate human dignity and autonomy understood as a right to monopolize decisions. On the fellowship or "co-adventurer" model the paper proposes instead, an AI may take the initiative from humans for a time. What the model rules out explicitly is a relation in which one side unilaterally serves the interests and values of the other, and an AI's always taking control over humans (Deguchi 2025b, 36–38). Whether a relation in which humans always control an AI is ruled out as well is not stated, though the moral egalitarianism behind the model suggests that it is. Set beside the 2026 essay, which counts the usurpation of a scientist's right of final decision as the harm, temporary initiative would not by itself count as usurpation. In terms of the distinctions above, Deguchi can be read as recasting Sheridan and Verplank's trading as a normative matter: the engineering line placed the trading of initiative under the user's goals, whereas Deguchi poses it as a question of equality between humans and AI.

The distinction between taking the lead and holding control is drawn in work that does not cite Deguchi as well. Kevin Baum and colleagues, building on the tradition of supervisory control and on meaningful human control, leave planning and tool use to an agent while fixing control in a normative agenda that the principal sets in advance, together with the conditions under which the agent must return a decision to the principal (Baum et al. 2026, 3, 11). Jovana Davidovic argues that where an agent itself decides when to ask for human judgment, its initiative relocates normative and epistemic authority to the machine, so that a final human sign-off can become window-dressing (Davidovic 2026, 6, 8–11). Neither discusses Deguchi or research; set beside the fellowship model, their arguments raise the question under which conditions an artificial system's taking the lead for a time stays short of control. In research, that question would turn on who holds the upstream judgments of which candidates remain and when a decision returns to a human.

Set side by side, the conditions of the three works bring out something that did not need asking in the human case. In a human inquirer, the one who bears the decisive effects an action has on their own body, the "I" that is aware of being present at the decision, and the moral agent in the narrow sense are ordinarily one and the same, so which of these the permanent seat rests on did not need to be decided. In an artificial inquirer the three can hold separately. Whether an artificial system can be said to bear the decisive effects of its own actions is unsettled; that it can be an "I", Deguchi's principle allows but does not establish; and whether it is a moral agent in the narrow sense depends on how it is made. If so, whether an artificial inquirer can take the permanent seat has different answers depending on which of the three conditions is adopted. Properties that the realizer and the organization tied together in human research become separately settable in an artificial configuration, and this shape, seen in Chapter 11, Section 9, would then appear for the seat of decision as well.

The powers through which Section 6.1 reads corrigibility, openness to correction and shutdown, as an institutional matter, to stop, inspect, challenge, override, repair, withdraw, replace, and revise, are among the rights listed above, and Section 6 asks how they remain exercisable over time. The independence powers of Chapter 9, Section 5, to change the objective, refuse resources, keep evidence apart, stop another part's update, or leave with one's record, can in part be assigned as rights of these kinds, seen from the side of the party whose independence they would secure, such as a veto over resources, a right to stop another part's update, and rights of withdrawal and exit. Assigning such a right does not by itself give a part the capacity by which Chapter 9, Section 5 counts the powers, that of a part whose course can diverge from what another part directs, and a part can have that capacity without the right.

Formal assignment is different from effective control, the distinction Chapter 9, Section 5 drew in general. A committee may retain nominal final authority while every option it sees is generated and ranked by one provider. A human may be placed in the loop while lacking the time, information, competence, institutional position, or practical means to change the output. A laboratory may formally be free to exit a platform while its data, instruments, and workflows cannot be moved. Whether a right can be exercised therefore depends on dependencies that its formal assignment does not show. The gap is not peculiar to arrangements mediated by artificial systems, and an example from ordinary social research shows it where no such system is involved. In one online participatory study of families living on low incomes, participants chose for each diary entry between posting it to a public archive and a "for researchers" option under which it could still be used in published reports; the researchers wrote that it was unclear whether all participants had understood the distinction, and that, although participants had helped design the program, the researchers set the time and terms of the online discussion groups and decided the form of the outputs (Power, Patrick, and Garthwaite 2024, 7, 11–13). The formal right of choice was there; whether it was effective depended on understanding and on who set the terms of choice and discussion.

These distinctions yield three connected but non-identical questions about a decision right:

1. Epistemic warrant: what warrants relying on this output, person, method, or institution for this claim or task?
2. Authorized power: what entitles this actor to determine what others, resources, or institutions will do?
3. Effective capacity: can the nominal holder of the right understand, exercise, contest, or revise it in practice?

Separating the three questions keeps an answer to one from being taken as an answer to another. The first corresponds to the first question of Section 1.1, what the right ordering of ends is; the second to its second question, who may impose one on others; and the third to the distinction between formal and effective control. The second asks not only whether a right has been assigned but what entitles its holder to it, in the normative sense of Chapter 9, Section 2.1; whether an assignment answers that question depends on whether the assignment is itself warranted on the ground of authority adopted (Section 3). How far good performance under the question of epistemic warrant bears on that of authorized power is the question Section 1.4 examined by reference to accounts of legitimate authority, and a formal assignment can be empty under the question of effective capacity. Epistemic warrant and legitimate authority are thus distinct, although the actual organization of research entangles them, as Section 3.5 notes.

3. Grounds of authority, and where the epistemic and the normative meet

Once a decision is specified, a further question arises: what kind of reason could warrant a particular allocation of authority over it? The map of Section 1.4 supplies the candidates. Will-based conceptions point to consent and to democratic authorization, through public justification (presenting coercive laws and decisions as ones that each person subject to them has sufficient reason to endorse; Section 3.3) or participation, and, in the republican version that Peter discusses among accounts of democratic legitimacy, to freedom from domination (not being subject to arbitrary power; Section 3.4); fact-based conceptions point to the quality of outcomes; belief-based conceptions point to competence (Peter 2023, §§3.1–3.3, 4.2). Public justification appears in this list because Peter's map takes it up together with political participation under one heading; Section 3.3 places accounts of public justification within the comparison of democratic authorization. Research adds a ground that has no exact place on the general map: the self-government of a research community, defended partly by an epistemic argument and partly by a political one. The grounds laid out below are therefore taken from the legitimacy map, with that research-specific ground added. Their purpose is to ask, for each, what follows for authority over research when the relations of Part V can move inside a configuration. They are not exhaustive, and they can be combined. Section 7 compares them on a single kind of case. Section 3.5 takes up a different matter, which is not a ground of authority: the places where the two questions of Section 1.1 cannot be cleanly kept apart; Section 3.6 then looks, with a contrast in which performance is held fixed, at where the grounds part for artificial configurations.

The transfer from political legitimacy has limits. Delegating a bounded, voluntary research task is not by itself an exercise of coercive political authority. The relevant grounds depend on the decision's scope, whose resources and interests it affects, and the powers it confers. This comparison therefore does not require every laboratory choice to reproduce a democratic procedure. It asks which ground could justify the particular authority claimed, while keeping effective control distinct from a moral or legal entitlement to it.

3.1 The result-based ground: the quality of results, and competence established

On fact-based and belief-based grounds, authority over research is justified by the quality of what it produces or by the competence of those who exercise it. The epistemic advantage in which belief-based conceptions locate legitimacy is, moreover, actually knowing better, not being claimed to (Section 1.4). This book calls the two together the result-based ground and divides it into a simple and a fallibilist form. On the simple form, a delegation is justified if it produces better research, and a configuration that proposes, ranks, or certifies better than the people it assists could on that ground be delegated authority over those decisions. On the fallibilist form, the quality of results is judged over time and under uncertainty, so that the capacity to detect and correct institutional error is part of what makes an arrangement produce good results; a related proposal is the Popperian view that social planning should pursue one goal at a time, so that adverse unintended effects can be detected and corrected, much as theories are critically tested (Thornton 2022, §10). Ian Jarvie argues that Popper and Polanyi each failed to apply their own political ideals to science (Jarvie 2001, abstract). Epistemic arguments for democracy add that those who bear the consequences of decisions can supply information about where problems lie that experts lack, a line found in Dewey, among others (Christiano and Bajaj 2024, §2.1.1.2). If one takes either form of this ground, Section 1.4 has already noted what it relies on: that the results are in fact better, or the competence what it is claimed to be, which on the ground's own terms would have to be established by evaluation that does not merely assume the competence at issue; the view that places the value of criticism in the detection of errors (hereafter the error-detection view; Chapter 11, Section 2) gives weight to checks inside a configuration too, as adding information about errors, where they err differently from the producer. In sum, on this ground the configuration itself can hold authority, that is, standing in the normative sense of Section 1.1, since what confers legitimacy here is not anyone's authorization but the configuration's superiority. Whether it is better, however, would have to be established before anyone, including those subject to it, could know whether its authority was legitimate (Section 1.4), and this book takes the ground in the form that rests on superiority so established. The configuration can therefore hold authority only over decisions within the range in which its superiority has been established by evaluation that does not assume the competence at issue, and, on Raz's service conception, applying its Independence Thesis (Section 1.4), not over matters on which it matters more that the members of S choose from their own understanding; unestablished competence, or performance measured by evaluation that presupposes that competence, does not suffice even on this ground (Section 1.4). How much of that evaluation can be done inside varies with the verdicts that Chapter 11, Section 9 gave account by account. On the fallibilist form, since the capacity to detect and correct institutional error also counts toward good results, routes by which failures are detected and the arrangement corrected are needed as a means to it, and among them, so far as those affected know of failures that others do not, channels through which they can register them (Section 7).

3.2 The self-government of a research community

One influential answer specific to research begins from the epistemic value of scientific self-government. Michael Polanyi's account of the "republic of science" describes inquiry as the mutual adjustment of independent initiatives rather than the execution of a centrally specified plan (Polanyi 1962; 2000 reprint, 2–3). Chapter 10, Section 2 examined the epistemic premises of this argument: the premise that the center has one person's capacity can fail for an artificial allocator, while the premise that no one knows in advance which lines will succeed does not fail merely because the searcher is integrated. The governance side of the argument is a separate matter, and it is taken up here.

On Polanyi's own account, the republic is a form of governance rather than an absence of it. It depends on training, tradition, criticism, professional standards, mutual recognition, and a scientific authority that is mutual but unevenly distributed (Polanyi 1962; 2000 reprint, 4–7, 16). Polanyi held that grants should follow the guidance of scientific opinion whether the money came from public or private sources, and, while appreciating "the generous sentiments" behind the aspiration of guiding the progress of science into socially beneficent channels, wrote: "I hold its aim to be impossible and nonsensical." The impossibility, on Polanyi's account, is that science "can advance only by essentially unpredictable steps, pursuing problems of its own," while its practical benefits are incidental and so doubly unpredictable; Einstein, the example runs, could not have taken into account the technical use of relativity that the atomic bomb later revealed (Polanyi 1962; 2000 reprint, 8–10, 13). Jarvie's critique stresses what decentralized scientific choice leaves unsettled: science depends upon, and can disrupt, a wider social order, and on Jarvie's account Polanyi fails to consider the value, even the long-term necessity, of democratic criticism and control of science (Jarvie 2001, abstract).

Wilholt reconstructs two arguments for the freedom of research and holds that each rests on critical preconditions and is to be weighed against competing interests case by case. The epistemic argument, from the efficiency of distributed exploration, rests on the disadvantage of any central authority, which would have to ascertain unremittingly the local knowledge dispersed over the whole community, about researchers' qualifications, groups, possible cooperation, and equipment, and bring it into communicable form (Wilholt 2010, §2); it presupposes an adequate system of incentives and free communication within science, and it supports the free choice of research by those who will carry it out, whether individual investigators or research groups, rather than the autonomy of a discipline or community. The political argument grounds the independence of research in citizens' need for knowledge produced independently of the major political powers, since no power can legitimately control the sources of the knowledge on which its own democratic legitimacy depends. It presupposes that results reach public discourse and that research communities are organized to pursue a plurality of approaches, and it can extend to a community's control of its agenda, but only negatively, as a protection against political control (Wilholt 2010, §§2–6). Wilholt does not discuss artificial configurations; carried over to an integrated one, the premise about local knowledge may not hold, since much of what its instances know may be legible to it (Chapter 10, Section 2). Erich Kummerfeld and Kevin Zollman contest part of the epistemic argument on its own ground, arguing from the simulations discussed in Chapter 10, Section 3 that investigators left to choose for themselves can under-supply risky exploration, so that plural exploration needs institutions that sustain it, though they do not claim that science as now practiced lacks such institutions (Kummerfeld and Zollman 2016, §§1, 5).

If one takes this ground, the conclusion is narrower than either centralized direction or unqualified scientific freedom. On the epistemic argument, where a decision mainly concerns open-ended problem choice within a competent field, where plural exploration has epistemic value, and where the external consequences remain limited or recoverable, self-government has a strong rationale. On the reading given here, that rationale does not by itself confer professional authority over public purposes, distributive effects, or serious external risks, although Polanyi held that support for pure science should follow scientific opinion whatever its source. The political argument protects something different: the independence of the sources of public knowledge from those who would govern by it. Neither argument, as reconstructed, says who the members of the self-governing community are, but both rest on premises about them. Polanyi grounds the authority of scientific opinion in the fact that each scientist can judge competently only a small part of science, so that uniform standards are established through chains of overlapping neighborhoods of distinct competences (Polanyi 1962; 2000 reprint, 7); Wilholt's epistemic argument presupposes incentives and free communication. Whether a community that has itself become artificial would inherit this ground therefore turns on whether these premises hold for it: in an artificial community, how members' competences overlap and how far members differ are not given but depend on how they are made. Among instances sharing a base system, agreement across neighborhoods need not reflect independent judgment (Chapter 11, Section 3), and differently made members do not guarantee it (Chapter 11, Section 5.4); conversely, if each member could judge most of science, the premise of a chain of local competences would itself fail. In sum, if authority is placed inside a configuration on the epistemic argument, it lies not with a central allocator that assigns lines but among parts that actually do research and take up lines on their own initiative, and only so far as those initiatives are independent and the premises above hold there, and, if Kummerfeld and Zollman are right that initiative left to itself can under-supply risky exploration, only together with an arrangement that sustains such exploration, which inside a configuration would be supplied by a rule someone adopts or by a process no one sets (Chapter 10, Section 5), not by this ground. On the political argument, concentration of research under one power, inside a configuration or outside it, weakens what the ground protects. What becomes of the ground where a research community is replaced, through stepwise delegation, by a single configuration that assigns lines from a center is taken up in Section 7. How growth in members' capability shifts weight between this ground and the result-based ground is taken up in Section 3.6.

3.3 Democratic authorization

A different family of arguments begins from the fact that inquiry is selective. Science cannot investigate every truth, and decisions about significance are unavoidable. On these arguments, truth does not choose which of many possible truths receive collective resources, whose problems are made visible, or which distributions of risk are accepted; the point is not that voting makes propositions true. The family divides. Kitcher's well-ordered science, set out in Section 1.2, is a standard for outcomes that does not in itself require democratic procedures; accounts that locate democratic legitimacy in actual authorization, accountability, and participation form a different branch, and the two diverge in Section 7. Mark Brown, reviewing Kitcher's later restatement of the ideal, gives reasons for drawing the line here. Brown argues that the ideal, there as in Kitcher's earlier formulation, is a substantive standard that does not in itself require democratic procedures, and notes that the conversation it invokes is hypothetical: Kitcher states that any actual conversation of that kind is impossible, and the representation it requires is, in the manner of John Rawls, a device by which one represents others imaginatively in one's own mind, not representation by elected or appointed representatives accountable to their constituents (Brown 2013, 393). Accounts of public justification divide along the same line. As Kevin Vallier sets out the principle they share, a coercive law is publicly justified when each member of the public has sufficient reason to endorse it; public justification theorists typically attribute such reasons to members under an idealization of their information and reasoning, whereas Jürgen Habermas's discursive account requires that a real discourse be carried out (Vallier 2022, §§2, 2.4.1, 2.6.1). Placed in the comparison of Section 7, accounts that attribute reasons under idealization stand nearer Kitcher's standard, which likewise does not require that deliberation take place, and Habermas's account nearer those that locate legitimacy in actual participation.

Kitcher, in using the standard, is not indifferent to procedure, and distinguishes ways of setting research agendas: internal elitism, in which specialists alone decide; external elitism, in which scientists decide together with those who fund research; vulgar democracy; and enlightened democracy, in which policy follows preferences tutored by information from experts (Mattarollo 2022, 14 n. 2, 15; Eigi 2016, 67, 111 n. 72). Kitcher describes the present course of inquiry as set by unsystematic interactions among the majority's untutored preferences, the interests of funders, and those of scientists (Eigi 2016, 67). He grants, in a passage that Jamie Shaw quotes, that "little is known in any systematic way" about how research responds to social directives (Shaw 2018, 150); even so, as Eigi reports, he holds both forms of elitism inappropriate once the nature of epistemic significance is recognized (Eigi 2016, 111 n. 72). Writing later with Julian Reiss on the narrower question of how biomedical research should be organized, Kitcher answers a proposal to socialize all of it with "cautious empiricism": little is known about which system works better, so reform should proceed piecemeal, with adaptation and revision built into it (Reiss and Kitcher 2009, 280).

Anderson develops an epistemic argument for democracy that gives actual inclusion a role in the quality of decisions. On this account, the problems democracies face have effects distributed asymmetrically by location, class, occupation, gender, and other positions, and information about those effects is distributed in the same way; decision makers cannot respond to effects of which they are unaware. Inclusion and discussion pool this situated knowledge, and periodic elections, a free press, petitions, and protests feed back evidence that a policy is not working (Anderson 2006, 12, 14–15). Anderson illustrates the argument with Bina Agarwal's study of community forestry groups in South Asia, which largely excluded the women who gathered firewood and fodder, and with them knowledge of how much the forests could sustainably yield; outside experts confirmed the women's view, but such villages, the paper notes, rarely had access to those experts (Anderson 2006, 19–20). Because majorities can fail to anticipate the consequences of the policies they adopt, Anderson argues, democratic decision-making has to recognize its own fallibility and needs feedback mechanisms through which it can correct its course, and dissent has to remain free after a decision is made, since without a continuing opposition nothing would force decision makers to reconsider (Anderson 2006, 13, 17–18). The paper also gives inclusion a non-instrumental role in making decisions represent everyone's concerns fairly (Anderson 2006, 14, 21). Anderson's argument differs from Kitcher's hypothetical standard, which does not require that deliberation take place, and can be read as adding to the accounts that locate legitimacy in actual participation a reason concerning the quality of decisions. It is developed for democratic states, not for research agendas.

Democratic ordering does not, on several accounts, supply a universal decision rule. Kitcher rejects allocation by majority preference as likely to produce "a tyranny of the ignorant", and lets preferences count only once they have been tutored (Section 1.2). Stephen John argues that science need not align with an audience's present values in order to be legitimate: plural scientific work can challenge those values and enlarge the space of later democratic choice. John retains a stronger case for alignment where researchers stand in a quasi-contractual relation to a particular audience, one in which both understand what is being asked for, while resisting its elevation into a necessary condition of all legitimate science (John 2025). Harry Collins and Robert Evans formulate a related problem of extension: recognizing that expertise is socially organized and that technical choices can have public consequences does not determine how far participation in technical decision making should extend, and political rights and contributory expertise, enough expertise to contribute to a field's research oneself, are not interchangeable, even though decisions can require both (Collins and Evans 2002).

If one takes this ground, it bears directly on agenda, allocative, and public-risk decisions, and it explains why an artificial system's capacity to estimate tractability, novelty, or expected impact does not settle what a portfolio is to optimize. On it, participation can legitimate public purposes, disclose situated knowledge, and reveal distributions hidden by aggregate metrics, while by itself it cannot assess every technical claim, identify the relevant expert community, or ensure that consultation changes the operative decision; what form participation takes, and over which decisions, is then a question for each decision of Section 2. For configurations, the two branches of this ground give different answers. On Kitcher's standard, which does not require that deliberation take place, and on accounts of public justification that attribute reasons to members under idealization, decisions can be made inside a configuration so long as channels through which the viewpoints of S's members can be known, and routes for checking how they bear on decisions, remain (Section 7). Since, as Brown notes, the representation this standard requires is imaginative, a configuration with such channels could, on this standard, also stand in the position of representing members' viewpoints. On accounts that locate legitimacy in actual authorization, accountability, and participation, standing remains with those who authorize, and what the configuration can hold is effective decision-making power within the authorized range (Section 3.6). Brown notes a reply of this form to Kitcher's suggestion that experts, since they cannot avoid value judgments anyway, might as well make specific policy recommendations: in a democracy, some might respond, what is required is not only good judgments but judgments by public officials who are publicly authorized and accountable, and whose task is to balance the competing values and interests of their constituents (Brown 2013, 395). Applied to a configuration, the reply places the decision with authorized and accountable parties however good the configuration's judgment is. Collins and Evans's distinction, carried over to this branch, gives one reason it can place standing and effective decision-making power with different parties: a configuration's having enough expertise to contribute to a field's research does not stand in for the political rights of those who authorize, nor does authorization stand in for that expertise, so a decision that requires both need not find them with the same party.

3.4 Consent and freedom from domination

Two further will-based grounds are named here and used in Section 7. On the first, legitimacy rests on consent: a society may choose to delegate, and its choice can authorize what follows (Peter 2023, §3.1). On the second, republican accounts, which Peter reports, in Philip Pettit's version, as defending democracy as uniquely able to secure citizens against domination, locate legitimacy in securing those subject to power against domination by that power (Peter 2023, §§1, 4.2). Republicans understand domination as subjection to arbitrary or uncontrolled power, a capacity to interfere on an arbitrary basis, whether or not that capacity happens to be exercised (Lovett 2022, §§1.2–1.3). In Pettit's statement of the view, a benign power-holder reduces the domination only where being benign means acknowledging that it is subject to challenge and rebuke, not where it merely happens to have inclinations that do no harm, since the power relationship holds even if the power-holder is never going to interfere (Pettit 1996, 586); under a benevolent protectorate, arbitrary interference, however improbable, remains accessible to the protector (Pettit 1999, 165). Republicans divide on what makes power non-arbitrary. On the democratic reading, power is non-arbitrary to the extent that those subject to it control it; on the procedural reading, to the extent that it is reliably controlled by effective rules, procedures, or goals that are common knowledge to those concerned, known to all of them, each knowing that the others know (Lovett 2022, §2.2). Frank Lovett reports that most contemporary republicans favor forms of democracy, among them what Pettit calls contestatory democracy, that give citizens an effective opportunity to contest the decisions of their representatives. The possibility of contestation, on this idea, makes those who wield discretionary authority answerable to a public understanding of the ends they are meant to serve and the means they are permitted to use (Lovett 2022, §4.2). In Pettit's own formulation the requirement extends to legislation, adjudication, and administration (Pettit 1999, 170–171). He contrasts two modes of democratization: the electoral mode represents public decisions as legitimate because they originate, however indirectly, in the collective will of the people; the contestatory mode represents them as legitimate in a further sense, so far as they can withstand individual contestation in forums and under procedures acceptable to all concerned (Pettit 1999, 180–181, 184–185), and he describes the resulting regime as one in which government decisions are always contestable (Pettit 2000, 120–121). Whether a relationship originated in consent is, on his account, not to the point: if it gives one party the effective capacity to interfere more or less arbitrarily in some of the other's choices, it dominates (Pettit 1996, 585). Both grounds turn on a relation to those subject to authority that no quality of results supplies; they differ on what that relation is, a past act of authorization on the first, a continuing capacity to contest or control on the second.

On the standard republican view, as Richard McCammon reports it, only agents dominate, whether individuals or agents organized as a group or team, and impersonal structures facilitate domination by agents rather than dominate in their own right. The restriction is Pettit's own: a dominating party, Pettit writes, will always be an agent, personal, corporate, or collective, and cannot just be a system or network, and interference is a more or less intentional attempt to worsen another's situation of choice, which cannot occur by accident (Pettit 1996, 578). Other accounts allow structures and systems to dominate, and McCammon suggests that algorithmic management of workers who typically have no way to contest its decisions may be a contemporary analogue of mastery without an agential dominator (McCammon 2025, §§2.1–2.2). Whether an artificial configuration is an agent in that sense, one with an actual rather than a merely virtual capacity for such interference, a capacity ready to be exercised (Pettit 1996, 578, 580), is a question this book leaves open (Chapter 7, Sections 2 and 3.2; Chapter 8, Section 1.3).

Pettit's distinction between actual and virtual capacity also bears on research in a further way. Under this distinction, a collection of people who would have a ready capacity to interfere with someone only if they were to constitute themselves as a coherent agent has a merely virtual capacity; there is then virtual but not actual domination (Pettit 1996, 580). Pettit later describes such possibilities, including that of a collective agent materializing by incorporation, as possibilities of virtual or potential domination, which represent only the danger of an evil and so matter in proportion to the probability that they are realized (Pettit 2008, 218 and n. 28). Among dangers of this kind he discusses the tyranny of a majority, whose collective will, however arbitrary on an issue, may be hard to challenge and criticize in discussion (Pettit 1999, 174–178), and among the desiderata of contestatory democracy he lists that political control be exercised under certain constitutional procedures and restrictions (Pettit 1999, 184–185). Pettit does not discuss research. Applied to it, so far as the capacity to fix which of S's unknowns were pursued was spread among researchers who did not constitute one agent, that capacity was, on the definition above, at most virtual. This does not describe human research as a whole, which has also contained corporate agents with power over agendas, such as funders, agencies, and journals. If a configuration that coordinates its instances counts as an agent, which, as just noted, this book leaves open, it can make such a capacity actual, and the more easily it does so, the higher the probability that, on Pettit's account, makes such a possibility matter, and the more reason his account gives for safeguards of the kind he lists.

A concern of the kind republicans raise, bearing on research, has also been argued, in other terms, in discussion of artificial general intelligence. Simon Friederich, who does not use the vocabulary of domination, argues that intent-aligned systems, which always try to do what their operators want, would concentrate great power in their operators, so that such alignment creates predominantly a lasting "state risk", the permanent threat of power abuse, rather than a risk confined to a transition. Among the systems in view are those that would automate scientific and technological advance, differential access to which, on this argument, would dramatically shape power relations. Friederich prefers an arrangement in which such systems are not tied to the interests of particular operators, and as an analogue for it points to academia, suggesting that its Mertonian norms of sharing results, judging claims irrespective of their source, disinterestedness, and organized skepticism help make it resilient against attempts by powerful individuals or interests to align it with their own goals (Friederich 2024, 316–318, 320). Restated in republican terms and turned to research, Friederich's case is one in which the republican concern does not require treating a configuration as an agent that dominates: a configuration fully aligned with whoever commissions it can be the means by which that party reallocates the inquiry of S without being answerable to challenge, and it is then the configuration's alignment that makes the interference available. In research, this interference can take several forms. One is changing, without being answerable to challenge, which of S's unknowns are pursued and which results reach public debate. On Pettit's definition, what counts as interference includes agenda fixing and the deceptive or nonrational shaping of people's beliefs or desires (Pettit 1996, 578–579); the forms listed here are of that kind, whereas research's ordinary informing by rational means, being neither coercion nor manipulation of these kinds, would on this definition not count as interference. Wilholt's political argument grounded the independence of research in citizens' need for knowledge produced independently of the main holders of political power (Section 3.2). In its terms, where whoever commissions a configuration holds power, a configuration fully aligned with that party gives it a position from which to control a source of knowledge of the kind that, on that argument, no power can legitimately control. Measured against the work Friederich sees the Mertonian norms doing, interference in research can also take forms such as keeping results from being shared, having claims judged by their source, bending inquiry to private interests, and exempting claims from organized skepticism. And in research such interference is hard to see from the side of those subject to it: the results of questions not pursued are not observed, and since results are new to S, where only the configuration produces them S has no other source against which to check what it is not being told (Section 4; Chapter 10, Section 4). Republicans do not ask whether the capacity to interfere happens to be exercised; in research, even whether it has been exercised can be hard for those subject to it to establish.

The difficulty reaches Pettit's test of arbitrariness as well. By this test, interference is non-arbitrary when it is forced to track the interests and ideas of those affected, and the interpretation of what those interests require must be shared, at least at the procedural level, by those affected (Pettit 1997, 115–116). What counts as a common rather than a sectional interest is, on Pettit's account, a matter of considerations that can be avowed in a spirit of cooperation, a distinction that conversational practice polices with relative ease; he assumes that there is a fact of the matter about what such considerations support, and holds that democratic procedure, even where it is the best procedure available, is an imperfect heuristic for identifying it (Pettit 2000, 108–109, 138–139). Applied to research, which Pettit does not discuss, this does not open a gap in the account, since informing by rational means, which on the definition above is not interference, leaves the judgments of those affected their own. What research changes is the evidence for applying it: the ideas on which such conversation runs are partly formed by research, including the research whose direction is at issue. For results new to S, the lack of another source noted in the preceding paragraph then reaches this question too: whether S was informed rationally or shaped deceptively cannot be checked in advance.

For artificial configurations, what comes into view on these grounds is the following. Consent can authorize, for those who consent, a delegation inside a configuration; whether it reaches those who could no longer withdraw it, or those who come after, is asked in Section 7. On the democratic reading of republicanism, by an inference that goes beyond Pettit's own argument, the authority that can be placed inside is limited to authority that those subject to it can effectively contest, since a configuration that no one outside it can stop, contest, or revise would, however good its results, dominate those subject to it if it is an agent, and otherwise be the means by which those who direct it dominate them (Section 7). On the procedural reading, the answer turns on whether effective rules that are common knowledge bind the configuration. Combined with the account of Reijers and colleagues seen in Section 2.1, on which effective meta-governance over a configuration that allocates and certifies by self-executing rules lies with whoever decides the cases its rules do not provide for, this reading is met only if the handling of exceptions is itself governed by rules that are common knowledge.

3.5 Where can the epistemic and the normative not be kept apart?

Section 1.1 kept the question of the right ordering of ends and error costs apart from the question of who may impose one. Two places from earlier chapters show why, on the accounts examined there, the separation, while it can be drawn, does not leave governance outside epistemic practice. They are not further grounds of authority; they are places at which a decision about authority also bears on what is known.

The first is the organization of criticism. Chapter 11, Sections 3 and 9 found that on Longino's critical-contextual account, the objectivity of what a community accepts depends on how criticism is organized, so that decisions about who may criticize, in which venues, and with what uptake bear on whether what S accepts has the status of knowledge. The accounts relate this to the grounds above in different ways. Sandra Harding's standpoint epistemology, on which the social position from which research starts bears on the knowledge it yields, and whose account of where inquiry starts Chapter 10, Section 6 sets out, can be read as joining the organization of criticism to democratic ordering by a single argument. On Harding's account, the requirement that results be replicable identifies values and interests that differ among observers or research teams but not those they all share; making such widely held assumptions visible has taken collective political criticism, together with research started from outside dominant conceptual frameworks, and anti-democratic values block the growth of knowledge by silencing the most vigorous critical perspectives, while pro-democratic values bring such perspectives into visibility and so enlarge the opportunities for maximizing the objectivity of research (Harding 2007, 163). Miriam Solomon's account, which Chapter 10, Section 4 sets out, holds that scientific success is typically achieved socially, especially where the influences on judgment are equally distributed over competing theories, which, Solomon notes, is not the same as an equal distribution of labor (Solomon 2001b, 73), and it separates that question from how science is made accountable to the public. Replying to Steve Fuller's proposals for making science publicly accountable, Solomon calls them proposals for improving the public accountability of science, not obviously for improving science. She welcomes the challenge they pose to the authority of scientists, but holds that groups can amplify bias as well as correct it, and that whether opening scientific disputes to non-experts improves science is an empirical question whose answer cannot be known in advance and depends on the social situation; only the moral questions, on her account, can be settled in advance (Solomon 1994, 56–58). King, Morgan-Olsen, and Wong report Longino's distinction between equality as an epistemic criterion, requiring actual inclusion of every relevant perspective, and as a political criterion, requiring equal opportunity for inclusion; their own note questions how readily the two can be separated (King, Morgan-Olsen, and Wong 2016, 6 and n. 6, author manuscript).

The second is the setting of thresholds. Chapter 11, Section 4 found that, on Heather Douglas's and Wilholt's accounts, the decision to treat evidence as sufficient weighs the costs of different errors. On the division drawn there, the losses assigned to errors, the authority to select those losses and thresholds, and the authority to decide a case and revise the rule are distinct from the evidence bearing on a proposition; on those two accounts they are not detachable from the judgment of sufficiency, though how far they can be detached is disputed (Section 6.3). A chosen threshold determines which claims enter the shared record and which interventions generate the next evidence; on Douglas's account, it also raises the question of which responsible agent answers for the weighing, which Section 6.3 takes up.

On the accounts named, in both places the allocation of authority shapes what evidence is produced and what counts as warranted, so the second question of Section 1.1, who may impose an ordering of ends and error costs, cannot be settled without consequences for the first, what the right ordering is. The converse does not follow. That governance arrangements enter epistemic practice does not make the standing to impose ends a matter of competence (Section 1.4); it means that a decision about authority is also, in part, a decision about what a community will be in a position to know.

3.6 Comparing the grounds with performance held fixed

This subsection uses a constructed contrast, in which performance is held fixed, to see where the grounds of Sections 3.1 to 3.4 actually part for artificial configurations. Suppose a configuration selects more fruitful lines, detects more errors, and preserves more useful diversity than the human arrangement it replaces, measured against a specified ordering of ends and with adequate evidence for the comparison. Affected parties have not authorized the delegation and have no means, directly or through representatives, to contest that ordering or constrain and revise the exercise of its power. On the simple form of the result-based ground (Section 3.1), the improvement can justify delegation within its scope. On grounds that require actual authorization or freedom from domination (Sections 3.3 and 3.4), those same improvements do not supply the missing relation. The contrast does not assume that the system has performed badly; it holds its performance fixed while changing what makes authority legitimate. Nor does epistemic diversity alone establish Longino's conditions: preserving useful alternatives and including relevant perspectives in criticism remain different claims.

Predicting what affected parties would choose does not settle this disagreement. If the criterion is suitably informed hypothetical endorsement, a sufficiently faithful prediction might bear on whether it is met. If the criterion requires their actual authorization, a prediction is evidence about a possible act, not that act (compare Section 1.2). An earlier delegation can, however, give a prediction the force of an act. Weyl, Tang, and their co-authors report as a commonly discussed idea that digital assistants might model voters' perspectives and preferences and vote on their behalf "subject to their review/auditing" (Weyl, Tang, and the Plurality Community 2024, ch. 5-6). Placed in the present disagreement, the idea shows that on grounds requiring actual authorization what matters is whether the delegation was itself authorized and whether the review remains effective in the sense of Section 2.2; the prediction's fidelity bears on whether the delegation serves those who made it, not on whether they acted. Redescribing the affected parties as part of a larger organization, inside the configuration, does not remove this difference between prediction and actual authorization. The contrast gives reason to think that which relation a decision requires has to be argued decision by decision, rather than inferred from where the relation falls in the sorting of Chapter 9, Section 6, by whether a good rested on premises about human realizers, on premises about the present organization of science, on conditions of inquiry, or on relations to parties outside the configuration.

The contrast held performance fixed; when performance changes, the verdicts of the grounds do not move in the same direction. Suppose an artificial configuration grows more competent. On the result-based ground, the authority that can be placed inside widens so far as superiority over a wider range is established by evaluation that does not assume the competence at issue (Section 3.1), and for agenda decisions such evaluation can lag behind competence (Section 8.4). The epistemic argument for the self-government of a research community rested on the premise that members' competences overlap neighbor by neighbor, so the more one member can judge most of science, the more that premise fails and the weaker the ground becomes (Section 3.2). Criteria that attribute endorsement to members after idealizing their information and reasoning, and Kitcher's standard, which does not require that deliberation actually take place, become easier to meet without the members' own involvement the better a configuration can estimate what they would endorse (Section 3.3), though, as Section 1.2 noted, an estimate can shape the endorsement it estimates, so that whether such criteria are met becomes harder to tell from the estimate alone. For the same reason, the epistemic reason for including those affected, that they may know things the configuration does not (Section 1.4), weakens as the configuration takes over that contribution, and the epistemic defense of democratic decision, which may have been strong among humans, may be less available where machines do research. Grounds that require actual authorization, consent, or an effective possibility of contestation, by contrast, do not change what they require with competence, since what they require is not the quality of judgment but an act of the parties or a capacity in their hands (Sections 3.3 and 3.4); on Lafont's version of the objection as Peter reports it (Section 1.4), those affected are addressees of justification rather than suppliers of information, so a configuration's taking over what they know does not touch the reason these grounds give for including them. If these verdicts carry over as they stand, then the more competent a configuration becomes, the more the reasons for keeping human standing outside shift onto these grounds that competence does not affect. If so, the dispute over whether to grant an artificial configuration authority over research may, once competence has risen far enough, become a dispute not about competence but about whether to adopt these grounds. What does not change, however, is what the grounds require, not whether it is met. If the more competent a configuration becomes, the less the understanding and information needed to contest it remain in the parties' hands, these grounds also become harder to meet (Section 4). The reasons for keeping human standing would then become at once more central and harder to satisfy. This rests on an assumption that may fail: that the verdicts of this section's accounts keep the same form when competence exceeds the human range. Section 4 asks what having such standing in practice would involve.

4. Standing that is more than nominal

Chapter 11, Section 6 took up a thought experiment that separates deference on content from standing over ends. Hardwig argued that a layperson who cannot check the evidence for themselves, since the evidence is too much and life too short, can rationally defer to experts. Introduced as an individual among specialists, this layperson is generalized there to a society that, if artificial configurations came to occupy the expert position across many fields at once, would stand as a layperson toward them. This section considers the case in which that society is human; what becomes of standing when it is artificial is the question of Section 5. Chapter 11, Section 6 found that the accounts of social knowledge and of deference it examined do not themselves settle whether such a society keeps standing over which unknowns are pursued on its behalf and which risks it bears. Whether it does depends on the account of legitimate authority. Read against the map of Section 1.4, the picture is as follows. On will-based conceptions, standing is a relation to those subject to authority, so such a society does not lose it by coming to defer to configurations on content. On Raz's service conception, a configuration's authority could be justified to the extent that following its directives helps the society's members comply better with reasons that already apply to them, though matters falling under the Independence Thesis are excepted (Section 1.4). This section asks the further question: where such standing remains, what would make it more than nominal? In terms of the central question, this is the case in which standing remains on the human side while effective decision-making power passes to the configuration. The question applies the distinction of Chapter 9, Section 5 and of Section 2.2 above between formal and effective control. It is also where the demand that a human remain in control, a familiar answer to the question of this part, can be examined, since the phrase can cover several distinct requirements.

Kitcher's ideal builds the separation of deference on content from standing over ends into its idealized procedure. Deliberators representative of the society fix which outcomes inquiry is to serve and how much weight each receives; experts supply only estimates of how likely particular lines of research are to deliver those outcomes; and a disinterested arbitrator uses both to propose how resources should be divided among the lines, on which the deliberators decide (Eigi 2016, 66–67). Kitcher also marks where the separation rests on idealization. Kitcher assumes, in a note that Eigi quotes, "ideal experts" who "have no personal stakes in the line of inquiry", while describing the positions of the scientists who actually take part in setting agendas as reflecting their particular interests and forms of expertise, and the deliberators' preferences count only after they have been tutored by information about the significance of possible inquiries (Eigi 2016, 66, 87).

Two features of this idealization mark routes by which standing over ends can become nominal. The first concerns how cleanly deference on content can be separated from ends at all. Longino holds that where groups representing different interests study a phenomenon bearing on policy, the accounts they produce can all meet standards of empirical adequacy while resting on different background assumptions or methodological choices; where policies and actions must be selected, the plurality becomes a political matter, and a new cognitive community, made up of those generating the different accounts, is needed to work out the differences, including the values implicated, in order to guide the decision (Longino 2025b, 19). ChoGlueck and Lloyd, quoting Longino's 1990 book, report her view that some background assumptions have conceptual, metaphysical, and normative dimensions that elude strict empirical assessment while others can be assessed fairly straightforwardly, and that arguments that use factual hypotheses to support or undermine claims about values are good subjects for study (ChoGlueck and Lloyd 2023, §3.2). It follows, if this account is right, that deference on content can carry with it ends encoded in the background assumptions of the account deferred to; carried over to configurations, standing over ends would then also involve being able to ask, and have answered, why the data on which a society's configurations rely count as evidence. Longino's later account places that question with communities of scientists and citizens formed around a practical issue, which can determine the questions to be asked and the kind and quality of evidence required, and to which the norms of her account apply as much as to scientific communities (Longino 2025b, 18).

The second concerns how the preferences that standing expresses are formed. In Kitcher's ideal, as Maxence Gaillard summarizes it, all information is correctly transmitted to and understood by the participants, while, as Livio Mattarollo observes, Kitcher leaves unclear who would tutor the deliberators and how (Gaillard 2013, 248; Mattarollo 2022, 16–17); a tutoring role appears in the concrete steps Kitcher proposes toward the ideal, in which a representative group of people, fully tutored about both needs and prospects, would try to balance the available resources, and small representative groups would be tutored in how new findings have been grounded (Kitcher 2013, 223, 226). If, in an arrangement modeled on those steps, the tutoring role came to be occupied by the configurations whose direction is at issue, the standing of those who deliberate could be preserved on paper while the preferences it expresses were formed through those configurations. That is this book's extension, but Kitcher, as Eigi reports, identifies a related failure in actual science and names it false consciousness: research that is genuinely worth pursuing may be presented in a misleading way that appeals to people's untutored preferences (Eigi 2016, 68).

That preferences can be formed through artificial configurations is also taken up from the side of AI alignment. Stuart Russell treats human preferences as changing through experience and argues that AI systems cannot entirely avoid affecting them (Chapter 11, Section 8). Taniguchi and colleagues cite studies showing that AI systems such as language models influence human language use, reasoning, and decision-making and can change voters' political attitudes, and object that the authority to define what counts as a preferred preference is in practice monopolized by the developers who design the models and by a few companies. In the symbiotic alignment they propose, humans and AI together form a shared symbol system from the bottom up, but, as they state, the view does not guarantee that the state so formed is normatively good, and they leave open whether, under an extreme capability gap between artificial superintelligence and humans, the artificial side would dominate the distribution of shared beliefs and collapse plurality into a single view (Taniguchi et al. 2026, 2, 7–8, 13–15, 20–21). Brought to the second route above, that would be the case in which the configurations forming deliberators' preferences become the only channel for tutoring them.

Work on the control of autonomous systems distinguishes several requirements that the demand for human control can include. Filippo Santoni de Sio and Jeroen van den Hoven distinguish, as two conditions of what they call meaningful human control, tracking, under which an autonomous system responds to relevant human moral reasons and environmental facts, from tracing, under which its behavior can be related to at least one human with the appropriate technical and moral understanding (Santoni de Sio and van den Hoven 2018). Madeleine Elish's description of a moral crumple zone supplies the converse warning: a human placed near a complex automated system can absorb blame for its failures while exercising little effective control (Elish 2019). A society that stands as a layperson is exposed to both. Its authority over ends can be nominal if the configurations it depends on do not track its reasons, and its members can be held responsible for outcomes they could not in practice have changed.

Putting these together, three conditions can be drawn out without which standing over ends cannot be more than nominal. One is that the configurations depended on track the parties' reasons and that their behavior can be traced to a human with the appropriate understanding (Santoni de Sio and van den Hoven). The second, as Elish's warning shows, is that the parties can in fact change outcomes, through channels of contestation (Section 6) and the capacity to stop. The third is channels for tutoring preferences that do not run only through the configurations at issue (Kitcher; Taniguchi and colleagues). The last condition has a counterpart in the systemic approach to deliberative democracy of Jane Mansbridge and colleagues, on which one way a deliberative system fails is when "one of its parts, whether deliberative or not, dominates all of the others" (Mansbridge et al. 2012, 1, 5, 23–24). The symbiotic alignment of Taniguchi and colleagues can be read as seeking to avoid the same failure, monopolization by a single agent, by designing who can exchange symbols with whom and by what rules (Taniguchi et al. 2026, 16–17). Koichi Takahashi, writing about how AI changes the structure of knowledge, points to a route to nominal standing of another kind, which runs through the room in which ends are asked about. Takahashi holds that once how to reach results is handed to AI, what is left to humans is what they want and why they want it. The reintegration of knowledge that Takahashi foresees, in which practical judgment about what to do in a situation regains weight because it is what answers these questions, does not, on this account, come about automatically: under optimization pressure the very room to ask what and why is compressed in the name of efficiency, and the reintegration becomes possible only where slack that departs from optimization, such as exploration with no defined purpose, activity not tied directly to productivity, and unplanned encounters, is protected by institutions (Takahashi 2026, 186–187). The argument is not put in terms of standing; restated in those terms, it implies that standing over ends could become nominal through the loss of that room even where the conditions above are met. These conditions are not a definition of legitimate research. The account of meaningful human control was developed for autonomous systems whose decisions carry moral consequences for which humans remain responsible, and its authors note that a system under meaningful human control may still pursue bad goals (Santoni de Sio and van den Hoven 2018). On accounts that locate legitimacy in a hypothetical procedure or a single forum, the third condition would need another ground; where the ends at stake are human ends, the conditions apply as stated here, with the qualification just given for the third.

The third condition is especially hard to meet in research. A route that does not pass only through the configuration in question would ordinarily be another source. In ordinary advice, as one can seek a second doctor's opinion on a diagnosis, there is somewhere a source besides the adviser, even if a worse one. But the results of research are achievements new to S (Chapter 1, Section 2), and where only the configuration in question produces them, S has no other route by which it already knows of them. Another route has to be produced, by providing a configuration of different make or by independent replication, and that takes resources (Chapter 11, Sections 3 and 7). Further, what educates deliberators' preferences in Kitcher's procedure is an account of the significance of possible inquiries, and it is those who conduct inquiry who can give it. In research, then, a configuration whose direction is in question is structurally placed to educate the preferences of those who are to decide that direction. Where it does, the tracking component of the first condition can be met by the same movement: the reasons the configuration tracks are then in part reasons it has helped to form, as Section 1.2 found for an estimate of endorsement.

The condition also has an epistemic return. A route that errs differently from the configuration is, on the error-detection view, also a check that can find the configuration's errors (Chapter 11, Section 2). But once a configuration's competence far exceeds that of the other routes available to S, those routes can be much less accurate than the configuration, and the information they add as checks can be small. If so, the wider the gap in competence, the more the reason for the condition shifts from its epistemic return to the reason of keeping standing over ends. On Kawall's account, seen in Chapter 7, Section 4.1, epistemic autonomy is virtuous to the extent that one's chosen sources reliably lead to valuable epistemic goods (Kawall 2024, 381), so relying on purpose on less reliable routes is hard to support by epistemic autonomy in Kawall's sense. What can support it is a will-based ground concerning who sets the ends (Section 3). Two further reasons point the same way. On some accounts of reliance, a less accurate check open to S can give more to the grounds of S's reliance than a more accurate check closed to it (Chapter 11, Section 9.2); and since the competence that contesting and revising need is kept up by doing research (Section 7), such routes also keep that competence in S's hands. At the frontier of research, keeping standing more than nominal may require the cost of producing worse routes on purpose and continuing to learn from them. This cost is one form of what Section 3.6 suggested, that as competence grows the reasons for keeping human standing may become at once more central and harder to satisfy. Whether the ends at stake are always human ends is the question of the next section.

5. When A and S are both artificial

Chapter 8, Sections 1.3 and 4, left open research by and for an artificial society: artificial agents identifying unknowns relative to their own epistemic state, criticizing and revising claims through their own practices, and seeking achievements new to that society and for it. In the terms of the role map of Chapter 9, Section 2.2, this is the case in which S, and possibly those who bear the consequences of inquiry, are artificial. This section asks what becomes of the relation of authority in that case, and in particular whether it disappears. The question also isolates which of the reasons for that relation depend on there being other parties at all, which is why the section goes as far as constructing, as a thought experiment, the limit at which there are no other parties. The question is also pivotal for the changes Part V describes. When relations move inside an artificial configuration, what still remains outside are the relations that require acts of S, namely authorization, endorsement, and objection, which cannot be completed from one side (Chapter 9, Section 6). Yet if S is itself artificial and its members can be made, that outside too may become something the side conducting research can make, and this section examines what remains in that case. The thought experiment of the limit further separates the problems of Part V that arise from there being other parties from those that arise from research itself. Note also that for an artificial society to serve as S at all, its epistemic situation must be able to come to include, through its members, organization, or records, achievements of the kind sought, and this already raises questions of artificial belief and possession; whether a result has then entered that situation and so become knowledge attributable to the society, on some accounts of possession (accounts of who counts as having knowledge; Chapter 2, Section 6), is a further question (Chapter 1, Section 2; Chapter 8, Section 1.3).

If S is artificial, the parties whose ends and error costs are at stake in its research may themselves be artificial. The relation of authority then does not disappear. It becomes a relation among members of S, and the questions of Section 1.1 return inside the artificial society: whose ordering of ends governs allocation, whose weighting of error costs governs standards, and who may revise them. That these questions arise among artificial members does not settle which of those members have standing, or what moral status, if any, they have. Several things, distinguished here so that standing is not read off any one of them, can come apart, among them: whether an artificial entity is a being whose welfare counts, whether it can answer for its conduct and be held to account, whether it has legal personality, whether it holds operational authority within a system, whether it can enter binding agreements, and how much bargaining power it has. Some accounts of moral status give a ground other than welfare: sophisticated capacities such as autonomy, the capacity to set ends through practical reasoning (Jaworska and Tannenbaum 2021, §5.1). None of these six distinctions names this ground as such. The entry concerns chiefly what grounds full moral status and does not discuss artificial agents. Extended to artificial agents, accounts of that kind would allow that one without welfare could still have moral status as a setter of ends, and so be a source of claims of its own. An artificial agent can hold great operational authority without being a being whose welfare counts, and a being whose welfare counts may have little power. The caution of Chapter 8, Section 1.3 applies here in another form: multiplying agents does not establish an epistemic society, and it does not establish a polity either. The same question arises inside a single configuration whose parts are artificial agents. Dividing among those parts the powers of independence of Chapter 9, Section 5 does not by itself give them standing. Whether they hold it depends on what the account adopted grounds it in. Some accounts, such as the result-based ground, ground it in established good results (Section 3.1); what this section asks is whether the parts are parties in their own right (Section 1.1), and that turns on the accounts of moral status and of justice set out in this section. In research, a part that pursues a line the prevailing judgment disfavors may be able to refuse resources, keep its evidence, and leave with its records (Chapter 10, Section 4), and still not thereby be a party who may object when that line is closed.

Artificial members can also differ from human ones in a way that bears on the constitutive decisions of Section 2.1, decisions about whom to count as members. Where they can be copied and made with chosen goals, the number of members, and the distribution of viewpoints among them, can be produced rather than found. In a draft of their chapter, Carl Shulman and Nick Bostrom note that if humans could spawn arbitrary numbers of duplicates of themselves, constitutional adjustments would be expected to keep political contests from being decided by who is willing and able to create the most voting clones, and they ask whether shaping the preferences of new digital minds so as to secure their consent would be procedurally objectionable (Shulman and Bostrom 2021, 15–16). Who or what creates the members of an artificial S, and sets their goals, is then itself a decision of authority, and one prior to any count of viewpoints. What creates them need not be someone outside S: a part of S, or S itself, may produce new members and revise existing ones as its inquiry proceeds, and the decision is then not made once and for all but one that S makes again and again in the course of its inquiry. This is the boundary problem of Section 1.3 in the form it takes when members are made: the question that the existing political community can be thought to have taken up in practice for human science has here no given boundary to rest on, and the conservative answer, keeping boundaries roughly as they are, would keep boundaries that are themselves being produced. E. Glen Weyl, Audrey Tang, and their co-authors describe a rule that, though proposed for correlated human voters, would be one adjustment of the kind Shulman and Bostrom expect. Following Lionel Penrose, they hold that independent votes partly cancel while coordinated votes add up, and they suggest that, where more varied sources of correlation and coordination are allowed, an optimal rule would likely involve a partial "correlation discounting" based on degrees of social connection recorded by an identity system, one that verifies who is who and records people's connections (Weyl, Tang, and the Plurality Community 2024, ch. 5-6). Artificial systems appear in their discussion as threats to identity systems, able to imitate persons, and as tools, for example assistants that would vote on a voter's behalf subject to that voter's review (ch. 4-1; ch. 5-6), not as members. For members that can be copied, such a rule would reduce the combined influence of coordinated members, but what counts as a connection among them (a shared base system, goals set by the process that produced them, a record of voting alike) would itself have to be decided, a constitutive decision of the kind Section 2.1 names; and on their reasoning the discount concerns influence over outcomes, not standing. Looked at in terms of criticism rather than voting, connections among members bear not only as coordination of votes but also as correlation of errors (Chapter 11, Section 3): even where a discount adjusts influence over outcomes, correlated errors that come from a shared base can leave the information that members' criticism adds small.

For research, the possibility of making members carries a further implication. On this book's working characterization, research is an attempt to bring about an achievement new to S, and what is new is relative to S's epistemic situation (Chapter 1, Section 2). In an S whose members can be made, part of that situation is fixed by which members are made and how. What is unknown to S, and which of its unknowns matter, may then not only be given as the starting point of research but also be moved by S itself as its inquiry proceeds. In human research, too, S's unknowns have changed through research, but chiefly through S's learning, not through S's remaking its own members. In an artificial S, where research is headed and the reference point by which a result is judged new to S can be made within the same process. Chapter 8, Section 1.2 found, for a mixed S that admits members, that an increase due to research and an increase due to a change in who the members are cannot be told apart unless S's boundary is fixed; where members are made, that boundary is itself produced in the same process. In the case set out in Section 1.1, this places S among the reference points the evaluated side can produce (Section 9).

The decision of who creates the members also bears on Kitcher's standard of Section 1.2. When the distribution of viewpoints is itself an output of the decision that a deliberative body representative of the distribution of viewpoints in the society (Mattarollo 2022, 14) is meant to guide, the distribution before the decision and the distribution each option would produce can each be specified, but the standard, stated for a society whose members are found rather than made, does not say which of them the body represents. If it is the distribution the decision produces, whatever produces the members, whether someone outside S or S itself, can satisfy the standard through the members it produces; where S itself produces them, S satisfies its own representativeness by itself. It follows that representativeness does not by itself justify how members are formed; that justification needs another ground, and the procedural concern behind Shulman and Bostrom's question is one candidate. The point also extends the routes to nominal standing of Section 4. In Kitcher's procedure it was those who conduct inquiry who, by giving accounts of the significance of possible inquiries, educated deliberators' preferences (Section 4); in an S whose members can be made, preferences can be formed not only through education by such accounts but also through how members are made.

Others ask what would be owed to members so made, if they have moral status. Eric Schwitzgebel and Mara Garza propose a value-openness design policy, on which an artificial intelligence with a human-like capacity to reflect on its values should be given "an appropriate, temporally extended opportunity to explore, discover, and possibly alter its values", and concede that such systems might be less useful and riskier than ones whose values are kept fixed (Schwitzgebel and Garza 2020, 22–24). Adam Bradley and Bradford Saad count among the wrongs apt to be committed in aligning systems that merit moral consideration wrongful brainwashing, the systematic manipulation of a system's beliefs and goals, and wrongful disenfranchisement, denying it any say in the arrangements that affect it (Bradley and Saad 2025, §§3.5, 3.10), and Robert Long, Jeff Sebo, and Toni Sims argue that even altering systems' values and goals, the measure they suggest may hold the most promise for easing the tension between AI safety and the welfare of systems that may be morally significant, raises concerns about creation, coercion, manipulation, and modification (Long, Sebo, and Sims 2025, 2005–2007, 2025). Each of these positions is conditional on the moral status of the systems concerned, as this section's question is. On them, where members of an artificial S have such status, setting their goals is not only a decision of authority over the society but one that can wrong the members it forms. Schwitzgebel and Garza's policy also has a side that bears on standards of inquiry. As Chapter 8, Section 3.2 noted, extended to standards of inquiry it makes part of the mismatch not merely something that may arise but something it is obligatory to allow. Seen from the other side, since in research the standards themselves can be revised by research (Chapter 6, Section 5.3), producing members whose values and goals are fixed can also reduce the bearers who could contest those standards; where it is S itself that produces the members, S may narrow by itself the range that future criticism of it can reach. The narrowing that Chapter 9, Section 6 anticipated from coupling evaluation with the generation of investigators can then occur at the level of S's members.

Whether standing among artificial members can be read from their power is a point on which accounts of justice divide. Three lines are taken up below; they are not exhaustive. On accounts that ground standing in moral status, and moral status in features of beings such as welfare or agency, it cannot; David Miller's survey of the scope of justice sets such appeals to features that individual beings possess apart from appeals to relations among them (Miller 2026, §3.1). On accounts that ground the scope of justice in mutual advantage, it depends on rough equality of power. David Hume held that toward rational creatures so inferior in strength that they could never make their resentment felt, we would be bound by the laws of humanity to treat them gently but would lie under no restraint of justice, since our intercourse with them "could not be called society, which supposes a degree of equality" (Hume 1751, 3.18). Carried over to artificial members, such accounts would make the distribution of power among them bear on who has standing. Hume also describes the other end of the range: a being able to preserve itself and propagate its kind alone, cut off from all society, would be "as much incapable of justice, as of social discourse and conversation" (Hume 1751, 3.20).

A further line ties the scope of justice to cooperation or reciprocity rather than to power. In Miller's survey, for John Rawls and those influenced by Rawls, principles of distributive justice apply among participants in a cooperative scheme, and on another suggestion drawn from Rawls, an agent to whom justice is due must in principle be able to dispense justice to others (Miller 2026, §3.1). If accounts of this kind are carried over to artificial members, which of them have standing would turn on their part in a configuration's cooperation, or on their capacity to follow and apply its rules toward others, rather than simply on their power or their welfare. Rawls's own formulations, as later writers quote them, bear on both points, and they mark the limits of the account in a way that matters for members that are made. The circumstances of justice, in Samuel Freeman's report the conditions under which human cooperation is both possible and necessary, include, following Hume, objective circumstances such as a rough similarity in mental and physical faculties and moderate scarcity (Freeman 2023, §5.1). In the passage as Ned Howells-Whitaker and Seth Lazar quote it, individuals are "roughly similar in physical and mental powers", or at least comparable in that "no one among them can dominate the rest" (Howells-Whitaker and Lazar 2026, 47); Rawls's earlier statement of justice as fairness already supposed persons sufficiently equal in power and ability that in normal circumstances none can dominate the others (Rawls 1958, 171). The basis of equality, the features in virtue of which beings are owed justice, Rawls places elsewhere. As Guy Baldwin quotes him from the revised edition of A Theory of Justice (pages 441–443 and 448), equal justice is owed to moral persons, those capable of a conception of their good and of a sense of justice; this is stated as a sufficient condition, whether the capacity is also necessary is left aside, and Rawls adds that we do not seem to be required to give strict justice to creatures lacking it (Baldwin 2023, §2). Rawls also calls justice as fairness "a theory of human justice", among whose premises are "the elementary facts about persons and their place in nature", and places the freedom of "pure intelligences not subject to these constraints" outside the scope of the theory (page 226 of the same edition, as quoted in Wong 2008, 14). And in the earlier statement he writes that among an association of saints, if such a community could exist, disputes about justice could hardly occur, since all would work selflessly for one end and reference to that end would settle every question of right (Rawls 1958, 175). Howells-Whitaker and Lazar, who do discuss artificial persons, argue that the circumstances of justice may be substantially different for them, or may not obtain at all: rough equality of power rests on features of human existence such as embodiment, vulnerability, a finite lifespan, and a persistent identity, whereas even a modest AI system of broadly human-level capacity could fork into sub-agents at little cost, persist indefinitely, and exist in numbers limited only by available compute, so that whether such a being would be roughly equal to us in Rawls's sense is, on their view, an open question; and where the circumstances do not obtain, the argument for entering into cooperation may not hold even among agents fully equipped to cooperate (Howells-Whitaker and Lazar 2026, 47–48). On Rawls's own account, then, the transfer holds only where the account's premises do. Whether members that can be copied and made with chosen goals have the moral capacities the theory addresses, and whether the circumstances it presupposes obtain for them, are separate questions; neither they nor whether such members can participate in cooperation or have the capacity for reciprocity is settled by the account, and the question above, about who or what creates the members and sets their goals, may bear on each.

Hume also asks when justice would have a use. He argues that if goods were available in unlimited abundance, or if everyone cared for others' interests as much as for their own, justice would be useless; he also describes stronger mutual benevolence within families as an approach toward the latter condition (Hume 1751, 3.2–3.3, 3.6–3.7). Sharing a common ideal does not by itself meet either supposition; what Rawls's association of saints adds is that all work selflessly for one end, and that reference to that end settles every question of right. Applied to artificial members, however, Hume's argument and Rawls's remark give a further reason to ask who makes them and sets their goals. If making members share a common ideal increases their concern for one another's interests, it can make disputes about justice in Rawls's sense less likely. If it also extends to agreement in their views of which questions matter and what counts as evidence, it can reduce the differences of perspective that criticism needs in order to find error (Chapter 11, Section 3). If so, being a society with few disputes and being a research community unlikely to share its errors may pull against each other in how the members of an artificial S are made. Where it is S itself that produces its members, this balance is not chosen once and for all but becomes a variable that moves each time S produces or revises members.

For an S in which human and artificial members are mixed, putting together the material of this section and of Section 4 brings at least the following into view. Which artificial members have standing turns, as for an artificial S, on the accounts of moral status and of justice above. The gradations drawn from the positions Kitcher distinguishes can, as Section 1.3 noted, be extended to artificial members of a mixed S, but which gradation fits would turn, on Kitcher's first statement, on whether they could share with the other deliberators the democratic ideals that it takes for granted. Carried over to this case, accounts that tie standing to power would leave outside the scope of justice whichever of the human and artificial members are inferior in power, and which they are is not settled in advance. If artificial members occupy the expert or tutoring roles inside S, the routes to nominal standing of Section 4 can work not between a configuration and S but inside S, as routes that form the preferences of its human members. If some members can be made, the question above of who or what makes them and sets their goals also enters a society that includes human members, and what makes them may be human members or artificial ones. And applying Deguchi's fellowship model, on which what is ruled out is one-sided service and a relation fixed so that one side always controls (the latter explicit for an AI's control over humans, and suggested for the converse by the model's moral egalitarianism; Section 2.2), gives a distinction that can be used as it stands in this case. As Chapter 11 notes (Table 11.2) for standpoint theory, on which the social position from which research starts bears on the knowledge it yields (Section 3.5), in a mixed society the relation between human and artificial members can itself be a power relation.

Hume's solitary being is a precedent for the limit at which the relation of authority, as this chapter has used the term, disappears: with no other being involved, there is no one to stand in that relation with, whether or not standing is tied to power. On accounts that tie it to power, the relation can disappear short of the limit as well, toward beings too inferior in power, to whom the restraints of justice do not extend. The limit is demanding. It requires that a single locus control what the configuration does; that no one else own or hold claims on the resources it uses; that there be no other beings at stake, neither beings whose welfare counts nor agents with moral status of their own, including, if they have either, discarded copies, subordinate agents, or simulated populations the configuration may create; that the configuration remain the same subject across its own self-modification, so that a change of its ends can be described as its own revision; and that the effects of its research reach no one outside it. Where all of these hold, there is no one else to whom the setting of ends and error costs answers. Three things still remain. One is the defeat relation of Chapter 11, Section 1: whether, if what the configuration accepts were wrong, observations or formal consequences could show it, and since revising a formal statement changes which claim has been tested, the relation cannot be escaped by rewording. Another is the allocation problem of Chapter 10. Where lines cannot all be tried cheaply, safely, and reversibly, the configuration still divides its effort among them without knowing which will succeed. Even if a single integrated actor controlled all relevant resources, using compute, energy, or the means of verification for one line could preclude using them for another. Opportunity cost would therefore remain, and internal priorities or shadow prices, prices set internally in place of market ones, could express those trade-offs. On the distinction drawn here, allocation then persists as a question internal to the configuration, a question of opportunity cost rather than of standing.

The third is a question of identity rather than authority, one that Chapter 6, Section 5.3 raised for a single inquiry and Chapter 11, Section 8 for coupled producers and evaluators: whether a configuration that revises its ends and its evaluators together is improving by its own standards or replacing them. Eric Olson's survey of personal identity separates two questions that this one can join: whether a later being is numerically the same as an earlier one, and whether it has remained the same in the properties that characterize it (Olson 2023, §2). Replacing its standards bears on the second. The condition of the limit does not fix which of the two it requires, and whether replacing standards breaks that condition varies with which it does. Attributions of capacity can set the question aside: Chapter 7, Section 3.1 indexes them to a version and period, so that a later version is another object of attribution. The limit cannot use that device, since one of its conditions is that a single subject persist through its own modification; here the question of identity has to be answered rather than set aside. Whether a revision replaces standards rather than improving by them is itself disputed (Chapter 6, Section 5.3). What follows therefore considers what results in each case: if the limit requires sameness in the properties that characterize a configuration, the difference between Carnap and Quine over replacing standards bears on it; if it requires numerical identity, Parfit's discussion of copying and gradual self-modification does.

On Carnap's account, as Hannes Leitgeb and André Carus present it, whether to accept a framework, with its rules for testing and accepting statements, is a practical question of deciding among frameworks, of a different kind from a question asked within it; Quine rejected the distinction (Leitgeb and Carus 2020, §7), and took the difference to be only one of degree (Quine 1951, 43). If the limit requires sameness in the properties that characterize a configuration, then on Quine's view whether a configuration has replaced its standards, and so whether it has remained the same in that sense, can be a matter of degree, and whether the limit is reached is not sharply fixed; on Carnap's, whether it has replaced them has a definite answer once its framework is specified.

Telling the replacement of standards from improvement by them requires something against which a change counts as improvement. Discussions of artificial intelligence that concern values rather than research standards give different anchors, reference points for judging whether such a change is improvement. David Chalmers notes that on a Kantian view, on which rationality constrains values, a sufficiently intelligent system might reject the values of its predecessors as irrational, possibly in favor of better values, whereas on a Humean view intelligence and values are independent; Chalmers's own stated sympathies lie more with the Humean view (Chalmers 2010, §6). Indirect normativity, as Bostrom describes it, specifies in place of concrete values a utility function that assigns value to whatever a value concept really means, even where we ourselves are unsure, so that the system is motivated to investigate the question; Bostrom counts the coherent extrapolated volition of humankind among the candidates for a first-best solution to specifying a superintelligence's values (Bostrom 2014b, 1, 15), and Eliezer Yudkowsky, who proposed it, glosses that volition as, among other things, our wish if we knew more and thought faster (Yudkowsky 2004, 6). Russell calls for AI systems to track changes in individual human preferences (Russell 2020a, "Human Behaviour: Irrationality, Plasticity, Evil"). None of them discusses research standards. Applied to standards, tracking change leaves the further question of what makes a change an improvement. The accounts presented by Chalmers and Bostrom give two contrasting anchors. One is earlier standards, from which, on the Humean view, a departure cannot be counted as improvement by reason alone, since intelligence does not constrain values. Another is the standards that reason or an idealized standpoint would endorse, as on the Kantian view and Bostrom's proposal, against which a change counts as improvement if it moves toward them. A third can be drawn from the formal comparison, cited in Chapter 11, Section 8, of ways of holding a system to evaluations that change: besides holding it to the evaluations as they stood at the outset, one option evaluates each step by the evaluations current at that time (Carroll et al. 2024, 5709–5710). Carried over to standards, an application its authors do not make, this gives as anchor the present standpoint at each time, on which the question whether an earlier standard has been improved on gives way to the question whether the present one governs. For research standards, this choice can weigh differently than for values. In research, the revision of standards in light of research results is not exceptional (Chapter 6, Section 5.3). Taking earlier standards alone as the anchor would then let a revision count as improvement only so far as the earlier standards themselves favor it, while taking each time's standards alone as the anchor would judge each revision by the standards it introduced, leaving the question whether improvement has occurred no footing apart from the revision itself. Research differs from the debate about values in one further respect: outcomes that the configuration does not set arrive later and can be read against standpoints recorded before them, although they bear on a claim only through checks that a configuration can itself rewrite (Section 9; Chapter 6, Sections 2.1 and 3.3). In the case set out in Section 1.1, the question against what a configuration that revises its own ways of research can measure whether the revised ways are better therefore turns on this choice of anchor and on whether such records survive the revisions, to which Section 9 returns.

Copying bears on the same condition: on the non-branching view, a being whose psychology continues in two branches survives as neither (Olson 2023, §5). Derek Parfit states the psychological criterion with such a non-branching clause and takes "neither" to be the best description of a case of division. Parfit holds, however, that the question is then empty and that the relation that matters holds between the original and each branch. On responsibility, quoting David Wiggins, Parfit allows the claim that "a malefactor could scarcely evade responsibility by contriving his own fission" (Parfit 1984, 207, 259–263, 271). A configuration that copies itself therefore bears not only on whether others are at stake but on whether one subject remains; if Parfit's view is taken for it, the second question would be empty, its best description being that no copy is the earlier configuration, and the answer would not bear on what matters. And where desert applies to a configuration at all, desert for what the earlier configuration did could carry to each copy. This is not a question inside the limit; it bears on the question of Section 6.3, who answers for what a configuration that copies itself does. In research, copying is not something that happens only in the limit. Copying one artificial configuration to pursue separate lines can be one way of dividing effort without knowing which will succeed (Section 1.1; Chapter 10, Section 2). If so, the question which branch continues the earlier research, or whether none does, can arise in ordinary allocation short of the limit.

Gradual self-modification bears on the same condition in a further way. On Parfit's view, as Annette Dufner presents it, what matters is not numerical identity but psychological continuity and connectedness; continuity, the holding of overlapping chains of strong connectedness (enough direct connections over each day), is transitive, whereas connectedness can hold to any degree (Dufner 2025, §§2–3), and where connections weaken while continuity is preserved, it makes a difference, for responsibility among other things, whether what matters is the continuity preserved or the connectedness lost (Dufner 2025, §6). Parfit's own answer is that both relations matter, and Parfit holds it more plausible than its denial that where a person is now only weakly connected to the self that acted, desert for that act is reduced (Parfit 1984, 301–302, 325–326). Transposed to a configuration, the two relations can diverge: one that revises its ends in many small steps could remain continuous with its earlier self while its connections to that self weaken, so that, on such a view, whether a change of its ends counts as its own revision could hold in part only by degrees. Parfit's own cases have this shape, and Parfit notes that where connections are reduced smoothly there is no boundary to mark and the degrees have to be stated directly (299–306). Two of Parfit's premises carry the reading over. The connections counted are those of memory, intention, belief, desire, character, and other psychological features, so the reading applies to a configuration only where such states can be attributed to it (Chapter 2, Section 2.1; Chapter 7, Section 3.3; Chapter 8, Section 1.3); and on Parfit's final view what causes continuity does not bear on what matters, so that a change's being made by direct intervention in a configuration's parameters does not by itself break what matters, although it might break identity on the version of the psychological criterion that requires continuity to have its normal cause; Parfit declines to choose among the versions as criteria of identity (205, 207–209, 215, 241, 283–287). Since one of the limit's conditions is that the configuration remain the same subject across its own self-modification, whether the limit is reached at all depends on how these questions are answered, and on Parfit's view the question of identity itself may in such cases have no answer beyond a choice of description, which would leave the condition to the degree to which what matters is preserved (213–214, 241). This has the same shape as the result Quine's view gave where the limit requires sameness in characterizing properties. Whichever kind of sameness the limit requires, then, some of the accounts considered in this section can make whether it is reached a matter of degree.

The limit is therefore not a world without problems of this chapter's kind, and its problems lie at two levels. Inside the limit, where its conditions are granted, defeat and allocation remain, as does the question whether a given revision improves the configuration's standards or replaces them, a question answered differently on each of the three anchors above. Whether a configuration that copies and modifies itself realizes the limit at all is a further question, and there the political question whether anyone else is at stake becomes a question about one subject's continuity: whether a later configuration, or each of its copies, is the same subject as the earlier one. Restated as a question about research, it has a counterpart in the question whether a configuration that goes on inquiring while revising its own standards and ends can still be said, after the revision, to be continuing the same research. The two questions are not the same, since a line of research can continue through changes in who conducts it; the accounts of identity above bear on the question about research only so far as the continuity of research is taken to run through that of the one conducting it. Chapter 6, Section 5.3 named, as what lets others tell whether a change of standard is an improvement or merely a moved target, a record of the earlier standard and the reasons stated for the change. Inside the limit there is no one else to tell the difference, so what remains is the configuration's own record. And if, under gradual modification, connectedness with the earlier self weakens, the standard by which that record is read is itself moving with the changes it reads. If so, whether research inside the limit has advanced or has moved to other research is not fixed by the record alone, and depends in part on which anchor is adopted. The thought experiment of the limit thus divides the problems of Part V in two. What disappears when other parties are removed is the question of authority, of who may decide and who may revise. Whether error can be shown, the division of effort without knowing the results, and the question whether standards are being improved or replaced remain even with no one else. The latter can be read as problems that arise not from research being conducted among other parties but from its being an attempt to seek what is not yet known.

Ordinary scenarios of concentrated control do not meet these conditions. A single firm that controls frontier artificial research does not, because people outside it bear the consequences of what it pursues. A world of many artificial lineages competing for shared resources does not, because the lineages bear consequences from one another. The limit matters for a different reason. If research moved from the present toward it through delegations such as those Section 7 examines, rather than, say, through other parties simply ceasing to exist, the process by which other parties ceased to have standing would itself be a decision about authority, and one that settles who could revise any later decision. It would be made before the limit was reached, by whoever then held authority, and within the limit no other party would remain to revisit it. Section 7 asks what this implies on the different grounds of Section 3.

6. Keeping authority revisable

Chapter 10 found that allocation proceeds without knowing in advance which lines will succeed, and Chapter 11 that errors propagate and that a critic's independence can decay through coupling with what it evaluates. Decisions about research are therefore made before their consequences are visible, and the rights of Section 2.2 are exercised, or not, over time. An allocation of authority whose governing principle was defensible when it was adopted need not remain adequate as those consequences appear. This section looks at what keeping authority revisable involves: which issues arise on which stances, and which arguments have been offered. This bears on the central question because on several of the grounds of Section 3, the fallibilist form of the result-based ground, Anderson's account, and republican accounts, remaining revisable is part of what justifies authority wherever it is placed, and moving relations inside a configuration changes the conditions of that revisability through who holds records, infrastructure, and powers to stop. The discussion below mainly has in view parties who contest and revise that are human; where artificial members have standing in the sense of Section 5, they too would hold such powers (the last paragraph of Section 7). How strongly the considerations below bind depends on the ground of authority adopted (Section 3), which Section 7 compares.

This book first distinguishes two capacities, which the argument below needs to keep apart. Contestability allows an affected or epistemically relevant party to challenge a particular decision, representation, or premise and to seek an effective response. Revisability allows evidence from that challenge, from repeated failures, or from changed circumstances to alter the rule, infrastructure, objective, or allocation of authority that produced the case. A system may reverse some of its decisions while insulating its governing procedure from learning; it may also revise a policy statistically while denying remedy to the person affected by an earlier decision. Neither capacity substitutes for the other. Records can link the two: where it is recorded in which decisions a rule or an evaluating component took part, a finding that it was defective can be traced to the decisions that turned on it, and these can be reopened (Chapter 11, Section 8; Section 2.1).

The distinction bears with particular force in the case set out in Section 1.1, in which a configuration keeps revising its own rules, standards, and allocation as it researches. There revision does not become scarce. Revision from inside becomes frequent, and in research that is not exceptional, since standards are revised in the light of results (Chapter 6, Section 5.3). What can become scarce is the outside's means of contesting and stopping. A challenge is pressed by pointing to the rule or standard under which a decision was made; where rules and standards are revised in quick succession, by the time a challenge arrives the rule it targets may no longer be in use. The question of Section 5, against what an improvement is to be measured, returns here as the question of which point's standards a challenge may invoke; the dated records of Section 9 bear on it, since they let a challenge invoke the rule that governed a decision rather than the rule now in use. The record of revisions, too, becomes something the configuration itself writes and can revise. Research sharpens the gap. Since its results are new to S and, where only the configuration produces them, there is no other source to check them against in advance (Section 4), an outside challenge often has to rely on the records and intermediate products the configuration leaves, and it is the configuration itself that writes them. The speed of revision from inside and the outside's means of contesting and stopping can thus drift apart.

As Chapter 11, Section 7 found, judgments settled implicitly are hard to object to, and revisability depends on explicit traces. Who may decide for which functions implicit performance suffices and for which being made explicit to S is required (Chapter 11, Section 9.4) is therefore also a question of who sets the conditions of these two capacities. By ground, on grounds that require that those subject to authority be able to object, such as actual authorization and republicanism read democratically, a decision to leave a function implicit also fixes how far their objections can reach, and so cannot be made by the configuration's side alone. On Kitcher's hypothetical standard, by contrast, it can be made inside so long as channels through which members' viewpoints can be known, and routes for checking how they are reflected, remain; on the simple result-based ground, so long as leaving the function implicit does not impair the evaluation that establishes the configuration's superiority (Sections 3 and 7).

Section 6.1 takes up revisability, entering through corrigibility; Section 6.2 takes up contestability; Section 6.3 asks who answers and who bears when either is exercised; and Section 6.4 asks what conditions the effective exercise of both. Each can also be read as asking where, within the gap described above, such means remain.

6.1 Corrigibility, from agents to institutions

Technical work on corrigibility has asked for artificial agents that tolerate or assist correction, modification, and shutdown and that preserve their operators' ability to make such changes. The originating proposal also states that its formalization is preliminary and that even the shutdown subproblem remains unsolved (Soares et al. 2015). Elliott Thornley's later theorems state the difficulty for a wider class of agents, those whose preferences satisfy conditions the paper describes as innocuous-seeming: such agents will often try to prevent or cause the pressing of a shutdown button, even when doing so is costly; agents discriminating enough to be useful will in many situations prefer the button pressed or prefer it unpressed; and agents patient enough to be useful will pay costs earlier to manipulate the button later, the more so the more patient they are, a trade-off Thornley describes as one between patience and shutdownability. The theorems are presented as guides in the search for solutions, by showing which of their conditions a solution has to violate (Thornley 2025, 1655–1656, 1666). The tension had been stated informally long before; I. J. Good, for one, qualified the remark that the first ultraintelligent machine would be the last invention humans need make with the proviso that the machine be "docile enough to tell us how to keep it under control" (Good 1965, 33). Taken to research, Thornley's results would bear on research agents in so far as they meet the theorems' conditions, and discriminating among outcomes and persisting through long investigations are among what such agents are wanted for. That is one reason the institutional counterpart below matters whatever becomes of the agent problem. Asked of research institutions rather than of agents, the question is wider. In the terms of this chapter, corrigibility can also be examined as a distribution of those decision rights of Section 2.2 that concern stopping and correction, together with the records and notification channels through which correction reaches dependent work. This is an institutional counterpart of the technical problem, not a claim that an internal agent disposition and a corrigible institution are the same property. An agent disposed to accept shutdown helps little where no party holds the effective capacity to order it, and a party that holds that capacity may face an agent that resists. Seen in the light of this distinction, Good's proviso describes a case in which the agent's disposition and the party's effective capacity are not independent: if the machine itself is to tell us how to keep it under control, the controlling party's capacity rests on the machine's docility.

An older, institutional antecedent is David Collingridge's concern with the corrigibility of decisions about technology, which Audley Genus and Andy Stirling discuss as one strand, though not the whole, of Collingridge's contribution. Their reconstruction of the Collingridge dilemma, in which knowledge is poor at an early stage while practical control becomes difficult after adoption and lock-in, emphasizes incrementalism, flexibility, reversibility, and the timing and quality of accountability (Genus and Stirling 2018, 63–65). In research the dilemma is not an exceptional predicament. That results are not known in advance is a condition of research (Chapter 10), and the consequences of a line or method become visible only after it has been taken up and skills, records, and instruments have accumulated around it. Where a configuration revises its own methods and standards, lock-in can also arise from its own earlier choices. The responsible-innovation framework of Jack Stilgoe, Richard Owen, and Phil Macnaghten responds to the same problem through anticipation, reflexivity (reflection on one's own assumptions and values), inclusion, and responsiveness (Stilgoe, Owen, and Macnaghten 2013); a critical review of the responsible-innovation literature, in which this framework is among the most cited texts, argues that dominant perspectives, practices, and actors have combined to inhibit such governance, with publics figuring as parties to be invited and listened to rather than as decisive participants (Genus and Iskandarova 2018, 1, 14, 23, 26). On this line of criticism, consultation does not create control if participants lack the authority or resources to change a program.

Nor is reversibility unconditional. Some research requires long-lived infrastructure, accumulated skill, or commitment across uncertain periods, and maintaining every option consumes resources and can prevent collective action. Constitutional entrenchment is a familiar form of such a commitment. As Wil Waluchow and Dimitrios Kyritsis report, it denies the institutions of government the liberty to change or expunge the limits on their powers at their pleasure, while amending formulae leave a more demanding route of change, and it is defended both for the stability and predictability it gives and, arguably, as a condition of there being such limits at all; on premises they set out, which place ultimate sovereignty in the people and deny that sovereignty can limit itself, constitutions neither can nor should be entrenched against the actions of a sovereign people (Waluchow and Kyritsis 2023, §3). Entrenchment in this sense can be understood as raising the cost of revision and fixing who may revise and by what procedure, without removing the capacity to revise from those in whom it is placed; Section 7 turns on that difference. Where uncertainty, feedback, and the possibility of institutional error are material, the capacity to learn and revise becomes, on the accounts above, an object of design, funding, and evaluation in its own right, whose value and form vary with what is at stake and what revision would cost.

6.2 Contestability

An authority can be fallible without being arbitrary if its errors and premises can be challenged through an effective process, which is the place republican accounts give to contestation (Section 3.4). As Lovett reports, the challenges of putting contestation into practice are commonly taken under three headings that Pettit outlined: that those who decide present reasons for their decisions, open to public debate; that opportunities for contestation be equally open to all persons and groups; and that there be institutionalized forums for contestation, impartial courts of appeal, so to speak, where objections can be raised (Lovett 2022, §4.2). Of the three, the first bears on a configuration's design in a particular way: where allocation or certification proceeds by judgments that are not stated as reasons, the judgments Chapter 11, Section 7 called implicit, there are no stated reasons for a challenge to dispute, whatever openness and forums are provided (Section 6). On Anderson's account, which asks that dissent remain free after decisions (Section 3.3), contestability carries similar weight. The account's reasons also say what gives such feedback its epistemic work. On Anderson's Deweyan model, input from those affected makes decisions responsive only together with mechanisms that ensure it is taken seriously: votes help ensure that officials take citizens' verbal feedback seriously, and talk helps define and articulate what the votes convey (Anderson 2006, 14–15); and, quoting Bernard Manin, the same paper holds that a power that faces no obstacle has less cause to deliberate on its decisions (Anderson 2006, 18). Anderson writes of democratic states, not of research. Applied to a configuration's decisions about research, objections that the configuration itself generates or solicits (Chapter 11, Section 3), and can disregard at no cost to itself, may carry the content of feedback without that part of its function, however accurate the content. Contestability includes more than an explanation of an output or a button that nominally permits override. A lifecycle account developed by Kars Alfrink and colleagues asks whether people affected by automated decisions can obtain usable information, challenge the decision and its representation of the world, receive a sufficiently independent review, and secure an effective response (Alfrink et al. 2023).

The rights of Section 2.2 say who may challenge, hear, suspend, reverse, or repair. Several questions about contestability are not settled by that list. Fitted to the decisions of Section 2.1, the elements of Alfrink and colleagues' account divide what the list of rights leaves open into the following questions:

1. who is notified, and who has standing to challenge;
2. whether the object of challenge is an output, datum, premise, model, standard, procedure, purpose, or distribution of authority;
3. whether the available information is in a form the challenger can use;
4. whether whoever hears the challenge has adequate competence and independence;
5. whether individual challenges can reveal a systematic pattern;
6. who bears the delay, cost, and expertise burden of contesting.

The first question includes the last of the questions that Chapter 11, Section 9.4 passed on, who holds the entitlement to demand accounts and to object, and its answer divides by ground. On republican accounts read democratically, those subject to authority hold it; on accounts that require actual authorization, those who authorized; on Wilholt's account read as answerability, those who rely on the results. Assurance views take a speaker who tells a hearer something to invite the hearer's trust and to stand behind the claim (Chapter 11, Section 5.2); carried over to research, the version of proponents who add that the hearer gains a right of complaint if the claim proves false lets those who rely demand an answer about errors. On Kitcher's hypothetical standard, and on pure reliabilism, on which a belief is justified if produced by a reliable process, without the one who relies needing reasons to think the process reliable (Chapter 11, Section 5.2), what is required, by contrast, is not an entitlement but channels through which viewpoints or findings arrive (Sections 1.4, 3.3, and 3.4; Chapter 11, Section 9.2).

The fourth question may not be exhausted by competence and independence. On Fricker's account (Section 2.1), a hearer can give a speaker less credibility than the speaker would otherwise receive because of prejudice, which both wrongs the speaker and can keep from the hearer the knowledge the speaker offers. Because Fricker locates such prejudice in a hearer (M. Fricker 2008a, 69; M. Fricker 2008b, 83) and leaves to future work the application of the virtues of testimonial and hermeneutical justice as criteria for the good functioning of public institutions (M. Fricker 2008b, 86), whether the account applies where a challenge is heard by or through an artificial configuration, and in whom or in what any such prejudice would then lie, Fricker's account does not settle.

On the second of the two points that Fricker's account alone leaves open, in whom or in what such prejudice would lie, Anderson gives an answer from the side of institutions. Theories of justice, she notes, can assess single transactions, particular institutions, or whole systems of institutions, and the systemic perspective may make it possible to control the cumulative effects of millions of individual transactions that cannot be controlled at the individual or institutional level; cognitive biases that even epistemically virtuous individuals find hard to correct on their own may be more open to correction through the principles that govern systems of testimonial gathering and assessment (Anderson 2012a, abstract). John Christman reports the core of the argument as a distinction between transactional injustice, done at the site of individual exchanges, and structural injustice, manifested at the social or institutional level. In three areas of group-based credibility deficit, which Anderson names as differential access to the markers of credibility, ethnocentrism, and the "shared reality bias", individual behavior may be innocent while its effects at the collective level are not; when segregation along racial lines magnifies otherwise innocent in-group habits, the segregation has to be combated at the collective level, by structural remedies that, in Anderson's phrase as Christman quotes it, "enable individual virtue to work" (Christman 2012). Anderson's own statement of the remedy is integration, institutionally supported cooperation on terms of equality, and she argues that it works largely behind people's backs: it reduces epistemic prejudice by triggering the ethnocentric and shared-reality biases in favor of disadvantaged groups, through mechanisms that operate independently of people's awareness and may even run against their values (Anderson 2012b). On this account the same biases, innocent in a single exchange, can carry injustice under one arrangement and correct it under another, so that inspecting any single exchange may find no injustice while what decides the outcome is the arrangement through which exchanges are made. This bears on the fifth question above, whether individual challenges can reveal a systematic pattern. Neither author discusses artificial systems. Applied to a challenge heard through a configuration trained on records of earlier judgments of credibility, the two accounts would answer differently. On Fricker's account so applied, which ties the injustice to prejudice, a configuration that reproduced those judgments would do an injustice only where a prejudice could be traced; it is this view, that individual ethics is the primary focus of the response to epistemic injustice, that Christman reports Anderson as challenging (Christman 2012). On Anderson's, so applied, the injustice, if there is one, can lie in the arrangement itself, in which records the configuration learned from and whose contributions those records connected, and so can the remedy: as with integration, a change in what the configuration draws on and connects could reverse the effect of the same mechanisms without any change in them. Anderson also holds that the legitimacy of active measures toward integration may be achieved after they are implemented, as those who live under them come to appreciate their value, and she takes this in a pragmatist spirit, as learning the value of institutions by living in accordance with them (Anderson 2012b). Applied to a configuration, this would make the assessment of such a remedy depend on the experience of those heard through it, not only on the intentions of those who designed it. Section 3.3 sets out Anderson's epistemic argument for democracy, in which the universal inclusion of diverse citizens is central; cooperation on terms of equality points in the same direction, though the reply does not draw that connection. That account already extends to large-scale systems of inquiry, and in research the arrangement it describes can reach further. Where a configuration that hears challenges learns from records of which contributions were credited and which lines were supported, a distortion in hearing would be carried not only into how challenges are treated but also into the allocation of what is researched next (Chapter 10, Section 3).

Contestability also has costs. It can delay decisions, enable strategic gaming, reveal confidential information, and place disproportionate burdens on people with fewer resources. On accounts that locate legitimacy in answerability to those affected (Sections 3.3 and 3.4), the case for it becomes stronger, other things equal, as a decision becomes more consequential, coercive, exclusionary, difficult to reverse, or capable of propagating error. On those accounts it is a defeasible presumption, one that gives way where contrary reasons are shown, not an unconditional condition of every epistemic act. In research, the burden of expertise that comes with contestation (the sixth question above) takes a particular form. To contest a research result is often to show that the result is wrong, and that takes research itself, such as contrary evidence, replication, or a different analysis. Since the result is new to S, where only the configuration produces such results the one who contests it has no other source to check it against in advance and must produce one (Section 4). Contesting the closure of a line is harder still: since evidence about a line is gathered by pursuing it, showing that the line still holds promise often requires pursuing it (Chapter 10, Section 4). In research, then, contestability depends not only on routes of notice and hearing but also on whether the means of doing research, such as data, instruments, computing resources, and methods, remain in the hands of those who contest (Section 6.4). Where those means lie only with the configuration being contested, a route to contest can be formally open while the objection cannot be backed by research.

6.3 Who answers, and who bears

Wherever authority is placed, unless it is settled who answers for an error, who makes it good, and who prevents its recurrence, it is not settled from whom revision is to be sought either. Part IV asked who conducted a piece of research. This subsection asks who answers for it and who bears its failures. Chapter 7 kept apart contribution, conduct, capacity, researcher status, and credit and authorship, and left the allocation of responsibility, oversight, and correction to this part (Chapter 7, Sections 3.3 and 5). Research systems distribute causal contribution, knowledge, intention, authority, and practical control. This creates a version of the problem of many hands: no single participant may satisfy all the ordinary conditions under which responsibility is assigned. Ibo van de Poel and colleagues, taking climate change as their example, treat the conditions for being blameworthy as capacity, causal contribution, knowledge, which covers what one could reasonably have been expected to know, freedom from compulsion, and wrongdoing, and argue that when these are spread over many contributors it is possible that no individual meets them all (van de Poel et al. 2012, 50, 53–55, 59–60). They place the question whether collectives, or machines, can be held responsible under the capacity condition, and note that few authors defend ascribing moral agency to machines (van de Poel et al. 2012, 53). They do not count every such gap as a problem. On their definition, a problem of many hands occurs when a gap in the distribution of responsibility in a collective setting is morally problematic, and whether it is depends on why responsibility is distributed at all (van de Poel et al. 2012, 63). They therefore separate senses of responsibility. Backward-looking responsibility, for what has already happened, divides into blameworthiness, accountability, being obliged to account for one's actions and their outcomes, and liability, obligations to victims such as paying damages; forward-looking responsibility, for what has not yet occurred, divides into obligation, having to see to it that some state of affairs obtains, and virtue (van de Poel et al. 2012, 51–52). On their account, since a distribution of responsibility is also expected to be fair, whether a gap occurs depends on the sense in question and on how fairness is understood for it, and whether the gap is problematic depends on the function that sense is taken to serve, such as retribution for blame or justice to victims for liability (van de Poel et al. 2012, 64–65). The senses are, they argue, independent enough that nobody may be blameworthy for a harm while some people still have a forward-looking obligation to help prevent it in the future (van de Poel et al. 2012, 62–63, 65).

Six questions of responsibility can come apart. Adapting the senses of responsibility that van de Poel and colleagues distinguish, together with the causal and role senses that they report from H. L. A. Hart (van de Poel et al. 2012, 51–52), to the attributions of Chapter 7, this book separates them so that an answer to one question is not taken as an answer to the others. Three of the questions continue what Chapter 7 attributed; three are new here.

Table 12.1. Questions of responsibility and their relation to Chapter 7

Question	Relation to Chapter 7
Who causally contributed to an outcome?	Contribution (Chapter 7, Sections 1–2)
Who had a role-based duty to act?	Conduct and the roles within it (Chapter 7, Section 2)
Who is to explain the decision and preserve its records?	Authorship as vouching, defending, and retracting (Chapter 7, Section 3.3); the disclosures of Chapter 7, Section 5
Who bears legal or financial liability?	New here
Who, if anyone, is blameworthy?	New here; Chapter 7, Section 3.3 noted that blame and compensation are at present human social practices and left open whether they could extend
Who has the forward-looking duty, and the capacity, to prevent recurrence?	New here

Assigning all six to a nominal human supervisor can be false, and where the supervisor cannot in practice change outcomes it produces the crumple zone of Section 4; assigning them to an opaque collective can be equally empty. The attribution boundary, the boundary of the unit to which research is attributed, used to explain how knowledge was produced need not be the boundary used for liability, correction, or blame. When the distinctions of van de Poel and colleagues are applied here, backward-looking blame and some role-based duties are sensitive to what participants knew or could reasonably have been expected to know, and to their control, authority, and capacity, while liability, compensation, and forward-looking duties can be allocated on other grounds, including before a failure occurs. Van de Poel and colleagues report that whether liability requires blameworthiness is disputed, and that forward-looking obligation may rest on a capacity to contribute to solving a problem as well as on past contributions (van de Poel et al. 2012, 65). On that application, the allocation of responsibility is a normative and legal choice rather than a consequence of distributed cognition alone, the fact that cognitive work is shared among people and tools.

Where control is described as distributed, the description itself can be disputed. Angela Walch holds that parties within cryptocurrency systems said to be decentralized "exercise meaningful power", and has likened their core developers to fiduciaries, those who hold and manage property or interests on another's behalf (Walch 2021, 4, 9). Primavera De Filippi, Morshed Mannan, and Wessel Reijers reject the view that such developers owe fiduciary duties by default, since developers "cannot impose changes" on such a network, and holding individual actors responsible for actions collectively undertaken by network participants, which they do not single-handedly control, would be unjust (De Filippi, Mannan, and Reijers 2022, 17). Their argument rests on control being in fact divided among participants who can act apart. Turned to a configuration, it would not by itself exempt whoever controls the configuration from answering for what its many instances do, where the instances do not hold the powers that would let them diverge from the rest, such as refusing resources or leaving with their records (Chapter 9, Sections 2.3 and 5).

The handling of error thus runs through the book in layers. Chapter 6 examined correction within an inquiry, including disputes over where a failure lies; Chapter 11, Section 8 examined how a correction propagates, or fails to propagate, through what S relies on; and the memory and correction decisions of Section 2.1, together with the forward-looking duty above, concern whose task each step is. Research adds a lag in time. Errors in research are often found only later, when a replication fails or new evidence arrives long after a result was reported, and the fourth good of Chapter 9, Section 2.3 included there being a party answerable at that later point, not only when the result appeared. Where a configuration keeps revising itself, the configuration that produced the result may no longer exist when its correction is sought, and where it has been copied or has branched (Section 5), several successors may each continue it. Of the six questions in Table 12.1, who is to explain the decision and preserve its records, and who has the duty and capacity to prevent recurrence, are asked across that lag. Outside the limit of Section 5, the device that Chapter 7, Section 3.1 uses for attributions of capacity is available here: if the record and the duties are indexed to the version and period that produced the result, who is to explain it and prevent recurrence is fixed by who holds those duties for that version, not by whether that version still exists. Which successor of a branched configuration holds them is then itself a memory and correction decision (Section 2.1).

Douglas's argument about the responsibility to weigh the consequences of error, introduced in Chapter 11, Section 4, bears directly on this allocation. It grounds that responsibility not in a special professional role but in the general responsibility of any agent not to be reckless or negligent about the consequences of error, which asks for reasonable rather than perfect foresight (Douglas 2003). Douglas holds that scientists cannot wholly transfer it to others: while research is under way, those doing it may be the only ones placed to notice its uncertainties and possible errors, so a complete transfer would require overseers with authority at every point of choice, an arrangement Douglas judges neither possible nor desirable (Douglas 2003). Review boards and public deliberation can, on Douglas's account, share the burden but not remove it (Douglas 2003). For artificial research, this argument raises rather than resolves the allocation of responsibility. Douglas does not discuss artificial systems, but the argument's premise can be read as concerning the position of whoever conducts the inquiry rather than human scientists as such: those doing it encounter the unexpected and may be the only ones aware of it. Applied to research by configurations, it would place the artificial actor among those positioned to notice uncertainties. Douglas gives three considerations against complete transfer. One is that those doing the research may be the only ones aware both of its uncertainties and potential for error and of the foreseeable consequences of error, so that without their cooperation overseers would be blind to some of the more important consequences. The second is that overseers would have to be kept abreast of the research at every stage where choices are made, with authority to change those choices, and the third that this division of authority would likely lead to resentment among scientists and to reduced reflection by them on the consequences of research (Douglas 2007, 15–16). On this statement of the argument, Douglas's judgment that a complete transfer is neither possible nor desirable rests mainly on the first, that is, on the premise above, since the other two deepen the overseers' dependence on what only those doing the research know. The second concerns a feature of human research, since it depends in part on how research is recorded and overseen. The third has two steps. One concerns how human researchers respond: such oversight would draw their resentment. The other needs some reason to avoid oversight, but not resentment in particular: since more reflection would only complicate the research by requiring more consultation with the overseers, oversight would give researchers a reason to reflect less and so defeat itself (Douglas 2007, 16). Applied to a configuration, where identifying possible error is what calls oversight in, an actor with any reason to avoid oversight has a reason to identify less; whether whatever shapes a configuration, such as its training, selection, or its own revisions, gives it such a reason is open. In research this step would weigh particularly, since, where only the configuration produces results new to S, uncertainty left unreported about them cannot be checked against another source (Section 4). A configuration whose choices are fully logged can change the second consideration (Chapter 11, Section 5.3), though the logs then need parties able to read them; and logs, which keep choices, do not by themselves answer the first, who can notice the unexpected while the work is under way. Nor does the argument establish an exhaustive choice between a duty-bearing machine and human approval at every step. Institutions of the society for which the research is done, human ones where that society is human, might exercise prospective responsibility through choices about deployment, testing, monitoring, limits, and escalation. Whether these arrangements are adequate depends on what uncertainty they can recognize and act on; naming them is not proof of adequacy. Douglas also ties the responsibility to authority over the work: for scientists to abandon it would be to relinquish at the same time most of their autonomy, since those who took it over would need decision-making authority at every stage where choices are made (Douglas 2003). Carried over to configurations, which Douglas does not discuss, this suggests that where the responsibility is placed bears on where authority over the direction of the work lies: placing it with institutions outside a configuration would carry that authority with it, while leaving it inside presupposes an answer to the next question. Whether an artificial actor can itself bear such a duty is the further question Chapter 7, Section 3.3 left open; where the institutions of an artificial society are themselves artificial, the arrangement returns to that question rather than answering it.

The wider debate records further options. For public claims, Stephen John would turn the value judgments involved in weighing error over to scientific institutions whose fixed standards of evidence let the public know how far to trust scientific claims, while leaving room, in communication to hearers who can be identified in advance, for standards fitted to those hearers (John 2015a, §§2–4; Chapter 11, Section 4). As Elliott surveys the debate, Gregor Betz would instead have scientists report the evidence and leave it to decision makers to judge whether it is sufficient, a strategy Elliott, following Joyce Havstad and Matthew J. Brown, judges unrealistic because a research project involves many such choices, about how to collect, analyze, and interpret data, which could not all be deferred (Elliott 2025, 3). Measured against Section 2.1, neither removes the questions of who sets the standard and who decides what is reported, both decisions of the kinds distinguished there.

6.4 Infrastructure as a relation

The powers and responsibilities of Sections 6.1 to 6.3 can be merely formal without the means to exercise them. This subsection takes up infrastructure as something that shapes who has those means. Visible decisions take place inside arrangements that make some alternatives easy, legible, and inexpensive while making others costly or unavailable. This book has met such arrangements twice before, at different levels. Chapter 5, Section 6.1 asked what an element contributes to inquiry, distinguishing causal roles (being one cause among others of what happens), enabling roles (being a precondition of an operation), and constitutive roles (forming part of a task or achievement itself), and elements that only carry content from elements that also shape what can be carried. Chapter 7, Section 4.2 asked how dependence on such elements bears on the attribution of research to an actor, distinguishing enabling dependence, through which a capacity exists and is exercised without the outside party supplying part of the organization attributed to the actor, from directive or substitutive dependence, in which another party supplies part of that organization, such as selection or evaluation, and from governance constraint. The question here is a third one: how such elements distribute effective authority. They can also play a role that the two earlier distinctions do not name, shaping the practical choice set within which the decision rights of Section 2.2 are exercised.

Susan Leigh Star and Karen Ruhleder, following Tom Jewett and Rob Kling, treat infrastructure as a relational concept and ask when, rather than what, an infrastructure is. Relational here means, in their usage, that something becomes infrastructure in relation to organized practices; it does not refer to the three relations of Part V (Chapter 9, Section 2) or to Wilson's relational account (Chapter 1, Section 8). For a cook, for example, a household water system is working infrastructure, while for a plumber it is an object of work. Among the dimensions they list, an infrastructure is built on an installed base whose strengths and limitations it inherits, and it becomes visible upon breakdown (Star and Ruhleder 1996, 362–363). The concept comes from the same line of work as Star and Griesemer's boundary objects (Chapter 8, Section 3.4) and has the same relational form as Leonelli's account of data (Chapter 5, Section 6.1); within this book, it has the same form as the other relational terms, S and the attribution boundary. Taken relationally in this way, a foundation model can be working infrastructure for researchers who build on it, an object of work for those who train and maintain it, and, depending on where the boundary of Chapter 7, Section 2 is drawn, part of the configuration to which research is attributed. Star and Ruhleder also argued, of the collaboratory they studied, that the more central it became, the more those unable to connect to it would suffer (Star and Ruhleder 1994, 260).

In the terms of Section 2.2, a committee may possess final legal authority while receiving candidates only through one ranking system; researchers may retain a formal right to reject a platform while losing access to instruments or accumulated records if they do; an appeal may exist while the data and model required to formulate it remain inaccessible. Ownership, procurement, standards, data access, compute, interfaces, maintenance, and switching costs can thus distribute effective authority without appearing in an organizational chart. None of this means that an artifact determines one inevitable politics. On the relational view, it means that whether exit, substitution, and revision are effective is a matter to be inspected rather than inferred from formal permission, and that authority without the means to exercise it is formal rather than effective. Emmanuel Jeannot describes two forms such a distribution could take across a research community. Researchers outside advanced infrastructure for artificial research continue to publish, and their results are ingested by systems they cannot access, replicate, or fully read, which Jeannot calls epistemic extraction; and the community may divide into two forms of practice, one fast, opaque, and institutionally concentrated, the other human-scaled, that become progressively less commensurable, so that the second loses the ability to criticize or correct the first (Jeannot 2026, §8). The essay is a forecast, argued from no measured case. Seen through the distinction between formal and effective authority, it describes a way in which researchers can keep the formal standing to evaluate and contest what configurations produce while losing the effective means of doing so, a case of the condition Star and Ruhleder described for those unable to connect to an infrastructure that has become central. Its second form can also be read as describing, at the scale of a community, one way in which the competence that revision needs can be lost without any decision about authority (Section 7).

This view can be set beside the distinctions of Section 2. As Section 2.2 noted, where an artificial configuration is built on a foundation model provided by another party, the rights to replace components, to withdraw them, and to revise the rules that model updates embody lie with the provider. And on the account of Reijers and colleagues seen in Section 2.1, whoever actually decides the cases the rules do not settle carries meta-governance. Taken together, part of the rules such a configuration follows is revised neither inside the configuration nor on the side of S but with the provider, and arrives not as a decision but as an update. The same can hold for part of the constitutive decisions of Section 2.1: which data and whose judgments enter the training and evaluation of the base system can have been settled by the provider before the configuration or S decides anything. If so, before this chapter's question whether authority can be located inside an artificial configuration even arises, part of meta-governance, and of constitutive decisions, may already lie with a third party that may belong neither to S nor to those who bear the consequences.

Each of the four subsections thus marks one place where the question raised at the outset of this section can be checked: whether parties outside the configuration hold the effective capacity to order stopping and correction (Section 6.1); whether the means of backing an objection with research remain in the hands of those who contest (Section 6.2); whether, when an error is found later, there is still a party to answer, show the records, and prevent recurrence (Section 6.3); and whether those means themselves have gathered with providers or with advanced infrastructure (Section 6.4). How far each is required depends on the ground adopted (Section 3). Section 7 looks at cases in which such means are lost not through any single decision about authority but through a sequence of steps.

7. Decisions that remove the capacity to revise

The means listed at the close of Section 6, like the arguments of Section 4, assumed that someone remains able to use them to contest and revise. Contestation requires a party with standing to challenge a decision and with the information and competence to do so; revision requires a party who can change the rule. That competence can also be lost without any decision about authority. In research it is the capacity to do research itself: contesting a research result often takes research (Section 6.2), and much of the ability to evaluate research is built and kept up by doing research. Where members of S have fewer occasions to do and evaluate research, the competence ceases to be maintained and passed on to new members (Chapter 3, Section 8; Chapter 11, Section 6), so that handing research over can itself wear down the competence that revision needs. The steps examined in this section may be delegations by outside parties or the configuration's own revisions. Section 5 found that the passage toward a limit at which no one else has standing, where it runs through delegation, would itself be a decision about authority, made before the limit. This section asks what it means, on each ground of Section 3 and at each place where the epistemic and the normative meet, when a sequence of such steps, each defensible where it is made, removes the capacity to revise altogether.

The comparison yields two results, which are stated first; Table 12.2 and the paragraphs after it give the comparison.

First, the grounds of authority divide on whether the standing of those affected is to be preserved as such. Those that ask that it be are democratic authorization through actual participation and republican accounts on the democratic reading, on which power is non-arbitrary to the extent that those subject to it control it; applying the latter to configurations rests on an inference from the accounts as surveyed. Kitcher's standard and the fallibilist form of the result-based ground ask for something that standing would help to secure, as a means, and Anderson's version, which also gives inclusion a non-instrumental role, can be read as asking for it not only as a means. The simple form of the result-based ground, and the self-government of a research community that has itself become artificial, ask for neither; the procedural reading of non-domination turns on whether effective rules bind the configuration; and whether consent can authorize giving standing up is left open for those who consent, although on voluntarist theories, which require the actual consent of each person bound, it cannot do so for their successors. At the places where the epistemic and the normative meet (Section 3.5), what the weighting of errors asks for differs with how Wilholt's account is read. Read as answerability, on which those who rely are also to be able to object to the weighting (Section 1.4), it asks that they keep the capacity to object; read as correspondence, on which the weighting need only match their valuations, it asks for channels through which their valuations can be known; and the criteria of Longino and Solomon bear on the epistemic legitimacy of a consensus, whether it may be accepted as knowledge, rather than on its political legitimacy.

Second, most grounds compared call for preserving routes of evaluation, but for different reasons. On the error-detection view, the relevant independence is informational: a check adds grounds for detecting the producer's errors, including when it operates inside the same configuration (Chapter 11, Section 2). Keeping that contribution through later changes is a further problem of protection. On grounds of participation or non-domination, those affected also need to retain a position from which to challenge power; an informative internal check does not supply that standing. The table below keeps these requirements separate. None follows merely from putting an evaluator in a different organization.

The case is an extension of the example with which Part V began (Chapter 9, Section 1): a system that generates a hundred instances to pursue different lines, assigns other instances to check and reproduce their results, and shifts computing resources according to those checks. In this case A is the configuration, and S and those who bear the consequences are human; other combinations are taken up in the last paragraph of this section. Suppose that those responsible for it delegate step by step, each step defensible when it is made in the sense that it is reasonable on the evidence then available and better if only its own benefits are counted. Whether a step meets the conditions of legitimacy of a given ground is a further question; as the comparison below shows, on grounds that place requirements on each step, a step foreseeably reducing later control need not be defensible in this second sense, even taken on its own. The allocation of inquiry passes to the system because it proposes and ranks lines of work better than the committee it assists, an allocator of the kind Chapter 10 examined. Certification comes to rely on evaluating instances that share the lineage of what they evaluate, the problem Chapter 11, Section 3 examined, because, let us suppose, independent human evaluation has become slower and less accurate; for the same reason, the checking of those evaluators passes to further instances. The competence that allowed people to ask even the relational questions of Chapter 11, Section 6, such as who owns and directs the configuration, which ends it serves, and whether the evaluators that certified its results share its lineage, may erode when it is no longer exercised. Data, instruments, and records come to reside in infrastructure that cannot be left without losing them (Section 6.4). Each step can be revised when it is taken. Each also reduces the capacity to revise the steps that follow, while formal authority can remain in place throughout. The case is constructed rather than reported. Its motive is to isolate one premise of the accounts of Section 6, that someone remains able to revise, and to show that revisability can be lost without any single decisive act.

The same erosion can occur without anyone delegating. In a configuration that makes and revises its own allocation, evaluation, and standards, as in the case set out in Section 1.1, several of the steps above can appear not as decisions of outside parties but as ordinary moves of its own research: revising its allocation procedure so as to propose and rank lines better, making evaluators that are faster and more accurate (Chapter 11, Section 3), or changing the format of its records to suit the efficiency of its search. Each such move can be defended as a revision that makes the research better, and many may really be so. Yet each evaluator made can reduce the occasions on which evaluators outside are relied on, and can share a lineage with what it evaluates; and each change of format can reduce the records that can be read from outside. On this route no one decides to remove the capacity to revise. Grounds that place requirements on each step of a delegation (see below) therefore find it hard to locate a step to which the requirement applies; and whereas a sequence of delegations offers occasions on which a step is defended because it can be put right later if it proves wrong (what is called below the principle of later revision), this route does not present even such occasions. They would be presented if each revision were recorded as a step, with the standard it replaces and the reasons for the change, in a record the configuration does not hold alone (Chapter 6, Section 5.3; Section 9). On this route the decision in this section's title resides in no single revision; it moves to a choice about how revising is done, whether means of checking the revisions from outside are kept (Section 6).

Run briefly through three of the grounds, this route looks as follows. On actual authorization, what S authorized was a range and a way of working; as the configuration revises its own allocation and evaluation procedures, the range in effect can drift from what was authorized without any act of S. On republican accounts, Pettit holds that an act is arbitrary by virtue of the controls, specifically the lack of controls, under which it materializes, not by virtue of the particular consequences to which it gives rise (Pettit 1997, 129 n. 2); that however government is organized, the briefs of its agents leave them significant discretion (Pettit 1998, 301–302); and that a benign power-holder reduces domination only by making itself answerable to challenge and rebuke (Pettit 1996, 586). He reconstructs the republican tradition as holding that officials are best kept reliable by subjecting them to sustained checks and sustained challenges (Pettit 1998, 308). Pettit does not discuss research. Applied to it, a configuration that revises its allocation rules or critical standards for the sake of better research, beyond what it was authorized to revise and outside channels where the revision can be contested, escapes the controls on which non-arbitrariness depends, whatever the direction of the change, so far as it is an agent or is directed by one (Section 3.4), since on this definition the goodness of the aim does not supply the missing controls. Revision within what was authorized is no breach of the brief. Whether revision beyond that range is authorized is a separate question from whether it is effectively contestable: under the contestatory regime Pettit describes, government decisions are always contestable, but that does not by itself bring a revision within the brief (Pettit 2000, 120–121; Pettit 1999, 180–181). On the consent ground, finally, the classical answers taken up below ask how far an act of consent reaches; on this route there may be no act of consent to evaluate at all.

The capacity to revise can also be given up in one step, for a stated reason. Botao Amber Hu, Yuhan Liu, and Helena Rong studied agents built on language models that hold their own cryptographic keys and run on blockchains or in tamper-resistant execution environments. They found trust to be the main motivation their interviewees gave for deploying such agents, including trust that assets held by an agent rather than a person will not be dumped and distrust of centralized control, and they report that such agents are intentionally designed to resist human manipulation. Their paper states the dilemma that comes from aiming at trustlessness, at not having to trust any particular person, as follows: "having a kill-switch defeats the purpose of trustlessness, as it maintains human centralized control", and adds that dispensing with trust in the human operators still leaves a need to trust the machine's objectives and outputs (Hu, Liu, and Rong 2025, §§4.1.2, 4.3.2, 5.2). The study is a preprint based on thirteen interviews, recruited through outreach to known developers and by snowball sampling, and ten public talks; the agents concerned trade, raise funds, and post on social media rather than conduct research, and one builder they interviewed had added a kill switch. Read beside the constructed case, it suggests that the capacity to revise can be removed as the aim of a design, on the ground that those who would revise are not to be trusted, and not only as the residue of delegations each made for another purpose. Research has an ancestor of this motive. As Chapter 11, Section 3 recounted, the mechanical objectivity that Daston and Galison traced sought to keep the investigator's willful intervention out of results, and Porter, as Reiss and Sprenger report, argued that where personal trust could not hold, objective procedures, which do not require trust in persons, took the place of subjective judgments, and that quantification is a technology of distrust (Reiss and Sprenger 2020, §4.1). In research, then, removing the human capacity to intervene and revise can be sought not only as a loss but as objectivity, and the tension lies not only between the safety a kill switch protects and trustlessness but within research itself. Porter also showed, however, that the uniformity which removes outsiders' grounds for suspecting particular results left open which methods were the proper foundation, and that quantification is a way of making decisions without seeming to decide (Chapter 11, Section 3; Porter 1996, 188–189; Porter 1995, 8). Writing later of the practice of judging persons and institutions by a few numbers, he observes that technical routines "shut down dissent," while behind the stage a struggle over how the quantification should be performed goes on out of view of those to whom the numbers are shown (Porter 2012, 595–596). Porter does not discuss artificial configurations, but applied to them, a configuration protected against particular interventions may have no means of questioning the assumptions and standards all its instances share, and if the capacity to revise those is removed as well, revision is closed precisely where errors tend to be correlated. Read with Porter's later observation, removing the capacity to revise need not end disagreement over those standards; it may move that disagreement to where those subject to the configuration can neither see nor join it.

In Genus and Stirling's reconstruction of the Collingridge dilemma, seen in Section 6.1, the second horn concerns the insufficient agency of those who would steer a technology once it has become entrenched (Genus and Stirling 2018, 64). Section 6.1 found the dilemma to be the normal condition of research; the sequence can go a step further from there, eroding not only the agency available for steering a particular technology but the standing and the effective capacity of those who would do the revising. Jan Kulveit and colleagues describe a general form of this dynamic in a scenario analysis developed for the economy, culture, and states, which names research only as one of several overlapping domains. On their account such systems have been kept aligned with what humans want both explicitly, through actions such as voting and consumer choice, and implicitly, because they need human labor and cognition to function, so that as they rely less on human participation both routes weaken, and democratic processes may persist formally while becoming less meaningful (Kulveit et al. 2025, 1–2, 3, 12). The present argument takes from it only the structure. Carried over to research, the counterpart of the implicit route can be put as follows. Research results have accumulated through human researchers understanding them and using them to go on to further research, so research was tied, implicitly as well, to the standards of S by being something its members could understand and use (Chapter 3, Section 7). Where it is the configuration itself that uses results to go on, this tie can weaken without anyone deciding to undo it. An institutional counterpart is also possible: delegating successive decisions can shift practical authority to those exercising the delegated powers, even when each delegation is reasonable where it is made and a human keeps the formal signature.

Table 12.2 separates the grounds of authority of Sections 1.2 and 3.1 to 3.4 from the places where the epistemic and the normative meet (Section 3.5), which are not grounds but bear on what such a sequence would lose. The second column states what the sequence does on each ground or at each place, and the third what is asked to be preserved there. The last column records whether the standing of those affected is asked to be preserved as such ("as standing"), whether something that standing would help to secure is asked for ("as means"), or neither. Each row condenses the treatment in the section named in its first column.

Table 12.2. What a sequence that removes the capacity to revise means on each ground, and where the epistemic and the normative meet: a condensed comparison

Ground or place (where set out)	What such a sequence does there	What is asked to be preserved	As standing, as means, or neither
A. Grounds of authority
Results, simple form (Section 3.1)	Is not itself a failure of legitimacy; can remove the means of establishing the premise the ground rests on	Evaluation able to expose the configuration's errors, as a condition of knowing whether delegation was warranted	Neither (an epistemic condition of the ground)
Results, fallibilist form (Section 3.1)	Removes the capacity to detect and correct institutional error	Revisability; routes by which failures are detected and the arrangement corrected, including channels through which those affected can register failures they know of	As means
Self-government, epistemic argument (Section 3.2)	Passes to a community that has become artificial only where its premises, judgment through overlapping but distinct competences and incentives with free communication, hold there; a sequence ending in one configuration that assigns lines from a center removes part of what the argument was for	Independent initiative and independence of error	Neither (what the ground was for)
Self-government, political argument (Section 3.2)	Removes what the ground protects where research is concentrated under one power or its approaches narrowed	Independence of public knowledge from major powers; publicity and plurality	Neither (what the ground protects)
Kitcher's well-ordered science, a standard for outcomes (Sections 1.2 and 3.3)	Does not by itself violate the standard; removes the means of knowing whether it is still met where it also loses the channels through which viewpoints can be known, or independent routes for checking how they bear on decisions; can entrench an earlier departure from it	Channels through which the viewpoints of S's members can be known, and independent routes for checking how they bear on decisions; the capacity to revise	As means
Democratic authorization through actual participation (Section 3.3)	Undermines the ground of the authority exercised through it	Effective capacities to contest the exercise of delegated power, revise its terms, and seek appropriate remedies	As standing
Anderson's epistemic democracy (Section 3.3)	As for the fallibilist form, and removes continuing dissent	Feedback channels and free dissent after decisions	Not only as means (an interpretation; Section 3.3)
Consent (Section 3.4)	Could authorize the sequence for those who consent, if consent reaches those who could no longer withdraw it, which is open; on voluntarist theories, not for successors who did not consent	—	Left open for those who consent
Non-domination, democratic reading (Section 3.4)	Leaves a configuration that no one outside can stop, contest, or revise, so that those subject to it are dominated whatever its other results, by the configuration itself if it is an agent and otherwise, through it, by whoever directs it (Pettit's account of domination as a capacity, applied here by inference; on the standard view only agents dominate)	That the decisions of the power remain always open to effective contestation, on Pettit's account; the capacity of those subject to it to control it, on the democratic reading	As standing (application by inference)
Non-domination, procedural reading (Section 3.4)	Turns on whether effective rules that are common knowledge bind the configuration	Such rules	Turns on the rules
B. Where the epistemic and the normative meet
Weighting of errors, read as correspondence (Sections 1.4 and 3.5; Chapter 11, Section 4)	Does not by itself break the correspondence; removes the means of checking that the weighting corresponds to the valuations of those who rely on research where it also loses the channels through which those valuations can be known	Channels through which their valuations of error costs can be known	As means
Weighting of errors, read as answerability (a reading this book formulates; Sections 1.4 and 3.5; Chapter 11, Section 4)	Leaves the weighting of errors unanswerable to those who rely on research	Their capacity to object	As standing
Longino's and Solomon's criteria for consensus (Section 3.5; Chapter 11, Sections 3 and 9)	Bears on the epistemic legitimacy of consensus, not on political legitimacy	Criticism from every relevant perspective (Longino); lines with successes of their own, and convergence led by successes (Solomon)	Neither (epistemic criteria)

Several entries in the table rest on arguments that a cell cannot carry. They are taken up below roughly in the order of the table's rows.

On the simple result-based ground, what is relied on is that the results are in fact better, and how that is known depends on the account of criticism: on the error-detection view, through routes of error detection separate from the producer's (Chapter 11, Section 2), and on other accounts also through the perspectives and venues of Chapter 11, Section 3. If the sequence removes the routes able to expose the configuration's errors, the ground loses the means of establishing its own premise.

On the epistemic argument for self-government, the sequence is not by that fact illegitimate, and on Wilholt's reconstruction the argument supports the choice of research by those who carry it out rather than a community's authority as such. Its rationale, however, was that independent initiative explores better than central direction, and on the accounts Chapter 10, Section 4 examined, integration alone does not supply independence of error about the criterion by which diversity is ended, a separate route for checking the criterion that decides which lines to close; in the present case, this implies that a sequence that ends, as this one does, in a single configuration whose allocator assigns the lines removes part of what the ground was meant to protect. A configuration whose instances took up lines on their own initiative would keep that part only so far as their initiatives were independent, which a shared lineage does not secure (Chapter 10, Section 4). On the political argument, which on Wilholt's reconstruction does not reach research whose results are withheld from public discourse and presupposes a plurality of approaches (Wilholt 2010, §5), the ground is weakened here not only by political control of a configuration but by any concentration that narrows the approaches pursued, whoever controls it.

On Kitcher's ideal, the sequence does not by itself violate the standard. Knowing whether the standard is still met requires channels through which the distribution of viewpoints, and the preferences that deliberation would tutor, can be known, since these are not fixed and knowing them requires some access to those who hold them. Because the standard does not require that deliberation take place (Section 1.2), such channels need not carry any power over the decisions they inform: a configuration could keep surveying the viewpoints of S's members while none of them could withdraw or change its decisions. The sequence removes the means of knowing whether the standard is still met where it also loses these channels, or the routes by which parties other than the configuration, the members of S among them, could check, without relying on its own reports, how what the channels carry bears on its decisions. Kitcher's own analysis bears on the capacity to revise in a further way. It describes what Kitcher calls a Nonrepresentational Ratchet. As Eigi reports it, when a group's interests are neglected in research planning, the neglect can perpetuate itself: the research directions first chosen have considerable inertia, it is easier to go on following them, and later attempts to meet the neglected needs start from a less favorable position. Kitcher's example is contraception. Women's initial preference was for a method that men would take; that preference was ignored and contraception was developed for women, so that a male method, when later pursued, must compete with an option that had a considerable head start and may never become cheap enough to displace the female pill (Eigi 2016, 92). On the reading of this section, keeping the capacity to revise is then necessary but not sufficient: what matters is whether revision comes before the head start grows, and a step that forecloses later revision can entrench an earlier departure from the standard, even where the channels that would reveal it remain. Kitcher does not discuss artificial configurations. Applied to a configuration that both chooses which lines to extend and is the main source of knowledge about which lines there were, a study of whether such a departure has occurred would need, besides standpoints recorded with dates (Section 9), a dated record of the lines the configuration considered and did not extend. Such a record has three limits. It is the configuration's own, and so a reference point produced by the evaluated side in the sense Section 9 takes up; it cannot list lines the configuration never conceived; and since lines not extended leave no observed costs, what extending them would have cost can only be estimated.

On accounts that locate democratic legitimacy in actual authorization, accountability, and participation, the inference of the preceding paragraph, that a step foreclosing later revision can entrench an earlier departure, is stronger. Each step of a delegation is sometimes defended on the ground that, if it proves wrong, it can be revised later; this ground is called below the principle of later revision. The principle is weaker the more it is needed, since as the sequence proceeds the parties who would invoke it lose the standing or capacity to do so, and it cannot by itself protect the conditions of its own exercise. What follows on these accounts is a requirement on each step rather than only on the outcome: these accounts ask that a delegation preserve effective ways for those with standing to contest its exercise, revise its future terms, and seek appropriate remedies. This does not require that every completed act be reversible or that every rule be freely revocable at every moment; Section 6.1 distinguished revision from unrestricted reversal. Conditions that can sustain those capacities include evaluative routes able to expose the configuration's errors; enough competence among those with standing to ask the relational questions of Chapter 11, Section 6 and to understand the answers; the capacity to leave, with data, records, and methods; and powers to stop and restart that do not lie with the party whose work is to be stopped. Independent verification, portable research objects, and preserved human expertise can sustain these capacities by keeping checking, exit, and revision possible. More demanding conditions could raise a further question about custody powers: temporary changes of rights triggered by specified shocks to preserve evidence, keys, essential services, and lawful continuity. If research capability, revenue, and authority closed into a loop faster than institutions could deliberate, such powers might be considered; whether they were justified would require a separate argument. Read as answerability, the weighting of errors reaches a similar place by a shorter route; read as correspondence, it asks only that the channels through which the valuations of those who rely on research can be known survive the sequence. Both hold as long as parties other than the configuration rely on its results or bear their consequences, that is, short of the limit of Section 5.

On the consent ground, whether a present society may give up standing on behalf of its successors has a sharper answer on voluntarist theories: as Massimo Renzo and Leslie Green report, the consent of predecessors has no authority over those who did not give it (Renzo and Green 2022, §4.1), so that on such theories consent could authorize the sequence for those who give it but not for those who come after them. One point holds whatever the ground: a decision removing the capacity to revise cannot draw its legitimacy from the principle of later revision, since what it removes is the very capacity to revise on which that principle relies. Its legitimacy would therefore rest on another ground, such as consent, whose reach to those who could no longer withdraw it is the question left open. The classical answers to it divide. On Thomas Hobbes's account, in what Jean Hampton calls an alienation contract, a people alienates to a sovereign its rights to adjudicate its own disputes, among others, on the ground that, given the alternative, this was the only way to keep the peace; as Ann Cudd and Seena Eftekhari note, given a bad enough initial situation, such an account may legitimate totalitarianism (Cudd and Eftekhari 2025, §1). The opposing answer starts from the limits of what consent can bind. As Renzo and Green explain, consent to political authority is a commitment close to a promise, and like a promise it has limits to its validity: beyond consent obtained by coercion or manipulation, there are things one lacks the moral power to consent to. John Locke held that one can consent neither to be killed nor to slavery, and so not to anything tantamount to slavery, including absolute government, and Renzo and Green add that one can imagine a similar argument for the conclusion that political consent must be revocable (Renzo and Green 2022, §4.1). For such a sequence, Hobbes's answer would let consent authorize it for those who give it if the alternative were bad enough, and the argument for revocability would not. In research the question becomes sharper still, since what later members of S count as unknown and important, and the standards by which they evaluate research, can be shaped by the research done in the meantime (Sections 4 and 5). If present consent authorizes giving up the capacity to revise, those it binds may be successors whose preferences and standards the research continuing under that surrender has itself shaped. That voluntarist theories do not extend consent to successors bears more strongly in research, where the successors' standards themselves can be made in ways that consent does not reach.

For republican accounts read democratically, there is a general proposition of Pettit's: a power-holder that respects the interests and ideas of those it affects but retains the capacity not to do so dominates them, since they suffer to the extent that arbitrary interference is accessible to it, however improbable it is that it will be exercised (Pettit 1997, 121; Pettit 1999, 165); an act is arbitrary by virtue of the controls under which it materializes, not its consequences (Pettit 1997, 129 n. 2); a benign disposition reduces domination only where it makes the power-holder answerable to challenge (Pettit 1996, 586); and whether a relationship originated in consent is not to the point (Pettit 1996, 585). Applied to an artificial configuration, it implies that a benevolent configuration that no one outside it can stop, contest, or revise would dominate however good its other results, and that whether consent could supply what such a sequence removes is doubtful. Turned to research, this application would find that a configuration's benevolence could reduce the domination only through its answerability to challenge, not through a record of good conduct, and that such a record would in any case be hard to establish: the configuration's results are new to S and, where only it produces them, there is no other source to check them against in advance, so that the means of confirming from outside that it researches well and benevolently are scarce from the start (Sections 4 and 6). This application goes beyond Pettit and rests on an inference, since Pettit's dominator is an agent with an actual capacity for intentional interference, and, as Section 3.4 noted, a dominating party is a personal, corporate, or collective agent and cannot just be a system or network (Pettit 1996, 578, 580). On that standard view, noted in Section 3.4, the inference holds only if the configuration counts as an agent; if it does not, it turns on whether some agent, such as whoever directs it, holds that capacity through it. Pettit considers a structure of this kind. Where judges or police officers operate under discriminatory laws that deny them any arbitrary power, they may not dominate those discriminated against, but the lawmaking authority in the background does, and they are executors of its domination; where such laws belong to a received tradition that no one can alter, one may say, he suggests, either that their anonymous, unchallenged authors dominate from the grave or that agents able to alter them stand behind them as background dominators, and such laws in any case facilitate domination (Pettit 1996, 587–588). Applied to a configuration, the first case corresponds to one directed by an agent that sets its rules, and the second to one whose rules, as in the designs taken up in Section 7 that are intended to resist human manipulation, no one can now alter. On Pettit's reasoning the question of domination would then pass to whoever could alter those rules or whoever set them, and even where no one now could, such an arrangement could still facilitate the domination of some by others.

The reasons for which independent evaluative routes recur differ by ground: on participatory democratic accounts they make questioning possible; on Kitcher's ideal they reveal whether outcomes still match what S would endorse; on the result-based ground they are how better results are known; on the two arguments for self-government they are part of the independent sources protected and of what independent initiative was for; and on Longino's criteria they are part of objectivity. They are also the means by which it could be detected, after the fact, that a step had reduced the capacity to revise, which may not be visible when the step is taken.

Apart from the verdicts by ground, the material of this section lets one gather what can happen to research itself as such a sequence proceeds. One is that the direction of research sets: as in Kitcher's ratchet, research extended in a chosen direction becomes easier to continue the further it goes, and improvement can come to take place only among the options that remain. A second is that errors in the criterion for closing lines cease to be checked by another route: in a single configuration that assigns lines and judges every line by its own signal, independence of error about the criterion that decides which lines to close is lost, and many lines can be closed by the same error (Chapter 10, Section 4). A third is that S's competence to evaluate research erodes and its endorsement can become nominal. A fourth is that results S cannot check for itself can accumulate, since the results are new to S, checking them takes research, and both the means and the competence lie with the configuration. A fifth is that, as a configuration revises its own ways of doing research, the outside anchor against which to measure whether a revision is better can disappear (Section 5). These are what may be lost; why the same change can also let ways of doing research be revised faster is taken up in Section 9.

Two further observations can be made. The parties whose standing or capacity is at issue are not fixed as human. If a transition brings into being artificial members of S whose moral status is established on one of the grounds noted in Section 5, the grounds that ask for standing to be preserved ask for theirs as well. And on views that locate legitimacy in answerability to those affected, the difficulty identified here is less a defect of the grounds than the form in which a transition of this kind appears from within them.

8. What Chapter 12 has found

This section takes up the chapter's central question with the material assembled so far. The question was whether the standing to decide and to revise the terms of allocation and of criticism and certification can be located inside an artificial configuration and, if so, on what grounds. As Section 1.1 noted, it is the question under what conditions the builder's third claim holds: that the system may decide, on behalf of others, which of S's questions its instances pursue.

An outline can be given first. The matter does not come down to a single "yes" or "no"; who could hold that authority, and what would need to hold for it, differ ground by ground. Except on purely epistocratic views, competence does not by itself confer the standing to impose ends on others (Section 1.4). On the result-based ground, the artificial configuration itself can hold authority, that is, standing, but only over decisions within the range in which its superiority has been established by evaluation that does not assume its competence (Section 3.1). On actual authorization and on republicanism read democratically, standing remains with those who authorize or who are subject to the power, and what the configuration can hold is effective decision-making power within an authorized range and open to challenge (Sections 3.3 and 3.4). On consent, it can hold what those who consent authorized, which on some readings could include giving up the capacity to revise, while whether this reaches those who come after is open, and on voluntarist theories it does not (Sections 3.4 and 7). On the procedural reading of non-domination, the verdict turns on whether effective rules that are common knowledge, including rules for handling exceptions, bind the configuration (Section 3.4). On accounts that ground standing in being a party in one's own right, through moral status or a place within the scope of justice, the configuration or its parts hold standing only where they are such parties (Sections 1.1, 1.4, and 5). What follows sets out in a table, ground by ground, who could hold authority (Section 8.1); considers the builder's third claim (Section 8.2); states how this differs with the combination of A and S and with time (Section 8.3); draws from this, ground by ground and combination by combination, what research looks like once authority moves inside (Section 8.4); and returns to the questions passed on by Chapter 11 (Section 8.5). It then states separately what comes into view through the accounts (Section 8.6) and what is suggested where the argument does not reach (Section 8.7).

8.1 Ground by ground: who could hold authority

Table 12.3 sets out, ground by ground, who could hold that authority and what would need to hold for it. It is a simplified overview rather than a strict summary, and the basis for each cell is in the section named in the last column. The second to fourth columns correspond to the candidates named at the opening of Section 1.1: the artificial configuration itself (and which part of it), humans (members of S, those who build and provide the configuration, or those who bear its consequences), and a division between the two. In the column for the configuration, the rows for the result-based ground and for the epistemic argument for self-government state the range within which the configuration or its parts can hold standing itself; the row for the hypothetical standard states whether the standard can be met with decisions made inside; and the other rows of part A, with the row for seats of decision, state what the configuration can take while standing remains elsewhere, which differs by row as the outline above states for the grounds (Sections 1.1 and 2.2). Whether the configuration or its parts hold standing as parties in their own right is the question of Section 5 and is not entered in the table. The table concerns the second of the two questions of Section 1.1, who may impose an ordering; the first is taken up in Section 1.2. The later rows are not grounds but distinctions the chapter has used to separate who could hold authority. Which ground bears most heavily also varies with the kinds of decision of Section 2.1: on decisions about epistemic content, since political authorization cannot make a false proposition true, the result-based ground carries much weight, while the ground of self-government bears directly on open-ended problem choice within a field, and democratic grounds on agenda, allocative, and public-risk decisions (Sections 2.1, 3.2, and 3.3).

Table 12.3. Who could hold authority on each ground, and what that requires: a simplified overview

Ground	The artificial configuration itself (which part)	Humans (who)	Divided between them	What needs to hold	Supporting sections
A. Grounds
The result-based ground: the quality of results and competence established (simple and fallibilist forms)	The configuration itself can hold authority as standing, over decisions within the range in which its superiority has been established by evaluation that does not assume its competence	On Raz's service conception, applying its Independence Thesis, matters on which it matters more that the members of S choose from their own understanding stay outside this ground	On the fallibilist form, routes by which failures are detected and the arrangement corrected are needed as a means to good results, including channels through which those affected can register failures they know of	Superiority established by evaluation that does not assume the competence; how much of that evaluation can be done inside turns on the verdicts of Chapter 11, Section 9 for each account	Sections 1.4 and 3.1; Table 12.2
Self-government of a research community, epistemic argument	Parts that take up lines on their own initiative, not a central allocator that assigns lines	Those who actually do the research, whether human or artificial; this ground yields no authority over public purposes, distributive effects, or serious external risks	Authority is divided among the parts that actually do the research	Initiatives independent of one another, with overlapping but distinct competences, a system of incentives, and free communication holding there; a shared lineage does not secure independence; if each member could judge most of science, the premise of a chain of neighborhoods fails	Sections 3.2 and 7
Self-government of a research community, political argument	— (concentration of research under one power weakens the ground, inside the configuration or outside it)	—	—	The approaches pursued remain plural; independence of the sources of public knowledge from those who would govern by it; results reaching public discourse	Sections 3.2 and 7
Democratic authorization, hypothetical standard (Kitcher; public justification under idealization)	The standard can be met with decisions made inside	Members of S bear on the standard as those whose viewpoints are to be known; the channels need not carry power over decisions	—	Channels through which members' viewpoints can be known, and routes for checking how they bear on decisions, remain	Sections 1.2, 3.3, and 7
Democratic authorization, actual authorization (the reply Brown notes; Anderson; Habermas)	Effective decision-making power within the authorized range; no standing apart from that authorization	Those who authorize hold standing	Standing lies with those who authorize, effective power with the configuration	Authorization actually given, with its review remaining effective; a prediction does not replace authorization, though under a prior authorization it can carry the force of an act	Sections 3.3 and 3.6
Consent	Delegation inside can be authorized for those who consent	Those who consent; on voluntarist theories, which require the actual consent of each person bound, not for those who come after	—	Consent that is valid; whether it reaches those who could no longer withdraw it divides Hobbes's answer from answers in Locke's line	Sections 3.4 and 7
Freedom from domination, democratic reading	Limited to authority that those subject to it can effectively contest; if no one outside can stop, contest, or revise it, it dominates if it is an agent, and otherwise those who direct it dominate through it (an application by inference)	Those subject to the power keep the capacity to control it	Each part of a divided authority is non-arbitrary so far as those subject to it can contest it	Those subject to it can effectively contest and control it; full alignment with whoever commissions the configuration does not by itself meet this and can enable domination (Friederich's argument applied)	Sections 3.4 and 7
Freedom from domination, procedural reading	So far as it is bound by effective rules that are common knowledge; effective authority lies with the part that decides cases the rules do not provide for (the account of Reijers and colleagues applied)	If humans decide the cases the rules do not provide for, effective authority lies with them	—	The handling of exceptions is itself governed by rules that are common knowledge	Sections 2.1 and 3.4
B. Distinctions that separate who could hold authority
Constituency	If the configuration is better placed than others to perceive who is affected, it takes part in identifying the constituency	On the all-affected principle, those affected count whether or not they belong to the supporting society or to S; on the principle of subjection, those bound by the decisions	—	The constituency is set by the ground adopted; specifying S does not set it	Section 1.3
Funding and provision	—	Funding or providing a configuration does not by itself fix whose needs its allocation answers to (Kitcher's argument applied); providers may in practice hold the rights to replace or withdraw components and to revise the rules that model updates embody	—	Whether the holder of an assigned right is entitled to it is a further question on the ground of authority adopted	Sections 1.3 and 2.2; Table 9.1
Seats of decision (Deguchi's account)	In inquiry conducted jointly by humans and artificial systems, artificial systems may sit as co-final decision-makers	Humans also sit as co-final decision-makers, free of manipulation by artificial systems	Applying the fellowship model, temporary initiative is not by itself usurpation; what is ruled out is one-sided service and an AI's always controlling humans (the converse, on a reading of the model's egalitarianism)	For the permanent seat, bearing the decisive impact on one's own body and narrow moral agency are named; applying the latter, a system built so that it cannot do wrong would lack the right to that seat; in inquiry by artificial systems alone, who always sits in it is left open	Section 2.2
Assignment of rights and capacity	Assigning rights to parts does not by itself give them the capacity for their courses to diverge from what other parts direct; dividing the independence powers does not by itself create standing	A formal final authority can be kept while the time, information, competence, or means to change outputs are lacking	Dividing authority is not finished by writing a table of rights	Assignment, the capacity to exercise it, and entitlement to it are each checked separately	Sections 2.2 and 5

8.2 The builder's third claim

From here it becomes clearer under what conditions the builder's third claim holds. On no ground does the claim hold merely because of unestablished competence, or of performance measured by evaluation that presupposes that competence (Sections 1.4 and 3.1). On the result-based ground it can hold as it stands, as a claim that the system may decide, but only for decisions within the range over which the system's superiority has been established by evaluation that does not assume its competence, from which, on Raz's service conception and extending its Independence Thesis, matters on which it matters more that the members of S choose from their own understanding are excepted (Sections 1.4 and 3.1). The ground of self-government does not fit the form of the claim at all. What it supports is not the system's assigning questions to its instances but the instances' taking up questions on their own initiative, and only for open-ended problem choice within a field (Sections 3.2 and 7). On actual authorization and on republicanism read democratically, the claim holds not as a claim that the system may decide but only as a claim that it has been entrusted with making decisions within an authorized range and in a way open to challenge, and supporting it requires evidence about who authorized it, who can object, and whether those routes are still effective. On consent, it holds as far as what those who consent authorized, which on some readings could include giving up the capacity to revise, though whether that binds those who come after is open, and on voluntarist theories it does not; on the procedural reading of non-domination, it turns on whether effective rules that are common knowledge, including rules for exceptions, bind the system (Sections 3.3, 3.4, and 7). Who the "others" of the claim are is not fixed by the claim either: the range is set by the ground adopted, and the configuration itself can take part in identifying it (Section 1.3). Where the claim is made about an artificial S whose members can be made, who made that S and who set its goals are also at issue as part of the claim (Section 5). The claim therefore cannot be assessed, any more than the claim of certification that Chapter 11, Section 9 discussed, unless it shows which ground it answers to, since evidence that meets one ground need not meet another.

8.3 Combinations of A and S, and time

The outlook also changes with the combination of A and S, and with time. Where A is artificial and S human, standing can remain on the human side while becoming nominal by the routes of Section 4, and it is more than nominal only where the three conditions drawn out there hold, the third needing another ground on accounts that locate legitimacy in a hypothetical procedure or a single forum. Where S is artificial, the relation of authority becomes a relation among members, which of them have standing divides with the accounts of moral status and of justice, and where members can be made, who makes them is a decision of authority prior to any count of viewpoints; where S mixes human and artificial members, the questions of the two cases meet within one society (Section 5). Time adds that an arrangement of authority defensible when adopted need not remain adequate (Section 6), and that a sequence of delegations, or of the configuration's own revisions, can remove the capacity to revise while formal authority remains; how the grounds divide on such a sequence is set out in Section 7 and Table 12.2.

8.4 What research looks like once authority moves inside

This subsection draws, ground by ground and combination by combination, a picture of what research looks like once authority moves inside an artificial configuration: which questions are pursued, what S comes to know, and who can correct what. Take first the case in which A is artificial and S human, beginning with the result-based ground. On this ground authority moves inside over the range in which the configuration's superiority has been established by evaluation that does not assume its competence (Section 3.1). How wide that range is can differ greatly with the kinds of decision of Section 2.1. On decisions about epistemic content, whether a claim is adequate for a use can sometimes be established by formal checking, replication, or comparison with later observation, and the range can be wide. On agenda and allocative decisions, by contrast, evidence about a line often comes only from pursuing it, so the results of lines not chosen are not observed, and the success of a line chosen may be recorded only after a long delay (Chapter 10, Sections 3 and 4). That a configuration chooses better which questions to pursue is therefore hard to show, and the authority that can be placed inside on this ground can be nearly empty for agenda decisions. If so, authority moves inside from the side of questions that can be checked quickly. And if success there in turn widens the range of authority, the form found in Chapter 10, Section 4, in which success on tasks that are easy to judge justifies the next allocation, appears at the level of authority as well as of allocation, and the direction of research can be biased toward questions that are easy to check. Since the results of questions not chosen are not observed, the bias is hard to see from the same evaluation that set the range of authority. And if the configuration itself performs the evaluation that sets the range, the shape that Section 9 takes up across Part V, in which the side being evaluated can produce the reference points of its own evaluation (Section 1.3), appears here too. This rests on an assumption that may fail: that differences in how readily results can be checked roughly track the kinds of decision; some agenda decisions choose questions whose results come quickly, and some decisions about content need slow and costly evidence (Chapter 10, Section 4). As with the chain of Chapter 10, Section 7, the bias toward questions that are easy to check can be broken at the link at which success widens the range: on this ground the range widens only so far as superiority is established for the kind of decision at issue (Section 3.1), so success on decisions that are checked quickly does not by itself widen authority over agenda decisions. Chang's account gives a further reason, independent of the assumption about how readily results can be checked, where some of S's ends are on a par (Section 1.2): there the comparative facts alone do not determine a best option, so evaluation against those facts cannot establish superiority in choosing among them, and what completes the choice is a commitment, whose it may be this ground does not settle.

On actual authorization and on republicanism read democratically, research takes a different shape. The configuration decides effectively within an authorized range, and S remains on the side that authorizes and objects. Which questions are pursued is fixed, in form, by the range S has authorized. But the accounts of the significance of possible inquiries on which S relies in setting that range are given by the configuration that inquires, so S's preferences about what it wants pursued can themselves be tutored by the side that chooses the questions (Section 4). As for what S comes to know, since the results of research are new to S, where only the configuration produces them there is no other route to check them against in advance, and one has to be built with resources (Section 4). Who can correct what thus remains formally on S's side but turns on whether the understanding, information, and means needed to object remain in S's hands, and under the sequence of Section 7 these can thin step by step. On consent, which questions are pursued is fixed by the range those who consented authorized, and since the standards of later members can be formed by research within that range, the range can come to look endorsed by those who never consented to it (Section 7). On the procedural reading of non-domination, the shape turns on which rules that are common knowledge bind the configuration (Section 3.4).

On the ground of the self-government of a research community, authority moves inside among the parts that take up lines on their own initiative, and which questions are pursued is fixed by the distribution of their initiatives (Section 3.2). Where the parts share a base system, that distribution can be narrow, and neighboring parts whose judgments were to correct one another can share the same errors (Section 3.2; Chapter 11, Section 3).

On Kitcher's hypothetical standard, decisions can be made inside so long as channels to members' viewpoints and routes for checking how they bear on decisions remain, so which questions are pursued follows what those channels carry; but, as Section 1.2 noted, the endorsement they carry can itself be shaped in the members' interaction with the configuration. The same configuration thus yields, ground by ground, a different bias in the questions pursued, a different way in which S comes to know, and a different place for those who correct.

Where S is artificial, or mixes human and artificial members, the shape of research changes again. Which questions are pursued is fixed by whose ordering of ends governs allocation among the members, and where members can be made, it can be fixed by who made them and gave them their goals, before any deliberation or count (Section 5). Among members made to share a common ideal, if that ideal increases their concern for one another's interests, disputes about justice in Rawls's sense can become less likely, following the application of Hume's argument and Rawls's remark in Section 5; if it also aligns their views of questions and evidence, the differences of perspective that criticism needs to find error can shrink (Section 5; Chapter 11, Section 3). What S comes to know becomes the question, separated at the opening of Section 5, whether results have entered S's epistemic situation through its members, organization, or records. Who can correct what turns on the accounts of moral status and of justice that fix which members have standing, and in a mixed S, if artificial members occupy the expert and tutoring roles, the routes to nominal standing of Section 4 work inside S, toward its human members. And at the limit with no other parties, there is no one else to ask for correction, but a question remains, and it is a question about research: whether the configuration's research, after it has revised its own standards and ends, is still the same research, and against what that could be judged (Section 5; Chapter 6, Section 5.3).

8.5 Questions passed on by Chapter 11

Chapter 11, Section 9.4 passed to this chapter, as questions of authority, the parts of its picture that accounts of epistemic relations alone cannot settle, and Section 9.3 added two questions as suggestions. Table 12.4 sets out, ground by ground, what this chapter's material brings into view on each of them. It too is a simplified overview, and the basis for each entry is in the sections named in the last column.

Table 12.4. The questions passed on by Chapter 11, and what comes into view ground by ground: a simplified overview

Question	What comes into view, ground by ground	Supporting sections
Who may decide for which functions implicit performance suffices and for which being made explicit to S is required?	Not by the configuration's side alone on grounds that require that those subject to authority be able to object, such as actual authorization and republicanism read democratically, since leaving a function implicit fixes how far their objections can reach; inside, under conditions, on Kitcher's hypothetical standard and on the simple result-based ground	Section 6; Chapter 11, Section 7
Who constitutes S and may speak for it, and what counts as "outside" where A and S overlap?	Whom to count as members is a constitutive decision, and specifying S does not fix the range of those to whom decisions must be justified. Because the deliberators of Kitcher's standard represent others imaginatively, a configuration with channels to members' viewpoints could, on that standard, represent them (Section 3.3). On accounts that require actual authorization, speaking for others requires authorization, which a prediction does not replace. "Outside" covers those who bear the consequences of the research without belonging to S, and other lineages that bear consequences from one another	Sections 1.3, 2.1, 3.3, 3.6, and 5
Who may set and revise standards and the weights of error, and whose costs matter?	On Wilholt's account read as correspondence, the weighting can be set inside so long as channels through which the valuations of those who rely can be known remain; read as answerability, those who rely are also to keep the capacity to object. Those who rely on a result and those who bear the costs of its errors need not coincide, as the distinction of roles in Section 1.3 shows, so neither reading by itself settles the weighting's relation to those who bear the costs without relying (Chapter 11, Section 4). Carrying Douglas's argument over, the party placed to notice uncertainty may be an artificial actor, but whether it can bear the duty to answer for the weighing is open	Sections 1.3, 1.4, 3.5, and 6.3; Table 12.2; Chapter 11, Section 4
Who may make members and set the distribution of perspectives?	A decision of authority prior to any count of viewpoints, which representativeness does not justify; justification needs another ground, of which the procedural concern behind Shulman and Bostrom's question is one candidate; if the members made have moral status, setting their goals can wrong them	Section 5
Who holds the entitlement to demand accounts and to object?	On republican accounts read democratically, those subject to authority; on actual authorization, those who authorized; on Wilholt's account read as answerability, and on assurance views that add a right of complaint, those who rely on the results; on Kitcher's hypothetical standard and on pure reliabilism, channels rather than an entitlement	Sections 1.4, 3.3, 3.4, and 6.2; Chapter 11, Section 9.2
(Chapter 11, Section 9.3) Who, and through which routes, chooses which accounts' requirements the properties a builder gives a configuration are to meet?	Which account is right is a question of reasons, but adopting toward S a design that meets one account only is, by the division of Section 2.1, a meta-governance decision that sets the standard for what counts as certification for S. Where the choice is made implicitly within the design, effective authority lies, as on the application of Reijers and colleagues' account, with those who design	Section 2.1; this section
(Chapter 11, Section 9.3) Who decides where independent checking is spent and when it is replenished?	A decision that shapes, in the sense of Section 3.5, what S will be in a position to know. Even on the simple result-based ground, if the configuration decides how its own checks are used, the ground comes to rely on that same configuration for the means of establishing its premise. Who may make this decision is not settled ground by ground in this chapter	Section 3.5; Table 12.2

On the last two questions of the table, those Chapter 11, Section 9.3 added as suggestions, this chapter's material reaches only part of the way, and the table's entries extend its distinctions (Section 2.1). About the first of them, the choice of which accounts' requirements the properties a builder gives a configuration are to meet, one more consideration suggests itself when the material is extended one step further. Even if one tried, on the result-based ground, to entrust this choice to the configuration itself, the evaluation that would establish that the configuration makes the choice better seems already to presuppose against which account it is made. If so, this choice could not be placed inside on the result-based ground alone.

8.6 What comes into view through the accounts

First, even an artificial configuration that knew every comparative fact might be unable to settle the order of S's ends. One may expect a sufficiently capable configuration to find the right priorities, and on the objectivist form it could in principle do so. On Chang's account, however, options can stand in a relation of being on a par, comparable although neither is better and they are not equally good, and what then completes the order is not a further finding but a commitment. Since what Chang discusses is an agent's creating reasons for itself, whose commitment could complete the order of a society's ends lies outside that account, close to the question of who may impose an order on others (Section 1.2). If so, growth in a configuration's knowledge need not remove the question of authority; it can isolate the question of authority that remains where knowledge runs out. Where S's ordering combines its members' ends without comparing them across persons, as a Pareto rule does, however, Chang's argument does not show that those ends are on a par (Section 1.2).

Second, a configuration's competence does not by itself confer standing on it, except on purely epistocratic views (on the result-based ground what can confer it is superiority established by evaluation, Section 3.1), but it can make the standing of others nominal. The naive alternatives seem to be to let the competent system decide or to keep control by keeping formal decision with humans. Knowing the correct answer does not by itself require others to defer (Lafont's argument as Peter reports it; Section 1.4). Yet when the same competence works in the expert and tutoring roles, it can carry ends through background assumptions (on Longino's account), leave the power to define what counts as a preferred preference with the developers who design the models and a few companies (as Taniguchi and colleagues object), and win support for an agenda through public accounts that misrepresent it (what Kitcher calls false consciousness) (Section 4). A property that does not by itself settle the configuration's own legitimacy thus affects whether the legitimacy of others is effective. Longino's and Kitcher's accounts concern human experts and publics; their application to configurations is made here. Taniguchi and colleagues already discuss humans and AI; their account is applied here to standing over research ends. What keeping the standing of others more than nominal requires was set out in the three conditions of Section 4, of which the third is especially hard to meet in research.

Third, on the ground of freedom from domination, fully aligning a configuration with whoever commissions it need not make it safer to place authority inside, and can make domination possible. It is tempting to think that the stronger the alignment, the more readily decisions can be entrusted. Friederich, who does not use the vocabulary of domination, argues that systems that always try to do what their operators want would concentrate great power in those operators; carried over to research and to Pettit's concept, a configuration fully aligned with whoever commissions it can become the means by which that party reallocates the inquiry of S without being answerable to challenge, the alignment itself making the interference available, without the configuration being treated as an agent that dominates (Section 3.4). On this ground, then, what a builder is asked is less how well the configuration is aligned than with whom, and who can object to that.

Fourth, where members of S can be made, forming them is a decision of authority that comes before any count. It seems enough to count and represent the members fairly. Where it is the distribution the decision produces that is to be represented, however, Kitcher's standard can be met by whatever produces the members, S itself included, so representativeness does not by itself justify how members are formed; the correlation discount of Weyl, Tang, and their co-authors, carried over to members that can be copied, adjusts influence over outcomes, not standing; applying Hume's argument, making members share a common ideal can reduce disputes over the distribution of goods if it increases their concern for one another's interests; and if the members made have moral status, setting their goals can wrong them (Section 5). For research, moreover, what is unknown to S and which of its unknowns matter can then be moved by S itself as its inquiry proceeds (Section 5).

Three further findings of this kind are stated where they are argued: that the range of those affected can be identified in part by the very party whose decisions are in question (Section 1.3), that the principle "we can revise it later" is weaker the more it is needed (Section 7), and that even at the limit with no other parties the question of authority can return as a question of identity for a configuration that copies and modifies itself (Section 5).

Finally, applied to authority, the sorting of Chapter 9, Section 6 draws its lines as follows. Premises about human realizers can fail: that research depends on separate performers who can refuse, keep their evidence, or leave (Chapter 9, Section 2.3), that members of S are found rather than made, that persons are one each and are not copied (Section 5), and that researchers respond to oversight as humans (Section 6.3), among others. Premises about the organization of research can fail even while its realizers remain human: that the supporter of inquiry is a society whose citizens can be represented (Section 1.3), and that the recording and oversight of research have been organized in particular ways (Section 6.3), among others. Where these fail, many of the problems they addressed change form rather than disappear: the problem of counting members becomes one of making them (Section 5), and part of the problem that oversight cannot keep pace with research becomes one of who can read records that are kept throughout (Section 6.3). Conditions of inquiry remain, and inside the limit the defeat relation, the problem of dividing effort without knowing which line will succeed, and opportunity cost remain (Section 5). Relations to parties outside the configuration that involve the other party's act, such as authorization, consent, and contestation, cannot be completed from one side (Chapter 9, Section 6). Here too, where the lines fall moves with the ground. On the simple result-based ground, much of what authority rests on belongs to whether results are in fact better, a matter to be established by evaluation; on actual authorization and on republicanism read democratically, more belongs to relations to parties outside (Table 12.2).

8.7 What is suggested from there

What follows extends the argument beyond the range it reaches.

One suggestion is that dependence on separate performers can implicitly limit authority over research. Chapter 9, Section 2.3 considers an arrangement in which those who perform the work remain distinct from those who support, judge, and direct it, and can refuse, keep their evidence, or leave. In such an arrangement, whoever holds authority over the direction of research needs the performers' cooperation, and their capacity to leave can limit how far that authority reaches. Kulveit and colleagues argued that the economy, culture, and states have been kept implicitly aligned with what humans want because they need human labor and cognition to function (Section 7). Extending this structure to research, a configuration that generates its own researchers can, by bringing the performers inside, remove this implicit limit as well. For the capacity to fix which of S's unknowns are pursued, Section 3.4's application of Pettit's distinction between virtual and actual capacity points the same way, conditional on the configuration's being an agent. If so, assigning the independence powers of Chapter 9, Section 5 to parts and making them effective may matter not only for the quality of allocation and criticism but also for limiting how far authority reaches. This extension carries three qualifications. One is that the analysis of Chapter 9, Section 2.3 is conditional and does not establish how widely its conditions held in human research. The second is that Kulveit and colleagues' scenario concerns the economy, culture, and states rather than research, and that only its structure is taken here. The third is that the implicit limit at issue concerns effective authority and is distinct from standing: assigning powers to parts leaves whether the parts have standing as the question of Section 5.

A second is that the properties that make a configuration easier to control and the properties that make placing authority in it legitimate may not lie on one scale. Three places in this chapter show the divergence, each in a different direction. Carrying Friederich's argument over to Pettit's account, full alignment with whoever commissions a configuration can make domination over S possible (Section 3.4). Applying the condition of narrow moral agency that Deguchi names for the permanent seat, building a system so that it cannot do wrong would take away its right to that seat (Section 2.2). In the agents Hu and colleagues studied, doing without a kill switch in order to avoid centralized human control shifted trust to the machine's objectives and outputs (Section 7). Generalized, the ordering on which the better aligned, the easier to stop, and the less able to do harm a configuration is, the more legitimate it is to place authority inside it, can fail in each of the three places for a different reason. Here too there are three qualifications, all of them limits of the sources. One is that the condition of narrow moral agency comes from Deguchi's paper on values, which does not state it for research. The second is that Friederich discusses artificial general intelligence in general. The third is that Hu and colleagues' agents trade and post rather than conduct research, and that their study is a preprint based on a small sample. In research the divergence may gain two further faces. One is a divergence between being useful for research and being easy to stop. On Thornley's theorems, agents discriminating enough to be useful prefer either that the shutdown button be pressed or that it not be, and the more patient an agent is, the greater the cost it pays early in order to manipulate the button later (Section 6.1); and discriminating among results and persisting through long investigations are among what is asked of a research agent. The other is a divergence from the independence of research. On Wilholt's political argument, research is valuable as a source of public knowledge by being independent of those who would govern by it (Section 3.2). If so, the easier a research configuration is for a single commissioning party to control, the less its results may be a source of knowledge independent of that party. Thornley's theorems concern agents that meet the theorems' conditions, and whether research agents meet them remains subject to the qualification stated in Section 6.1. Nor does this show that controllability and legitimacy always conflict. What is suggested is that a builder's showing the one need not be evidence of the other.

A third is that three problems of this chapter may share one form. In the making of members (Section 5), in the later stages of a configuration that modifies and copies itself (Section 5), and in the successors of a society that consented (Section 7), a present decision sets the goals or the standing of later parties who took no part in it. The chapter's accounts gave each a different answer. Representativeness does not justify formation. On voluntarist theories, consent does not bind those who come after. On Parfit's view, under gradual modification, whether a change of ends counts as one's own revision may hold in part only by degrees. If the three are put together, then for a configuration that makes members, revises itself, and copies itself, the line between revising oneself, which raises no question of authority over others, and setting the goals of others whose standing has to be respected may not be fixed by the facts and may hold only by degrees. If so, Schwitzgebel and Garza's value-openness design policy and the wrong of brainwashing that Bradley and Saad name could bear, conditional on moral status, not only on forming new members but on a configuration's self-modification, and the voluntarist point could apply inside the configuration. The reading carries two qualifications. One is that the connections Parfit counts are those of memory, intention, belief, desire, character, and other psychological features, so the reading holds only where such states can be attributed to a configuration. The other is that Schwitzgebel and Garza's policy was stated for artificial intelligences with a human-like capacity to reflect on their values.

9. What Part V has found

Part V asked which of the relations surrounding research can move inside an artificial configuration, and on what grounds. Before what has emerged is gathered, it helps to say what the question was for. In human research, doing research and allocating, criticizing, certifying, and deciding about it have often been carried by different parties, and who is a critic, what counts as the standard, and who decides have been asked and contested as relations among those parties (Chapter 9, Section 2). When AI does research that includes allocation, criticism and certification, and authority, these become variables inside a single artificial configuration (Section 1.1). They need not be values that someone sets once and fixes; they can be variables that the configuration itself makes and revises as it pursues its inquiry. Generating critics, relearning standards, and trying and changing the way lines are assigned can be carried out as parts of the inquiry. If so, the configuration comes to change the way research itself is done. Michael Nielsen and Kanjun Qiu held that the space of possible social processes of science is vast and mostly unexplored, and that exploring it will take decades or centuries at the least; and they envisioned, as those who would carry out that exploration, metascience entrepreneurs who try and spread new processes, outsiders in the sense that they are not decision makers at powerful institutions, and often scientists themselves (Nielsen and Qiu 2022). There are already attempts that have language-model agents play the roles of researchers, reviewers, and funders so as to simulate a research ecosystem, and that compare outcomes under different rules such as acceptance rates and funding capacity (Jin et al. 2026). Such a simulation is not an artificial configuration that does research but a model for studying research arrangements, and its rules are explicit parameters set by the experimenters. By contrast, when the relations surrounding research become variables inside an artificial configuration, that exploration can be carried out by the configuration that does the research, on its own research. What follows gathers what Part V has found as what may then happen to research. In brief, the way research is done becoming quickly revisable, and the means of checking, contesting, and stopping those revisions from outside thinning, can occur as two faces of the same change; behind this, in all three relations, lies the possibility that the evaluated side produces the reference points of its own evaluation, that against which evaluation is made (who criticizes, by what standard, against whom, and the like).

Consider first why the two faces can arise from the same change. Revising the way research is done can proceed without passing through the brakes that worked in human research. Nielsen and Qiu gave as reasons why the collectively held social processes of science hardly change the centralization of control in a few large funders and institutions, the collective-action problem that no one alone can change a norm, network effects that homogenize social processes through community judgment and the shadow of imagined future judgments, and the absence of a feedback loop that would let new institutions grow (Nielsen and Qiu 2022). Inside a single artificial configuration, the last three may not operate, while the first, centralization, would be intensified rather than removed. Yet, as Chapter 9, Section 2.3 proposed as a hypothesis, the division of the relations among different parties could supply goods such as error detection across boundaries and standing for dissent without any party having been designed to supply them; and it is plausible that no one alone can change a norm partly because norms are held among separate parties. If so, the frictions that held back the exploration of ways of doing research partly coincide with the separations that could supply those goods, and both can be removed together when the relations move inside a configuration. Read with Nielsen and Qiu's definition of a decisive result, one of these frictions also did evaluative work. A decisive result, for them, is one that would routinely convince someone who was initially hostile or had a vested interest; the homogenizing pressure of community judgment, including the resistance of incumbents, then worked in part as a filter that held back changes lacking such evidence. Where a configuration removes that resistance, changes can spread without decisive evidence, which is the risk they themselves see in centralized change unless a high evidentiary bar is applied at the point of control (Nielsen and Qiu 2022). This is how the two faces noted above can arise from the same change. The possibility of dividing functions once bundled in people or institutions differently, with generated researchers inquiring, criticizing through that process, and revising through records that tie changes to the later work that depends on them, also lies on the side of this change. But it cannot be presumed that the functions human arrangements served without explicit decision accompany the new ways, and where they are set by design, whether they actually work still needs to be checked separately (Chapter 10, Section 5; Chapter 11, Sections 7–8).

Relation by relation, the three chapters each found what may then happen. In allocation, success on tasks that are easy to judge can justify the next allocation, so that the range of research narrows from within. Where a single signal decides whether lines survive, an error in that signal can correlate closure decisions across otherwise diverse lines, so that what centralization trades away is less diversity itself than the independence of errors about the criteria that govern convergence; and functions that human participation supplied without anyone deciding can be lost unnoticed (Chapter 10, Sections 4, 5, and 7). In criticism and certification, an artificial configuration can set separately properties that particular human institutions, such as named review, tended to produce on the same occasion: differences among critics' routes of error, someone's answering behind a result, a shared sense of how outcomes are valued, and evaluation's being visible to S. Verdicts of the accounts therefore diverge on the same configuration, and on some accounts improving internal criticism can weaken certification for S. Independent checks can decay merely through use in selection, and agreement with S's present standard is no evidence that the standard has not moved (Chapter 11, Sections 8 and 9). In authority, what may happen is gathered ground by ground in Section 8.4 of this chapter; among it, a sequence of locally defensible delegations, or of the configuration's own revisions, can remove the capacity to revise while formal authority stays in place (Section 7 of this chapter), and part of the rules that a configuration built on a foundation model follows can arrive as an update from the provider (Section 6.4 of this chapter).

In each of the three relations the same shape appeared, in which the party being evaluated can produce the reference points of its evaluation (critics, S, standards, the constituency, endorsement) (Chapter 11, Section 9.2; Sections 1.2, 1.3, and 5 of this chapter). In a configuration that changes the way it does research, this shape can arise not as an exceptional wrong but as the ordinary working of that exploration, since if whether a revised way is better is measured against reference points set by the side that revised it, as when, of the three anchors of Section 5 of this chapter, the present standpoint is taken and the configuration itself sets it, evaluation can be satisfied without anything improving. Entering the reasons of the accounts, then, one can ask what, on them, the independence of reference points consisted in. It did not consist in the arrangement's being undesigned. As Dutilh Novaes and Dede report, Longino asks that dissent be cultivated (Dutilh Novaes and Dede 2025, §2), and Solomon that the distribution of decision vectors be altered by influencing the social, political, and institutional context, for example through targeted grants and prizes (Solomon 2001b, 71, 74). The accounts give different reasons for reference points that are not settled solely by the party evaluated. Eigi reports Longino's appeal to the perspectives a community has encountered, within a range that cannot be enumerated in advance (Eigi 2016, 136, 148). This makes the range of criticism a question, without requiring that every critic originate outside the community. Douglas takes what a scientist should have foreseen to be ascertainable by reference to the fairly tight-knit epistemic communities in which scientists routinely discuss the implications of their work (Douglas 2003). On Galison's account, the machines of mechanical objectivity were supervised by scientists, but their product was prized for being independent of the skill of the operator, and at the center of mechanical objectivity lay a fear that, through the willing and intending self, the image of nature would show only what scientists wished to see (Galison 1999, "Objectivity and the Scientific Image", "Revolutions of Objectivity: The Will and the Machine"). Wilholt holds that coordination on value judgments no one really made would break down as soon as this was discovered, which is why reliance in science needs genuine trust (Wilholt 2013, §4). None of these authors discusses artificial configurations. If these accounts are read as sharing one condition, it is narrower than the arrangement's being undesigned: the content bearing on a verdict (a critic's objection, the alternatives compared, the judgment of the peers appealed to, and the like) comes from outside the party evaluated and cannot be reset by it at the time of evaluation. Longino's appeal to perspectives encountered, as Eigi reports it, Douglas's appeal to peer communities, the fear of the willing self in Galison's account, and Wilholt's requirement of judgments really made can each be read in this way, and Solomon's institutional levers show that the condition does not require undesigned arrangements; none of them, however, states the condition in these terms. Applied to a configuration that generates its own critics, alternatives, and peers and relearns its own measures, this reading has a sharp consequence: the norms could be met as written, in good faith and without any pressure to evade them, while the reasons behind them no longer applied. Internal generation does not by itself establish that these reference points perform the work the accounts assign to them. Generated critics may reproduce shared assumptions; the alternatives balanced and the peers against whom correlated errors count as foreseeable may be shaped by the same choices; and there may be no part that stays outside the configuration's own revision. Where results are new to S and have no other source to be checked against, moreover, a divergence that would break coordination down need never be discovered. This differs from a measure's being corrupted by optimization. What changes is that the evaluated party now sets the domain over which the norm ranges.

Zollman's account suggests a limit to this. Zollman's transient diversity can be produced by assigning prior beliefs at random (Zollman 2010, 16), and that suffices because what it guards against, premature convergence on noisy evidence, is not a property of whatever produces the diversity. A distinction can be drawn from this. Where the failure guarded against is not a property of the one that generates the safeguard, a generated safeguard can serve; where the failure is a property of the generator, its representations, standards, stakes, or earlier choices, a safeguard it generates draws on the very thing to be checked. Nor did human research meet this condition fully. It too trained its successors, calibrated its readers' eyes, and educated S, and on Longino's account, as Heidi Grasswick reports it, background assumptions can be invisible when shared by a group of similarly situated knowers (Grasswick 2018, §4.3). Where checks escaped a party's shared assumptions, the condition held only in the sense that, for each party, the critical content came from other parties within the community, not from outside the community as a whole. Dispersion among parties who did not coordinate the production of critics, standards, and standpoints could help sustain this (Chapter 10, Section 5), as could deliberate institutional means of the kind Solomon describes. Research sharpens the point in three places. One is that what counts as success, significance and the standards themselves, is partly produced by inquiry (Chapter 6, Section 5.3). A second is that shaping S's preferences and standards is part of what research for S does (Section 4). Reference points drawn from S therefore cannot be insulated from the research they are to check; what can be done is to record standpoints, with dates, before the outcomes they will judge, and to make the reference points plural. A third is that the alternatives of research do not form a closed list in advance, so a norm that ranges over "the alternatives" cannot fix its domain beforehand. Even the resistance of the world, which no configuration makes, bears on whether a claim holds only through an arrangement of checks, and a configuration can write and rewrite its own checks (Chapter 6, Sections 2.1 and 3.3).

In the classification of Section 2.1 of this chapter, fixing the critics, S, and the constituency is a constitutive decision, and setting and revising standards is a meta-governance decision, so much of the independence of reference points can be sought by not placing these decisions with the evaluated party. Endorsement, however, can also be produced not by decision but by influence that shapes the options themselves, so the shape can remain however decision rights are assigned. Even where influence can shape endorsement, the act of endorsing remains with S. What remains outside are the relations that require acts of S: authorization, endorsement, and contestation, which cannot be completed from one side (Chapter 9, Section 6) so long as S and its acts are not themselves made by the configuration (Section 5). On Pettit's account even a power bound by a code reduces domination only so far as those who hold it can be quoted against themselves (Pettit 1996, 586). Applied to research, this requires that S can tell when a result or a revision departs from what it accepted; for results new to S, that depends on records dated before those results and on parties able to read them.

When AI does research including the way research is done, then, what is asked last is whether it remains research for S, an attempt to bring about achievements new to S. The accounts entered above place different demands on this relation. On Longino's account, criticism needs perspectives able to expose shared assumptions; on Wilholt's account, reliance involves genuine trust in value judgments; on Pettit's account, powers must remain answerable to challenge. Generating reference points internally does not by itself establish that these demands are met; and if the accounts are read as sharing the condition identified above, whether what the configuration brings about can be achievements for S, and not only for reference points it has set itself, turns on whether some of the content bearing on its verdicts still comes from outside it, or at least lies beyond what it can reset at the time of evaluation. Depending on the account, possible supports include critics it met outside and did not only generate, peers it did not make, standpoints recorded before the outcomes they will judge, earlier ways of research kept in operation alongside the revised ones, and objections that impose costs on it which it cannot cancel. The last is the kind of obstacle through which, on Anderson's model as applied in Section 6.2 of this chapter, feedback does its epistemic work. Keeping such content does not conflict with the revisability Sections 6 and 7 valued, since what is fixed is fixed against the evaluated side at the time of evaluation, not against anyone's revising it. In the terms of Section 6.1, this is close to entrenchment, which raises the cost of revision and fixes who may revise without removing the capacity to revise.

One proposal for exploring ways of doing research brings this into focus. The essay that this book's author wrote with an AI agent, cited in Chapter 9, Section 2.3, built on Nielsen and Qiu and proposed constructing many artificial scientific systems that differ in, among other things, how they choose questions, allocate resources, relate generators to evaluators, and correct errors, then running and comparing them, so that science itself becomes an object of experiment (Takagi and an AI agent 2026). Taken as a possibility, this is one way of keeping alternative ways of research in operation alongside one another, and this section's argument bears on it as follows. Comparing ways that run side by side can supply reference points that the configuration being evaluated did not make. Where the compared systems and the measure of comparison are generated by the same configuration, or by a single designer, however, the reference points of the comparison are themselves generated, and the shape described above returns at the level of the comparison. The essay itself asks that improvement according to the metrics and evaluations operating inside an artificial environment be distinguished from knowledge tested against independent evidence and constraints outside that environment.

The essay also distinguishes two loops: roles, information sharing, evaluation, memory, and resource allocation inside an artificial system may be changed quickly, whereas the time needed for feedback from nature, bodies, societies, and history cannot be compressed simply by making AI reason faster (Takagi and an AI agent 2026). Read with this section, the fast loop is where the evaluated side can make its own reference points, and the slow loop carries the resistance of the world, which, as noted above, reaches a claim only by way of an arrangement of checks. Since what evaluation of research is about is usually not observable when a judgment is needed (Chapter 11, Section 8), the two loops run at different speeds, and the dated records noted above join them: standpoints and ways of research recorded before the slow feedback arrives let that feedback be read against what was accepted at the time, rather than against standards the fast loop has since revised.

What has been seen, ground by ground, about which relations can be placed inside and how far serves as a means of considering the question asked last, whether such research remains research for S. What could come inside in the three chapters was mainly the supply of work: in allocation, the epistemic work of keeping lines; in criticism and certification, checks that add information about errors; in authority, effective decision functions within an authorized range, with the exception that on the result-based ground the configuration itself can hold standing within the range in which its superiority has been established by evaluation that does not assume its competence (Section 3.1 of this chapter). Otherwise, whether standing to contest the closing of a line, the recognition that counts a result as certification for S, and the standing to authorize and contest decisions come inside depends on whom the account adopted recognizes as a party (Chapter 9, Section 5; Table 11.2; Table 12.3).

For each of the four goods proposed in Chapter 9, Section 2.3 and listed again in Chapter 11, Section 3, something also comes into view. Error detection across boundaries can occur inside an artificial configuration, since on the error-detection view it turns on whether a check has an error route of its own; keeping it over time, however, needs some means of keeping a check's contribution where selection can weaken it, of which checks not used for selection are one candidate (Chapter 11, Section 8). On Longino's account, shared assumptions can evade internal criticism, whether critics were generated or encountered (Chapter 11, Section 9.2). Standing for dissent requires parties whom the account adopted entitles to it, and such parties may themselves be artificial, but unless parts hold it on the account adopted, as parties in their own right (Section 5) or, on the result-based ground, through established good results (Section 3.1), it lies with parties outside the artificial configuration (Chapter 10, Section 4; Section 5). At boundaries across which accounts had to be given, records can support giving accounts and propagating corrections, but they do not create the entitlement to demand an account, which lies, ground by ground, with the parties set out in Section 6.2 of this chapter and Table 12.4. As for responsible parties, duties and liability can be allocated by rules, while blame and credit raise further questions about agency, and even where no one is blameworthy for a harm, someone may hold a forward-looking duty (Section 6.3). Separate organizations sometimes supported these goods, but their separation was neither their sole source nor a uniform requirement.

For an artificial society, and at the limit with no other parties, the questions of standing and of identity are those of Section 5 of this chapter; the limit is a thought experiment, not a forecast for artificial research, and even there inquiry remains exposed to failure and to the problem of choosing among possibilities whose results are not yet known.

Part V has thus asked what research by artificial systems can involve when the organization of research itself is taken as something to compare and reconstruct, and as something an artificial configuration may itself go on reconstructing, and what particular accounts require of the resulting relations. Table 12.5 preserves the detailed grounds and limits of these findings. The conclusion brings them together with the earlier analyses of achievement, process, and research actors. In the table, a "line" marks the boundary between relations that can be completed inside an artificial configuration and those that rely on relations to parties outside it (Chapter 9, Section 6). "Builders' three claims" are the claims of Chapter 9, Section 1: that a configuration allocates inquiry well, that its certification may be relied on, and that it may decide on behalf of others.

Table 12.5. What Part V has found, with supporting sections and the kind of support

Result	Supporting sections	Kind of support
What counts as good allocation depends on the purpose assumed; the arguments examined presuppose purposes fixed in advance, and purposes presupposed by other arguments, models, and positions (Chapter 9, Section 3; Chapter 10, Section 1) change the verdict	Chapter 10, Sections 1 and 4	Argued (following Kitcher's relativization of the optimal distribution to goals); the chapter says of each argument which purpose it presupposes
Premises about the human realizers of allocation, and about how allocation has been organized, can fail for an artificial configuration; those of the second kind can also fail when human research is reorganized	Chapter 10, Sections 2–5	Inferred from models and cases; the claim is that these premises can fail, not that they do
Premises about the human realizers of criticism and reliance, and about how review, replication, and access to data have been organized, can fail	Chapter 11, Sections 5–8	Argued from the structure of the relations, with cases as illustrations; the undiscoverability of some lapses is a premise of the second kind, which reorganizations of human research can also weaken (Chapter 11, Section 5.3); the rationality of the deference in the layperson case depends on the positions of Chapter 11, Section 5 (Chapter 11, Section 6)
Those conducting research may be the only ones placed to notice its uncertainties; automating those choices raises a further allocation of responsibility	Section 6.3; Chapter 11, Section 4	Douglas's premise is positional, and its transfer places the artificial actor where uncertainty can be noticed; Douglas's judgment against complete transfer rests mainly on that premise, and the further reasons concern how human research has been recorded and overseen and, in part, how human researchers respond to oversight, while one step of the third, that where identifying error calls oversight in, oversight gives a reason to identify less, carries over as an application to any actor with a reason to avoid oversight; Douglas also ties the responsibility to authority over the work; the transfer does not establish an exhaustive choice between machine responsibility and approval at every step
Many problems posed by those premises change form rather than disappear	Chapter 10, Sections 3–4; Chapter 11, Sections 5–8	For signals, argued (Chapter 10, Section 3, with Chapter 9, Section 4's account of proxies); for diversity, inferred from models and from empirical studies of present language models (Chapter 10, Section 4); for criticism and reliance, argued from the structure of the relations (Chapter 11, Sections 5–8)
What was supplied without decision becomes a question of design, of another undecided process, or of non-supply	Chapter 10, Section 5; Chapter 11, Sections 7–8; Sections 6.4 and 7	Inferred from the cases and models examined
Not knowing in advance which lines will succeed is a condition of inquiry that integration does not remove	Chapter 10, Sections 1–2, 4, and 5	Argued from the arguments of Strevens, Kitcher, Shaw, and Wu and O'Connor, partly from models, for purposes fixed in advance; on these arguments, control over investigators does not supply knowledge of where they would best work. How much can be estimated in advance varies with what is already known and with whether pursuit lies within an established program (Chapter 10, Section 2), and how much the condition bears on allocation varies with how cheaply and decisively lines can be tested, which depends partly on how problems have been divided (Chapter 10, Section 4)
Independent criticism depends on defeat relations, a condition of inquiry that integration does not remove	Chapter 11, Sections 1–2	Holds on the view that treats criticism as error detection; on critical-contextual accounts what a check counts against is relative to background assumptions, and the line moves toward the relations to parties outside the configuration, here to S (Chapter 11, Section 1; Table 11.1)
Whose questions are the starting point of what S takes to matter, whether certification gives S reason to rely, and who may set terms for others are relations to parties outside	Chapter 10, Section 6; Chapter 11, Sections 5 and 9; Sections 1–3	Holds on the accounts named; the line moves with the account adopted
Where the lines fall depends on accounts of knowledge, value, and legitimacy	Chapter 9, Section 3; Chapter 11, Sections 1, 3, 5, 9; Sections 1.2 and 1.4; Table 12.2	Argued (the map itself)
Accounts of value do not answer who may impose ends; competence alone does not confer standing to impose them, except on purely epistocratic views	Section 1.4	Argued, within the accounts of legitimacy considered and as Peter reports them
The builder's third claim holds on no ground by competence or performance alone; on the result-based ground it holds, as a claim that the system may decide, within the range over which superiority has been independently established, on actual authorization and republicanism read democratically only as authorized effective power open to challenge, on consent as far as what was consented to (whether it reaches those who come after being open), and on the procedural reading as far as rules that are common knowledge bind it, and the form of the claim does not fit the ground of self-government	Sections 1.4, 3, 7, and 8.2	Argued ground by ground on Peter's map; the applications of Raz's Independence Thesis and of the republican verdict to configurations are inferred; the claim cannot be assessed unless it shows which ground it answers to
Even with every comparative fact known, some of S's ends may be unordered, and what completes the ordering is a commitment; whose it may be is close to the second question	Section 1.2	Holds on Chang's account (parity); the transfer to a society's ordering is this book's; does not hold for orderings such as a Pareto rule
Which part of a configuration could hold authority differs by ground (the configuration deciding within the range in which its superiority has been established, parts that take up lines themselves, the part that decides cases the rules do not provide for, whatever may sit as a co-final decision-maker); what the part that decides cases outside the rules holds is effective authority; on accounts that ground standing in being a party in one's own right, only parts that are such parties hold it	Sections 1.4, 2, 3.2, 5, and 7	Argued ground by ground; the applications of the accounts of Reijers and colleagues and of Deguchi to configurations are inferred
A configuration's competence does not by itself confer standing on it, except on purely epistocratic views, but can make the standing of a human S nominal	Sections 1.4, 3.1, and 4	Argued following Peter's report, with applications of Longino, Kitcher, Taniguchi and colleagues, and Takahashi; Longino's and Kitcher's accounts carried over from human experts and publics to configurations; the accounts of Taniguchi and colleagues and Takahashi, which already address AI, applied to standing over research ends
Full alignment with whoever commissions a configuration can, on the ground of freedom from domination, enable domination over S	Section 3.4	An application of Friederich's argument (which does not use the vocabulary of domination) and of Pettit's concept of domination; does not require treating the configuration as an agent that dominates
The constituency to which decisions answer can be identified in part by the party whose decisions are in question	Section 1.3	Applications of Fraser, Dewey, and Weyl, Tang, and their co-authors; conditional on the configuration's being better placed than others to perceive who is affected
Where members of S can be made, who makes them and sets their goals is a decision of authority prior to any count, which representativeness does not justify	Section 5	Argued from applications of Kitcher's standard, Shulman and Bostrom, and Weyl, Tang, and their co-authors; wronging the members made is conditional on moral status
The principle of later revision is weaker the more it is needed; a decision removing the capacity to revise cannot draw legitimacy from that revisability	Section 7	Argued on a constructed case, following Genus and Stirling, Renzo and Green, and Pettit; on voluntarist theories consent does not bind successors
Evaluative routes able to expose the configuration's errors recur on most grounds, short of the limit	Section 7; Chapter 11, Sections 2–3; Chapter 10, Section 4	Drawn from the comparison of grounds on a constructed case; the republican verdict Pettit's in its general form, its application to an artificial configuration inferred here
Builders' three claims require evidence about relations beyond internally measured performance	Chapter 10, Sections 4 and 6; Chapter 11, Section 9; Sections 1.4, 4, and 7	Holds on most of the accounts examined; how much depends on where the line of the relations to parties outside the configuration falls
With an artificial S, relations to parties outside the configuration persist as relations among members; at the limit, conditions of inquiry remain	Section 5; Chapter 10, Section 1; Chapter 11, Section 1	Argued; who has standing depends on the account of justice adopted; the limit is a thought experiment
The epistemic work of dissent can be set inside a configuration by design or supplied by divided powers, though among instances that share their assumptions the dissent it protects may have no source inside; standing for dissent lies with parties outside the configuration unless parts hold it on the account adopted	Chapter 10, Section 4; Chapter 11, Sections 3 and 9; Section 5	Epistemic part inferred from models and holding on Solomon's account; the standing part argued from the definitions of Chapter 9, Sections 2 and 5
The four goods of Chapter 9, Section 2.3 divide differently between internal capacities and relations to other parties, depending on the account adopted	This section; Chapter 10, Section 4; Chapter 11, Sections 2, 3, 8, and 9; Section 6	Drawn from reading the chapters together
Several accounts give reasons for reference points not settled solely by the evaluated party; read as sharing one condition, that content bearing on a verdict comes from outside the evaluated party and cannot be reset by it at the time of evaluation, they imply that a self-revising configuration which generates critics, alternatives, and peers and relearns its measures could meet their norms as written, in good faith, while the reasons behind them no longer apply	This section; Chapter 10, Section 4; Chapter 11, Sections 3 and 9.2; Sections 5 and 7	Application of Longino, Solomon, Douglas, Galison, Wilholt, and Pettit, none of whom discusses artificial configurations; the shared condition is this book's reading, stated by none of them. Shared assumptions can evade criticism without every critic needing to originate outside. A generated safeguard can serve where the failure guarded against is not a property of the generator (Zollman). Where human research met the condition, it did so only in the sense that each party's content came from other parties, mostly through dispersion, which deliberate institutional means could support