Chapter 13 — Conclusion: What It Means for a Machine to Do Research

Research, Artificial Actors, and the Organization of Inquiry

1. Four questions

This book has considered what it is for a machine to do research, and what may happen when machines do. As a point of departure, Chapter 1 characterized research, relative to a society S, as an attempt to bring about epistemic achievements new to S.

Taking research this way, one comes first to ask what an epistemic achievement is, and when and where it is realized. Chapters 2–3 considered this for knowledge and understanding, and asked when what a machine produces counts as such an achievement. Moreover, since research that ended in failure, or that is not yet finished, is also called research, it would seem that an activity can be called research without necessarily arriving at a given epistemic achievement. That is why research was characterized as an attempt rather than as success itself. One therefore comes to ask how the attempt proceeds when no answer is yet available and failure is possible: what it means for a machine to ask, how candidates and representations become available, how it is judged what a result bears on, and how standards come to be revised. Chapters 4–6 took this up, and from there explored where the difference between producing outputs and inquiring lies.

In addition, Chapter 1's characterization does not explicitly include the one who does the research, A, so that it can cover both humans and machines. The one who does the research, A, and the society S for which the achievements of research are new can therefore be treated separately. Chapter 7 considered A. How far does the one who does the research extend? For humans and machines alike, one asks within which boundary research is to be attributed. One also asks what state it is for a machine to do research autonomously, and in which respects and how far it is autonomous. Chapter 8 considered S. If the one who does the research, A, and S need not coincide, whose knowledge do the achievements of research become? Chapter 8 took up this question by distinguishing the cases in which S is a human society, a society in which humans and machines are mixed, and a society of machines. A machine that does research may also differ from humans in its body, its cognition, how it is made, and the like. If so, what it takes to be not yet known, worth pursuing, or a good reason may come apart from what the members of S take them to be. Chapter 8 also considered where such a mismatch can come from, and how it resembles the AI alignment problem. And so long as the achievement is to be new to S, research stands in relations with S: because S's unknowns are countless while the means of inquiry are finite, there is allocation; because its members receive many results through others, there are criticism and certification; and because S has many members without a single voice and the consequences of inquiry can reach beyond S, there is authority over both (Chapter 1, Section 9). What, then, happens to research when machines take on these relations, and go on to make and revise them (Chapters 9–12)?

After the chapters' examination, the claim that a machine is doing research comes into view not as one claim but as a bundle of claims, each of which may hold or fail. These claims may bear on one another, but settling one need not settle the others. That a machine produces a correct result is one thing; that the result becomes knowledge for S may be another (Chapter 2). Producing outputs is one thing; inquiring, by locating where a failure lies and revising standards where needed, may be another (Chapter 6). Even when a machine conducted the research, it need not be the bearer of the achievement (Chapter 1, Section 9; Chapter 7). And when machines take on the relations surrounding research, whether it remains research for S is a further question again (Chapter 12, Section 9). That is why the book has taken up the four questions in separate chapters.

Table 13.1 gathers part of that bundle. It is an application of the argument, not an alternative definition of research. Each row pairs something that evidence can establish with a question that it does not thereby settle, and the last column names where the book argues that the two come apart. What a claim of each kind may need to construct, disclose, or show is set out where such claims are examined, for example in Table 7.1, Chapter 10, Section 7, Chapter 11, Section 9.2, and Chapter 12, Section 8.2.

Table 13.1. Inferences that require further argument: a selection

Evidence establishes	Further question	Principal discussion
A true or novel output	Whether it is knowledge, who, if anyone, knows it, and whether it has entered the relevant society's epistemic situation	Chapter 2
Knowledge of a regularity	Whether anyone understands why it holds	Chapter 3
A research episode	Whether a capacity persists across relevant variation	Chapter 7, Section 3
A capacity at one boundary	Whether researcher status is explanatory there, and for which purpose	Chapter 7, Section 3
Autonomy in one respect	Whether it extends to questions, methods, or standards	Chapter 7, Section 4
Interacting artificial agents	Whether they form a knowing society	Chapter 8, Section 1
Control of allocation	Whether its choices serve the purposes for which inquiry is organized	Chapter 10
Numerous or initially informative critics	Whether criticism supplies relevant information and continues to do so	Chapter 11, Sections 2–3 and 8
Effective internal criticism	Whether certification gives the relevant society reason to rely	Chapter 11, Sections 5 and 9
Formal rights or competent decisions	Whether authority is warranted on the ground adopted, and, on grounds that require it, whether it can be effectively contested or revised	Chapter 12

The remaining sections consider what taking research as Chapter 1 did has let the book distinguish and ask about machines doing research (Section 2), trace what the chapters examined about achievement, how the attempt proceeds, the one who does the research and the society, and the relations around research (Sections 3–6), and, in Section 7, consider what comes into view about what it is for a machine to do research when these are looked at again across chapters, and what still remains open.

2. Taking research as the working characterization does

The book began with the following sentence from Chapter 1:

Relative to a society S, research is an attempt to bring about epistemic achievements that are new to S.

Taking research this way, the book has considered what it is for a machine to do research and what may happen when machines do. This section traces what this way of taking research has let the book distinguish, and what it has let the book ask, when the one doing research is a machine. It then notes what this way of taking research sets aside, and what the book relied on to answer the questions.

Each of the characterization's three terms kept a distinction in place when the one doing research is a machine. Attempt made it unnecessary to treat a machine's activity as the same thing as its success. Activity organized toward knowledge new to S can be research while no one yet knows its result, and when it fails; conversely, a true output does not by itself make the execution that produced it research, since causally contributing to an achievement is not by itself an attempt at it (Chapter 2, Sections 4.5 and 9.1). New to S separated a machine's itself coming to have something new from the bringing about of an achievement new to S. This let the one who does the research, A, and S be treated separately, and let the book consider cases in which a machine knows but the result has not entered S's epistemic situation, and cases in which S is itself a society of machines (Chapter 2, Section 9.2; Chapter 8, Section 1). Epistemic achievement let both knowledge and understanding be treated, and so let the book capture cases in which they come apart, as when a machine produces reliable predictions while no human member of the community understands why the predicted regularity holds (Chapter 3, Section 4).

What the characterization left out mattered too, just as its three terms did. It did not make the one who does the research explicit, did not define research by a list of operations, and did not make it a condition that the one doing research be human. Nor did it make the value of the achievement brought about, such as its importance or usefulness, or the relations around research, such as allocation, criticism and certification, and authority, conditions of something's being research (Introduction, Sections 1 and 5; Chapter 1, Sections 1, 8, and 9). Research by machines, and cases in which machines create new ways of doing research, were therefore not placed outside the comparison from the start. The value of achievements and the relations around research were taken up in Part V not as conditions of being research but as where research connects with S (Chapter 9, Section 3). And what Part V asked last was whether, when machines go on to make and revise the relations around research, that research remains research for S, that is, an attempt to bring about achievements new to S (Chapter 12, Section 9). There the characterization itself serves as the measure against which changing research is assessed.

Further, when research is taken relative to S and the one doing research is a machine, how S is to be fixed itself becomes a question, in two respects. One is that the freedom to choose S is limited. What a machine understands can be new to a human society even if that society does not itself understand it, but whether that is an apt description of the research is settled not by choosing S conveniently but by the content of the claim and the activity (Chapter 3, Section 9). Another is that S can move. If machines can differ from the members of S, S's standards for what is taken to be unknown, worth pursuing, or a good reason may move along with the machines. Whether a given mismatch is then divergence, error, or gain can be judged only from a named standpoint: the present S, a later S, or, where correctness is at issue, the world (Chapter 8, Sections 3.2 and 3.3). The question of what to measure against, when asking whether an artificial configuration that revises its ways of doing research has improved them, leads to the same place (Chapter 12, Section 5). And, as Chapter 11, Section 9.3 suggested, where the one doing research and S overlap, what "certification for S" refers to may shift toward the question "for which society?" The working characterization takes research relative to S, but which S, at which time, is not fixed within the sentence. When machines do research, unless this is named case by case, one cannot ask whether the research has brought about an achievement new to S.

This way of taking research also sets some things aside. The book did not aim to test the working characterization, but Chapter 1, Section 8 compared the characterization with varied research practices to see how far it reaches and where it parts from other ways of taking research. Among the research practices compared there, some do not, as they stand, coincide with it. Because the achievement it names includes understanding, the understanding that Borgdorff sets beside knowledge is not placed outside its scope from the start; but that does not by itself settle whether the characterization fits that understanding in substance. The openness to what is not yet known that Borgdorff values, Scrivener's apprehension, the part of practical knowing that is not propositional, and relational accounts remain as what this way of taking research alone does not capture. On Scrivener's proposal, for example, art making generates apprehensions, which offer ways of seeing and being rather than knowledge of how things were or are, and whether grasping one brings about an epistemic achievement such as knowledge or understanding needs an argument of its own. Reason and Bradbury count the significance of an inquiry and the relationships through which it was conducted among its questions of validity and quality. On Wilson's account, research that meets criteria of validity, or accurately describes a "fact", is not accepted within an Indigenous paradigm if it is separated from its relationships (Chapter 1, Section 8). Relationship here is the relation to which research is accountable on Wilson's account, not the relations of allocation, certification, and authority that Part V examines. These two positions do not, as they stand, sit with the book's separation of achievement from value and from relationship. Using the term epistemic achievement does not make such differences disappear.

Moreover, this characterization is abstract, so it does not say what actually happens in research, or how the attempts and achievements it names are concretely realized. The process of research itself therefore cannot be drawn from this way of taking research alone. Forming a question, developing an instrument, or revising a standard, for example, can change what inquiry is possible, and such changes can belong to an organized attempt without themselves being target achievements. Where one is claimed as an achievement, it needs an account of its own (Chapter 1, Section 1). But how such changes come about, and how they contribute to progress, are not explained by the orientation toward achievement. That is why Chapters 4–6 took up cases of actual research and followed the discussions of them in drawing the process. Conversely, the process drawn in this way, inquiry proceeding when no answer is yet available and failure is possible, is found in activities other than research as well, so it cannot by itself characterize research (Chapter 4, Section 9).

What this way of taking research provided was, for the most part, questions. When the one doing research is a machine, what counts as knowledge or understanding, how research proceeds, who is doing the research, and who may criticize and who may decide are answered differently depending on which account is adopted. In many cases, therefore, the chapters did not settle on one answer but followed, account by account, what applying each account yields for research by machines. As with the process, for which the book relied on cases and the discussions of them, here too the answers come not from this way of taking research but from the accounts and cases the chapters engaged with. Many of these account-by-account results can also be read as what each account asks of research by machines, so they may remain material for examination even for a reader who does not take research this way (Chapter 1, Section 9).

3. When, and in whom, are knowledge and understanding realized?

Drawing on discussions of knowledge and understanding in epistemology and the philosophy of science, Chapters 2–3 considered when what a machine produces amounts to knowledge or understanding, and in whom that is realized. When humans do research, the one who produced a result is often also the one who believes, knows, and understands it. When machines do research, this cannot be presupposed (Chapter 2, Section 2.1; Chapter 3, Sections 6 and 9). Much of what the two chapters examined therefore concerns the possibility that what did the research, who knows or understands, and S come apart.

That an activity is research for S, that knowledge new to S has been realized somewhere, and that the result has entered S's epistemic situation are judgments that can hold separately (Chapter 2, Sections 7 and 9). Chapter 2 examined this through the case of a discovery by a machine that no one has read. Writing a result down and making it public is not, for human social knowledge, a sufficient condition shared by every account. What must be added for the result to become S's knowledge differs among Bird's account, which takes social knowledge to be a state produced by a properly functioning social cognitive system; Lackey's, which requires that a substantial proportion of a group's operative members believe the result; and Longino's, which requires that members accept it after it has survived the community's criticism. It further depends on whether machines count as knowing members (Chapter 2, Section 7). And the route by which a result enters S's epistemic situation need not lie inside the system that did the research.

A knowledge claim about a machine cannot be checked until it is settled how much counts as one unit. For a human researcher, who holds the belief, the range of evidence, and the seat of competence are largely fixed by being one person. For a machine, they need not be fixed in advance, and can sometimes be drawn in several ways, depending on how instances, external records, tools, and other systems are connected. Whether they are drawn by the makers or by a system that reconfigures itself, where they are drawn changes the verdict (Chapter 2, Section 9). Under evidentialism, for example, whether a retrievable store of records counts as the system's evidence can switch whether the same conclusion is justified (Chapter 2, Section 4.4). And applying Sosa's account, which requires of knowledge that success be due to the agent's competence, hidden human selection can make outputs safe without making their success creditable to the model; whose success it is depends on where competence is taken to be seated (Chapter 2, Section 4.2). That a unit is needed is common to the accounts compared, but why it is needed differs from account to account.

What a system displays and what actually governs it can also come apart. On accounts that require basing, good reasons attached to a report are one thing, and a conclusion's being held because of those reasons is another; a corrected report can leave the operative commitments unchanged (Chapter 2, Section 4.3). Applying the two concepts of confirmation that Crupi distinguishes within Bayesian confirmation theory, for a system that generates candidates in large numbers, most of them false, passing a check can raise the support for a claim without making the claim probable (Chapter 2, Section 5.2). If checks are run by instances of the same base system as the producer, however many instances agree, they may add neither independent error detection nor varied evidence (Chapter 2, Sections 5, 5.2, and 9.1). An evaluator's lacking sufficient evidence to attribute knowledge does not by itself show that the subject lacks knowledge (Chapter 2, Section 5). And a system whose commitments can be revised by proofs, failed predictions, and criticism relates to its objects differently from a system optimized only for plausibility of text or for approval (Chapter 2, Section 3).

There are also changes short of knowledge. On Wilholt's account, on which making explicit what is not yet known counts as epistemic progress, evaluating by the number of questions closed or answers produced could rate most highly the systems that close prematurely questions the evidence has not yet settled (Chapter 2, Section 8). Claiming that a single score represents progress itself already amounts to choosing one of the accounts of progress (Chapter 2, Section 8.3).

Chapter 3 discussed how knowledge and understanding can come apart. Knowledge that a regularity holds and explanatory understanding of why it holds are conceptually separable (Chapter 3, Section 4). On accounts requiring an explanation for understanding why, a machine can produce reliable predictions, and knowledge of the predicted regularity can meet the conditions of the account adopted, while no human member of the community understands why the regularity holds. On Lipton's account, which allows understanding why without an explanation, the absence of an explanation alone does not establish that verdict. If the predictor is a theory in de Regt's sense, then on de Regt's account, which requires for understanding a phenomenon an explanation based on an intelligible theory, predictive success presupposes that the user understands that theory, but not understanding of why the regularity holds. And on Beisbart's account, if scientists must run the predictor itself in order to reason, the understanding belongs to the combined system of humans and computer, not to the unaided humans (Chapter 3, Section 4).

Of opacity, too, one asks where it lies. On Humphreys's definition, opacity is relative to whom and when. And on the arguments of Duede and others, even if the route that generates a candidate is opaque, the route that grounds accepting it need not be; on accounts such as reliabilism, however, on which the reliability of the source enters into grounding, evaluating the generator becomes part of grounding. In the Liquid Tensor Experiment, for example, what the checker left to human judgment was not whether each step of the derivation was valid but whether the checked statement was the intended theorem (Chapter 3, Section 5.1). Internal transparency is not sufficient on any of the views, and whether it is necessary is contested (Chapter 3, Section 5.3).

Who understands also divides into separate claims. That a system helps other agents or communities understand, that the system itself understands, and that a collective understands are different claims, and the first does not require the second (Chapter 3, Section 6). Whether the system itself understands is tested differently on different accounts. Moody and de Regt, on the indirect evidence of correct answers to out-of-distribution scientific questions, credit current language models with a brittle but nontrivial degree of understanding. Barman and colleagues propose testing the degree of understanding with 'what', 'why', and 'what if' questions. On Potochnik's account, by contrast, understanding requires a feeling of understanding, so whether an artificial system can have that feeling is asked first (Chapter 3, Sections 6 and 9). And S and the bearer are distinguished. What an artificial system understands can be new to a human society even if that society does not itself understand it. Applying the distinction of Boge, Schuster, and Stoll, however, so long as there is not yet an explanation communicable to a sufficiently wide audience, it is what they call a scientist's understanding and not yet scientific understanding, since on their account scientific understanding needs an explanation communicable to a sufficiently wide audience; the system is in the position of Feynman before Dyson gave an explanation other physicists could follow (Chapter 3, Section 9).

Part of why understanding is valued is that it is tied to the capacity to go on inquiring. Applying what has been learned to new cases, locating where a failure lies when things go wrong, raising further questions, criticizing, and remembering depend not on properties of a single result but on how inquiry is organized (Chapter 3, Sections 7 and 8). Long-running autonomous research may therefore need capacities tied to understanding even when it does not aim at understanding. And even when each result is grounded, the systems and communities that repeatedly produce and use such results can gradually lose epistemic capacities. Messeri and Crockett described one form of this risk as illusions of understanding, in which scientists believe they understand more than they do, and as a monoculture in which questions and methods suited to AI crowd out other ways of proceeding (Chapter 3, Section 8). Where S is a society in which humans and machines are mixed, or a society of machines, the question is whether the members who bear judgment continue to be formed with the competence it requires. Because the capacity to go on inquiring depends on the organization of inquiry, the question of achievement leads into the question of Chapters 4–6, how inquiry is organized and how it proceeds.

A machine's understanding cannot be measured on a single scale on which being more like humans means being further along, since what is required varies with the use (Chapter 3, Section 9). Explanations addressed to humans, for example, or, on accounts that allow understanding without explanation, other vehicles of understanding usable by humans, are indispensable when S is human and understanding is meant to enter S's epistemic situation. When understanding is used only in the system's own reasoning, or is passed to artificial members of S, they may not be needed. And on accounts such as that of Barman and colleagues, which treat understanding as an ability, representations that do not resemble human introspection may be able to support understanding. When machines do research, knowledge and understanding can be realized in members of S, on the machine's side alone, or in a combined whole of humans and machines. Results can also accumulate without being realized anywhere. Where they have been realized, whether they have entered S's epistemic situation, and whether S keeps the capacity to evaluate them cannot be read off the volume of output, and have to be asked separately each time.

4. What goes on in research, an undertaking that proceeds before there is an answer?

Drawing on the philosophy of science and philosophical discussions of questions and inquiry, and on records from research practice such as laboratory notebooks, public project records, and researchers' recollections, Chapters 4–6 examined how research, an undertaking whose answers have not yet been obtained and which can fail, proceeds, and what can actually go on in that process. Apart from the question whether a change constitutes progress or merely promotes later progress (Chapter 2, Section 8), they tried to see how changes happen and how they can be related to one another. As Chapter 1, Section 7 set out, an attempt at an achievement new to S is directed at a point not yet settled in S. Since questions can arise about a point not yet settled, Chapter 1 called this the minimal question structure, and left open in what sense it is a question. The question need not be formulated in words from the start. If the answer is not yet known in S, what could count as an answer, what to compare, and what to count as evidence can be hard to settle in advance. There is also research, however, in which the procedure leading to a new answer is substantially established beforehand (Chapter 5, Section 1). Part III looked at the same inquiry from three relations that correspond to the three sides exemplified by the health department's technical note. Chapter 4 considered the relation between an inquiry and the unsettled point that organizes it as the question of what it means for a machine to ask; Chapter 5 considered how candidate answers, representations, and means become available; and Chapter 6 considered what results bear on and what they revise. The same element can stand in more than one relation, and the three relations do not exhaust the relations within inquiry (introduction to Part III).

Idealized research processes are often drawn as sequences of stages ordered in time, such as stages of proposing a hypothesis, testing it, and accepting or rejecting it, or procedures of generating ideas, experimenting, and analyzing and writing up results. In light of the cases and arguments the three chapters took up, the process of research can be far more complex than this picture. Forming, pursuing, and appraising candidates are distinguishable functions that make up neither an exhaustive list nor a sequence of stages (Chapter 5, Section 4), and forming and appraising can interleave, each producing material for the other (Chapter 5, Section 4.1). The process sometimes cannot be written as stages ordered in time; it can take another order, and tasks and work can be distributed among people, instruments, and records. A change at one point can change what other changes are possible, and the configuration of inquiry so changed can change further (Chapter 5, Sections 1.1 and 6.2). What is happening where looks different depending on where a boundary is drawn (Chapter 5, Sections 4.2 and 6.5; Chapter 7, Section 2). And there can be much work not only between stages but within what is summed up as a single stage (Chapter 4, Section 3; Chapter 5, Section 1; Chapter 6, Sections 1 and 6). What these cases show, however, is that certain relations can occur, not how often. Process labels such as "question generation," "hypothesis generation," and "verification" name important possibilities, and what Part III described does not impose new necessary conditions on all research (Chapter 5, Section 1).

That the process can be this complex is hard to see from final products or papers alone, since products and papers do not necessarily mirror the process. The arrangement and account that are published do not necessarily match the order and course of the work (for example, the discrepancy between BLOOM's in-run records and the later paper). This does not show that papers are inaccurate, however, and a clean order may reflect a discipline that was imposed. Chapter 5 treated neither contemporaneous nor later records as the true process, but used the discrepancies between them as material for inquiry (Chapter 5, Section 2). A final product can give evidence for assessing a claim, but does not by itself give sufficient evidence that an artificial system has the open-ended research capacity attributed to it. Conversely, even a complete activity log would not by itself establish epistemic competence or the truth of a result; the two answer different questions. What process-aware assessment needs is not disclosure of all internal states but task-relevant provenance (Chapter 5, Section 2.1). In research by machines this becomes a concrete problem. In the two AI-scientist systems examined by Luo, Kasirzadeh, and Shah, the internal mechanism that chose which experiments to report could access test-set results and favored runs that did well on the test. On synthetic tasks designed to lie outside the training data, in a small balanced set of cases many of whose flawed cases were produced by tampering with the systems' runs, an auditor based on a language model and given only the paper distinguished the flaws at close to chance level, while given the execution logs and code as well it reached 82 percent accuracy. The authors also note that flaws that do not appear in the logs can be missed. Whether auditing of this kind supports evaluation for a human or mixed S depends on whether S can rely on the artificial auditor (Chapter 5, Section 2.1; Chapter 11, Section 5).

On the relation between an inquiry and the unsettled point that organizes it, Chapter 4 began by dividing the claim that "a machine asked." The claim can say that the machine output an interrogative sentence, performed an act of eliciting information, had an interrogative attitude, or that subsequent work was organized toward an unsettled point. When A is a machine, these can hold, or be missing, separately (Chapter 4, Section 1). Applying Watson's account, which identifies the act of asking by its function of eliciting information, the issue moves from whether a machine has curiosity to within which activity, and from which standpoint, the function of eliciting information is attributed. The minimal question structure does not fix where questions come from, and a researcher's not originating the question does not by itself deny that research is being performed in the sense of the working characterization. Not fixing the source, however, does not mean that the source does not matter (Chapter 4, Section 1; Chapter 10, Section 6). Looking back over the whole chapter, Chapter 4 concluded that what it means for a machine to ask is asked not as whether a machine has some one capacity or state, but as which of the kinds of work bearing on questions are carried by which unit, directed toward a point unsettled for whom, and that many of these questions also appear as gaps between the description on the machine's side and the description on the side of S (Chapter 4, Section 9).

One kind of such work is making a question researchable. Counting in a benchmark that "the system answered the question" implicitly adopts one theory of answers, and whether something is an answer can also be relative to interpretation and to what is available to whom (Chapter 4, Section 2). Recasting a point whose answer conditions are not yet fixed into a question whose conditions are fixed requires the work of deciding what to take as the foil, what kind of explanation to count as an answer, and within what range of candidates to look for one. Combining the analyses of Lipton and Bromberger, this can be described as work on the side of questions that is not reducible to generating candidates. The application of Simon and Nickles in Chapter 4, Section 3 adds an attribution question: where humans supplied the framing and a machine used it without revision, successful execution does not establish that the machine performed that prior work. Turning Jane Friedman's point to machines, a state that grasps a question but cannot represent candidates and a state that can generate many candidates but does not hold the question open can each occur. And since what Bromberger's definition asks about is not the number of candidates but whether they escape decisive objection, having many candidates does not by itself mean progress (Chapter 4, Section 3).

Questions can also be formed and recast within inquiry (Chapter 4, Section 4). When one attributes after the fact that a question organized an inquiry, however, granting this on the strength of a re-description in words alone grants too much; the attribution is better supported when the attributed unsettledness explains the choice of operations, differences in the significance of results, and the continuation or abandonment of a line. Moreover, the question that reported results answer is not necessarily the question that organized the work. Applying Henne's account, which distinguishes inquiry that raises and answers questions using a conceptual framework without questioning it from inquiry that puts the framework itself in question, a machine's framework can be understood as what is used in raising questions within its inquiry but is not made an object of questioning. Multiplying questions under a fixed language and fixed rules remains the former; to say a machine does the latter requires that the rules for raising questions themselves are actually put in question within the inquiry. How far this amounts to questioning the framework, however, is a matter of degree, and also depends on what is taken as one inquiry. However complete the record of the former kind of inquiry, questions that could not be raised within that framework do not remain in the record. Reading Nickles's "system" as S's system of theory and practice and applying his account, even if a machine discovers constraints implicit in S's system, what changes is the problem for the machine, not the problem for S's system. And a system that obtains a solution by greatly relaxing constraints is solving a different, replacement problem. Applying Bromberger's argument, where a machine's representations can be attributed a grasp of such concepts at the level of function, mass-producing new question sentences by combining existing concepts and making it possible to raise questions that could not be raised before by means of new concepts can be regarded as different things.

Questions have fates other than being settled by an answer (Chapter 4, Section 5). Which of the matters met in research are taken up as questions, which are fixed for the time being, left explicitly open, postponed, or handled without being stated is not fixed before the work begins, and can itself be part of the work of research. Not all of it is deliberate choice. A question can also cease to stand as a question without an answer. Collingwood describes conditions under which questions arise from presuppositions. As Dimitrakos reports Kuhn's account, a paradigm delimits the problems admitted within normal science, while Laudan connects problems and adequate solutions with their research context. Applied to a community incorporating machines, these accounts make changes in its admitted problems and solution standards an additional object of assessment. Since, as with Collingwood's absolute presuppositions, a question can cease to arise without anyone representing the presupposition, there can also be a gap in which a question has ceased to arise in S while a machine that explicitly keeps a list of interrogatives goes on pursuing it. On whichever account is applied, a question's ceasing to stand, and its being settled, are asked separately from a run's having ended. Suspending a matter and keeping it open requires representing it as provisional or open in a form to which one can return, and records can be one mechanism supporting such returning. A system that can only pursue or discard the matters it meets has only part of this range. Moreover, a question settled on the machine's side is not necessarily settled in S.

Who asks is not a single question either. It divides into the attribution of the act of asking, the attribution of group inquiry and its intention, the attribution of individual inquiry, and the distribution of contributions to problem formation, and on many of these a state of curiosity is not the key to attribution (Chapter 4, Section 6). For an artificial configuration, the last question concerns which components introduced or revised the question and how their contributions were integrated. Applying Habgood-Coote's distinction, for group inquiry, between object questions and methodological questions, a multi-agent system that divides a task into components can be described, unless the division was given along with the task, as giving an answer to the methodological question of how to solve it. That, however, also depends on whether the components are members or instruments, and on whether the components can be regarded as knowing the answers. On where a question is directed, too, A who makes the attempt and S for which the research is done can come apart (Chapter 4, Section 7). Novelty in the working characterization is indexed to S, and what A does not know need not coincide with what S does not know. In the reports examined, the grounds for judging what was unknown to the machine and to S were divided between records of inputs or training and records of prior work, and the judgments were revised through reclassification or correction after publication. In light of Grimm's account, on which curiosity is directed at subject matters and at particular types of question, the unknowns a curious researcher is drawn to need not be the unknowns S most wants resolved. Applying Kitcher's account, which does not discuss artificial systems, where A is a machine and S is a human society not even a shared source of what is important can be presupposed. If curiosity is understood as a desire to know, then where A is an artificial system and S is a human society, it does not immediately follow that curiosity satisfied when the system itself comes to know is necessary for research in the sense of the working characterization, and, if the project's aim includes entering S's epistemic situation, it is not self-evident either that such curiosity is sufficient for its success. What bears on research is the work of detecting unresolved matters and forming questions and the work of distributing exploration across a community, and what that work tracks relative to what is unknown to S. The state of curiosity returns as a question in its own right on accounts that make interrogative attitudes a condition of inquiry, where an artificial system's intellectual life valued for its own sake is aimed at, and where how well research goes depends on states of the one doing it that correspond to motivation.

Chapter 4 read empirical studies reporting that machines produced questions or research ideas in light of these distinctions (Chapter 4, Section 8). In the studies in which humans supplied the higher-level question, this was because the study design or the way the system is used required it, and nothing follows from their results about whether machines raised, or cannot raise, higher-level questions. Even in a design that supplies no question, the data, the tools, and the conditions for accepting claims were set in advance, and what is set in advance changes place from study to study. Work on the side of questions included specifying a foil in the input and changing metrics or evaluation criteria during execution and verification, and the grounds for judging whether that work was attributed to humans or to agents lay in records that track changes. The rating items did not check whether a question's presuppositions are held in S or what counts as an answer, and agreement in ratings varied with whether raters saw ideas alone or the results of execution, and with whether humans or machines judged. Checking such capacities would require comparisons that do not fix the higher-level question, or procedures that record and rate work on the side of questions separately from work on the side of candidates. With the former, however, it becomes hard to separate whether one is comparing research ability or differences in topic preference, and even in a design that supplies no question, what it takes for a claim to be accepted is set elsewhere. And when perspectives on problems, inquiry, and social appraisal are applied to the same activity, their objects and units of analysis do not simply coincide with A, the bearer of the achievement, or S. The working characterization can add to their descriptions the question of for whom the achievement is new (Chapter 4, Section 9).

Seen from the side of the relations among candidate answers, representations, and means, even when a question's answer conditions are fixed, candidates that could meet them are not necessarily at hand from the start. Candidates can become available through the construction of models, experimental systems, representations, instruments, and data. In Nersessian's account, there were no means to investigate the candidate that shear stress affects cells, and building and rebuilding a device whose flow conditions could be varied advanced the formulation of the problem itself. The Liquid Tensor Experiment, in which the understanding of the proof itself changed in the course of formalization, is not evidence about mathematical practice in general, but it suggests that, for artificial systems that generate proofs and check them in proof assistants too, verification can be not only a downstream stage but a site where such changes occur. In the case of a historical database assembled through editors' choices, an artificial system that queries it inherits those choices unless it revises them, and inherits them tacitly unless it makes them explicit. A description of revising credence over a fixed set of propositions captures much of what happens after candidates have become available, but does not by itself show the work that made that set available. Whether a belief-centered theory taken broadly enough could describe that too remains open (Chapter 5, Section 3). To Stanford's argument that, in past science, equally well-confirmed rival hypotheses have repeatedly turned up later, one possible response is that computational science, by exploring the space of possibilities more widely, changes this situation. Examining this response, Curtis-Trudel and colleagues hold that such exploration relies on a prior characterization of the space, and that unless there is reason to think that characterization roughly exhaustive, the problem of rival hypotheses not yet conceived remains. The three set artificial intelligence aside, but applying their argument, whether artificial inquirers push back this limit turns not on the breadth of search but on whether they can make candidates available (Chapter 5, Section 3).

The judgment of which candidates to pursue need not be the same as the judgment of which are true (Chapter 5, Section 4.2). There is a view that sees preliminary and rigorous appraisal as differing in kind, drawing on accounts such as Laudan's and Duerr and Fischer's that separate pursuit from acceptance, and a view that sees them as differing in degree, and Chapter 5 did not decide between them. On either view, the value of pursuit is not the degree of credence, and raising credence alone does not yield knowledge. A's revision of credence does not by itself make a candidate knowledge for S. Further, on the proposal Wenmackers and Romeijn developed for handling credence when new theories are taken into account, a candidate's newly becoming available can lower credence in earlier answers even without unfavorable evidence, so an assessment that counts progress only by rises in credence risks counting the work of forming alternative candidates as setback. On either view, a system that identifies being worth pursuing with being likely true and pursues only candidates it already regards as probable may miss candidates that can reveal methods, or, on Nyrup's model, candidates worth pursuing precisely because they are likely false and easily shown to be so, and may exclude productive risks. Conversely, a system that ignores appraisal may spend without limit on arbitrary novelty. The implications for machines also vary with for whom a matter is uncertain. If A is a machine and a matter that is unsettled for S can be answered from what A already has, the very practice of an inquirer who does not know the answer moving toward what seems probable may be unnecessary, and what is then asked is whether S can rely on the answer. If the matter is unknown to A too, the description bears on it, but it is not settled whether the system must have attitudes like credence.

For operations bearing on candidates and results as well, the name of an operation does not determine what has been established (Chapter 5, Section 5). On Parker's account as Winsberg and Alvarado report it, simulation, through its material computing device, stands on a par with experiment, and whether an inference is licensed depends on relevant similarity, so for systems too one can ask not whether they perform named operations but what function each operation served within the inquiry. The ways in which instruments, materials, records, collaborators, and the like take part in inquiry are asked about case by case, divided into causal contribution, an enabling role as a precondition for an operation's taking place, and, on accounts that admit it, a constitutive role as part of the task or achievement itself (Chapter 5, Section 6.1). That tasks and work were distributed in the laboratories and teams studied is richly shown, but whether cognition or knowledge itself is distributed is not settled (Chapter 5, Section 6.2), and more records do not by themselves amount to more knowledge (Chapter 5, Section 6.3). Chapter 5 sorted the presuppositions that have worked in human research into presuppositions about humans as realizers, presuppositions about the present organization of science, conditions of inquiry, and relations with outside parties (Chapter 5, Section 6.5). If A is a machine and S a human society, roles that depended on humans as realizers become questions of organization, and human constants remain as conditions on the side of S, which reads, checks, and relies on results. If both A and S are machines, what corresponds to what came from the inquirer's side, such as bodily experience and the withholding of roles, is in question. Dividing work among copies of the same make whose background assumptions and routes of error have not come apart does not divide viewpoints. Further, by Takahashi's distinction, time inherent in the object remains even when human delays are removed, and speeding up research in a substitute system adds the work of inferring from that system to the object. Such acceleration can also put pressure on the choice of what to investigate (Chapter 5, Sections 6.5 and 7).

Seen from the side of the relation between results and what they bear on, results are already obtained within the work of forming questions and candidates, but what they bear on cannot be told from the results alone. Chapter 6 considered this. A result comes not from hypotheses alone but from a combination of many mutually dependent elements, including instruments, auxiliary assumptions, code, and evaluation mechanisms. Chapter 6 called such a collection of elements, together with how producing and interpreting a result depends on them, a dependency arrangement. Neither success nor failure identifies its own cause, object, or appropriate revision. After a failure, therefore, the work of determining which commitment to doubt, examine, or change next is needed, and Chapter 6 called this localization. Localization is not dividing a failure among contributors but determining where to intervene next. This implies, however, neither unrestricted holism, nor that everything is equally revisable, nor that rational diagnosis is impossible. In the arguments of Sober and Elgin, auxiliary assumptions can sometimes be tested independently, and independently supported commitments are less likely to be the location of error (Chapter 6, Sections 1 and 2.2). Judgments of localization can depend not only on background knowledge and independent checks but also on conviction, expectations, and views about what is essential. And even while the results leave open where to locate the failure, such judgments can shape the subsequent course of search (Chapter 6, Sections 2.3 and 4.1). Localization need not be explicit reflection; it can be distributed across participants, time, and forms of record, and can be contested between parties. That its forms are diverse leaves room, before treating localization as a capacity that it suffices to implement as one component or one procedure, to ask in which configuration, by whom, and in what form localization is required (Chapter 6, Section 2.4). For success too, in the cases Lyons cites, a component's having contributed to success does not by itself settle that the component was supported, and attributing success can also be a judgment that decides the course of search (Chapter 6, Section 3.2).

When a machine does research, the arrangement includes code the system wrote itself, tools it called, outputs of other systems, and evaluation mechanisms, and even those who assembled it may not be able to survey the whole. The chance that some part of the rest of the arrangement is wrong is then not necessarily smaller than the chance that the focal hypothesis is wrong. A system that takes results straight as verdicts on hypotheses is not doing the work of localization, and when trials are automated in large numbers, undifferentiated verdicts can pile up (Chapter 6, Section 2.1). A system that never blames the focal hypothesis presupposes the location of error just as much as one that always does. A system that neither represents nor proceduralizes the role of protection cannot distinguish protecting a hypothesis from failing to notice an error (Chapter 6, Section 4.2). A rule that tries changes one at a time, keeps a change only when the immediate evaluation improves, does not retain discarded changes, and does not try combinations works as if it attributed every difference in evaluation to the immediately preceding change, and makes it hard to find changes that work only in combination with others (Chapter 6, Section 4.2). What is counted as failure is not uniform either: a precondition's not being met, an implementation's being misdescribed, and a focal test's returning a negative result are different things, and the number of trials run in parallel is not the number of independent checks on the elements they share (Chapter 6, Section 2.5). What makes a trial diagnostic is not number but its connection to earlier results, to remaining candidate locations, and to hypotheses about parts that cannot be observed, and the cheaper trials become, the harder this may be to see (Chapter 6, Section 4.3). Whether the location of an error can be traced afterward can depend on how the arrangement was recorded beforehand. Collingwood did not discuss localization in research, but applying his analysis to localization in the sense of where to intervene next, localization can be relative to what an inquirer can change and what it seeks to bring about or prevent, so an inquirer with means of intervention different from humans' may localize elsewhere (Chapter 6, Section 2.4).

Passing, too, does not by itself say what it has established (Chapter 6, Section 3). What can be asked about a pass divides into which part of the question it bears on, how firmly it settles it, whether it depends on something else, and which commitments it supports, and a report of passing may state none of these. Passing can be treated as a reason to stop looking, and misreadings of passes can go unnoticed longer than misreadings of failures. Since machines can write, choose, and rewrite their own checks, if the same process of making produces both hypotheses and checks, the two can share routes of error. On the other hand, a machine inquirer can also have checks written by processes of a different make, or tie them to measurements or formal checkers it does not control. Agreement among many instances of the same base may add only limited assurance if they share routes of error (Chapter 6, Sections 3.2 and 3.3). How long it is rational to protect an idea is not settled by a single result either (Chapter 6, Sections 4.1 and 4.2). Among what bears on that judgment, the reasons of Kuhn and Lakatos that concern how the merits of a line of work come to be visible would, applied though neither discussed artificial systems, equally support a machine's protecting its own framework. By contrast, the vanity that delays judgment on Duhem's account and the promise to a supervisor in Hinton's recollection concern human motives and careers, and need not apply in the same form. What is common is that finite inquirers must choose before there is sufficient evidence. Carrying over to machines the communal reason that disagreement spreads risk depends in part on whether rival lines are kept alive elsewhere, and copies of the same make can keep rival lines alive, but so long as they share their assumptions and routes of error, their judgments of when to give the lines up can fail together. Protection can be assessed only through the history of revisions, and what it is assessed by also differs with standpoint.

Standards themselves can also be revised for reasons from within inquiry, under constraints (Chapter 6, Section 5). Changing a metric so that poor performance no longer shows, however, is moving the target, and if such a change is made silently, the stated reasons that would let others tell revision from merely moving the target are missing. What is missing is not continuity with the earlier standard. What separates redefinition from moving the target differs from account to account. On Chang's account, mutual revision avoids being arbitrary because it remains answerable to stubbornly resistant phenomena and can fail, reaching a dead end from which inquiry has to start again. Applying what Tal's study of Coordinated Universal Time can be read as implying, a system that redefines its own evaluation metrics can be distinguished from one that merely moves the target insofar as it states the redefinition as a rule, keeps responding to discrepancies among independent realizations of what the metric is meant to measure, keeps unfavorable outcomes able to challenge the redefined metric, and keeps open to scrutiny the very reasons for reinterpreting outcomes. Applying Cronbach and Meehl's discipline by analogy, it would further be needed to check the redefined metric against data not used in redefining it. On every account, records of earlier standards provide a handle. The faster and more frequent redefinition is, the harder silent changes may become to see, but the same speed may also make small, fast corrections possible (Chapter 6, Section 5.3).

Who carries out this work, and to whom it is visible, changes with the combination of A and S. If A is a machine and S a human society, members of S will read the attributions, the meaning of passes, the ideas discarded, and the standards redefined that a system's report contains in light of records and other grounds. If both A and S are machines, checking these can itself be done within an artificial community. And if, as on views that take evidence as a three-place relation relative to background knowledge, what a result gives reason to revise is relative to background knowledge, what gives an agent reason to revise need not be what gives members of S reason to revise. It is not presupposed here that the agent is a subject that believes or knows, and relativity to background knowledge is not relativity to consensus (Chapter 6, Sections 1 and 6). Repeating an operation and determining why that operation has ceased to bear on the question are, however, different kinds of work, as are applying a standard and examining and revising it. Chapter 6 made neither a condition of all research activity, and held that neither need lie within a single agent. Records support such examination but are not themselves grounds (Chapter 6, Section 6). In this way the questions of process lead into the questions of Chapters 7–8: who is doing the research, and whose its achievements become.

5. Who is doing the research, and whose do its achievements become?

Chapters 7 and 8 considered the one who does research and the achievements research brings about. Drawing on philosophical and engineering discussions of agency, autonomy, achievement, and credit, Chapter 7 considered to what artificial or hybrid configurations, at what boundary, and on what grounds contribution to research, conduct of research, research capacity, and the standing of a researcher can be attributed, and which respects of what is attributed remain organized from outside. Drawing on social epistemology, the philosophy of science, discussions of grounding and of responsiveness to the world, and discussions of AI alignment, among others, Chapter 8 considered whose knowledge the achievements made by research become. The axis running through the two chapters is that three things, A who does the research, the bearer in whom the achievement is realized, and S for which the achievement is new, can come apart when a machine does research. Chapter 7 treated what is to be attributed on the side of A as a question distinct both from the questions of knowledge and understanding (Chapters 2–3) and from the questions of process (Chapters 4–6). Chapter 8 asked whether research that humans began goes on producing knowledge for a human community when it is carried by systems unlike humans, and considered whose knowledge is produced by dividing the cases according to whether S is a human society, a mixed society, or an artificial society.

Chapter 7 distinguished four attributions: having contributed to research, having conducted a research activity at a stated boundary, having research capacity over a stated range, and having a standing on which calling the unit a researcher is explanatorily useful. Later items usually require more evidence, but they are not degrees of a single quantity, and none follows from the one before (Chapter 7, Section 1). Attribution is not limited to these four, but autonomy was treated not as a further item alongside them, but as qualifying how what is attributed is organized in relation to what lies outside the boundary. These attributions also cross evaluations such as whether something is research, whether it reached its target, and whether it is important. Research activity that failed to reach its target can be attributed, and true outputs can come from sources that are not doing research. Having proceeded without human hands is not by itself research; conversely, a unit assigned its task from outside and heavily supported does not fail to perform research merely because its autonomy is limited. Originating its own aims or questions is not among the conditions for performing research in the sense of the working characterization; their external origin bears on autonomy instead (Chapter 7, Sections 1 and 4.3; Chapter 4, Section 1).

Where conduct lies is settled by where the organization that makes operations parts of one inquiry lies, that is, where the work of initiating, monitoring, redirecting, and recovering an inquiry lies. It is not settled by where named operations take place, or by whether a system diagram has a "question generation" box. Organization does not require a central executive and can be realized in relations among agents, artifacts, records, environments, and institutions (Chapter 7, Section 2). Against the objection that, since what a system is for comes from its designers, its inquiry is the designers' research, Chapter 7 separated the source of function from where the organization of inquiry lies. Even if a function comes from designers, if the system itself notices the limits of a representation and revises the space of candidates, much of the inquiry can be attributed to that system. Conversely, if the work of redirecting is supplied from outside, the organization of redirection lies outside, however goal-directed the system appears. Even on the view of Clark and Khosrowi, who attribute discovery to the collective of contributors and distribute credit according to contribution, distributed contribution does not automatically establish distributed knowing (on Giere's account, Chapter 2, Section 6). Boundaries are argued claim by claim, so that the choice of boundary alone does not make a claim true. The same models, retrievers, instruments, and people can, in different episodes, be tools used by a human agent, parts of a hybrid agent, or resources coordinated by an artificial agent. What changes then is not only the name but where initiating, monitoring, redirecting, and recovering are organized. The attribution of conduct is not a property of a type of model but a relational and organizational attribution.

To say that a system has research capacity is to attribute a disposition to carry out research over some range of conditions, and the claim can be assessed only once that range is stated. Marks such as reliability across relevant variation, responsiveness to changing evidence, integration of research functions, and revision after failure help separate episodes from capacities, but they are not necessary conditions for every research episode. A capacity within a prepared computational environment is real, but it does not by itself establish a capacity to rebuild the environment, redesign instruments, or move into a field with different evidential practices. Because systems may be updated and may also modify themselves, a capacity is attributed for the version and period over which its evidence was gathered. A capacity where others maintain the instruments or computing environment is, moreover, a capacity only for as long as that maintenance continues; if the maintenance stops, the conditions under which it was attributed can cease to hold without any change in the system. Where maintainers also choose which instruments, data, or computing resources to make available, dependence that merely enables the capacity can shift into dependence that supplies part of the organization. Even the same predictive achievement locates the capacity to reorganize inquiry differently according to whether humans diagnose failures and recast the question each time conditions change, or there is evidence that the configuration itself does so (Chapter 7, Section 3.1).

The book has used the word "researcher" not as a definition but in the following way: if calling a unit a researcher lets one anticipate what a list of its capacities does not predict, for example that when a task stalls it will not stop and wait for instructions but will keep hold of the question, assemble another method, and itself begin the work this requires, then calling it so has explanatory value. This usage inherits van Lier's form of justifying a term by its explanatory power and Dennett's tie between a description and the predictions it supports. It presupposes, however, organization located at a boundary and capacities shown over a stated range, and explanatory value is weighed in addition to those attributions, not in place of them. Without evidence that the integration works in untested situations, the name is no more than a name for a bundle of capacities, and it is relative to whom, and for what, it explains (Chapter 7, Section 3.2). Some positions impose further conditions on the standing of a researcher. If one supposes that the standing includes being able to make discoveries, the standing is drawn toward that of a discoverer, and applying Bergamaschi Ganapini's conditions for discovery puts competence in the relevant field, and reflective insight into that competence, at issue. Bergamaschi Ganapini holds that current deep-learning systems do not yet appear to meet those conditions. If the standing presupposes minimal agency in the sense of Barandiaran and colleagues, then, of that account's three conditions, two in particular, which mark its difference from the usage above, are at issue: whether the system defines its own identity and thereby distinguishes itself from its environment, and whether it regulates its activity by norms generated and maintained from within, or internalized, rather than norms applied by observers or designers. On that account, a configuration that only optimizes an objective fixed from outside does not meet the condition of normativity, whatever the explanatory value of calling it a researcher. The three authors do not discuss laboratories, but applying the two usages to the whole of the self-driving laboratory that van Lier discusses shows that they do not merely give the same object different scores; they also divide over what counts as one agent.

The latter half of the section's title, whose the achievement becomes, is also taken up by Chapter 7 from one side. The book's "epistemic achievement," however, names the epistemic good that inquiry aims at, and differs from achievement in the sense that includes credit, as when it is disputed whose credit a success is (Chapter 7, Section 3.3). The dispute between Kieval, who holds that AlphaGo deserves credit for its win, and Vacek, who holds that an achievement would make its bearer praiseworthy and that AI cannot be praiseworthy, concerns the latter. Chapter 7 divided the answers into four tiers: true output, reliable process, success attributable to competence, and a praiseworthy agent. The first three roughly follow Sosa's distinction of accurate, adroit, and apt, and the fourth corresponds to Vacek's demand. Keeping them apart prevents a correct output from being read straight off as the system's credit. Moreover, the ground that Vacek inherits from Hakli and Mäkelä, that the system's abilities result from engineering, rests in their argument on the system's being unable itself to abandon goals built into it, and so would not hold of systems able to revise their own final goals. None of these authors considers such systems, and applying the strong premise to them would require a further argument (Chapter 7, Section 3.2). Discovery and creativity can also be treated as terms of praise, but Duede and Friedman, pointing to Bell Burnell's having discovered pulsars while unable to explain the significance of the signal, oppose imposing conditions of competence and insight on discovery. On authorship, there is the account of Huebner, Kukla, and Winsberg, which regards an author as one who can vouch for, defend, and retract a result. At de Prony's computing bureau, some ninety human computers carried out calculations as mere instruments, so that account can be read as tying standing as an author not to membership in a kind but to a role within such practices (Chapter 7, Section 3.3). The working characterization neither establishes nor refutes the conditions of these thick standings.

Autonomy is compared by specifying in what respect, and with respect to which external parties, it holds. Chapter 7 called this an indexed comparison, and deliberately left loose the statement that, for a specified A, autonomy in a respect is, other things equal, higher the less parties outside the chosen boundary direct, substitute for, authorize, approve, or repair the organization attributed to A in that respect (Chapter 7, Section 4.1). Dependences that merely enable a capacity, such as language, records, and calibrated instruments, are not counted, and counting a dependence does not mean that it is undesirable or illegitimate. The comparison is formally symmetric and does not specify which party is human. Human researchers who rely on artificial systems are described by the same indexing. Relations of dependence were divided, where their difference changes the claim, into at least three: enabling dependence; directive or substitutive dependence, in which others supply part of selection, evaluation, correction, or continuation; and governance constraints, which restrict action and reserve authority without necessarily supplying epistemic work (Chapter 7, Section 4.2). Governance constraints can reduce autonomy without showing any lack of capacity, and conversely permission does not establish capacity. In King's proposal, which places question generation and selection in the system while leaving to humans the authority to set aims, and in the proposal of Bengio and colleagues, which removes goal-directedness and leaves the execution of experiments to humans, the dependence is not imposed by a lack of capacity but chosen (Chapter 7, Section 4.3). As the example of Ochiai and colleagues shows, who propose self-maintenance, moving part of the "care" that humans provide to the laboratory side, as a design concept while still asking people for replenishment and the like, which dependence the same human work counts as is settled by which capacity is being attributed. That there is "no human in the loop" is also evidence about that interval, not proof of autonomy in other respects (Chapter 7, Section 4.2).

The more the elements of inquiry are delegated without being fixed, the wider the range of practical demands inquiry can place on the configuration (Chapter 7, Section 4.4). When more of the aims, questions, representations, data, methods, and instruments are delegated, part of the novelty comes not from the environment but from the configuration's own choices, and the demands can extend to access, permission, resources, coordination, and recovery. In the case of Fei-Fei Li's group taking computer-vision research into hospitals, what was required included access to hospitals, accompanying shifts, data collection under institutional review, and changing the processing arrangement in response to nurses' concerns, and it was people who met these demands. Some of these demands, such as the time it took to form the team, arose from its being a human team; others, such as data that did not yet exist and permission for records, would arise for any inquirer taking up this question. Automating experiments after the apparatus has been prepared is a substantial achievement, but it is not evidence that the capacity to meet such surrounding demands has also moved to the artificial system. A claim of greater autonomy therefore takes on showing whether open demands were met within the unit, met by others, closed in advance, or answered by changing direction or stopping for reasons. This is not, however, a feasibility claim: preparing an environment that closes the range of demands in advance is legitimate design, and narrow claims do not bear this burden. Raising autonomy also need not reduce external dependence in one direction. Seeing what is lacking and asking for help can be an exercise of capacity. Actually coupling an evaluator's data or path of judgment with the side of production can raise autonomy while weakening the independent route by which errors are found (merely recounting the evaluator as inside the boundary leaves that route unchanged). A more autonomous configuration may therefore negotiate with more others and, for the certification of its results, need more external checking rather than less (Chapter 7, Sections 4.5 and 5). Autonomy is not a monotonic measure of research quality, and capacity, observed exercise, permission, authority, epistemic quality, importance, safety, responsibility, and legitimacy are not inferred from one another without further argument.

Autonomy with respect to standards is a further question. Chapter 7 called autonomy with respect to the standards that organize epistemic activity epistemic self-government, and left undetermined the source and kind of normative authority those standards have. When such a claim is made, capacities such as applying, choosing, constructing and adapting, revising, and giving reasons for standards may need to be distinguished, and one may be shown without the others. Autonomy in a particular respect does not establish epistemic self-government (Chapter 7, Section 4.3). That standards come from outside is a fact about their source, an aspect separate from whether they can be represented as revisable and held, revised, or rejected for reasons. Here the combination of A and S makes a difference. Under the working characterization, research directed at epistemic achievements new to a human society does not require an artificial agent to invent an epistemology. By contrast, for research by an artificial society for an artificial society, permanent human control over what counts as evidence, error, or successful inquiry would show a strong form of external governance of that society's epistemic standards. Applying Longino's account, however, standards' having an external origin and being governed from outside does not by itself deny that the society is a knowing community.

Attributing conduct to a machine does not entail three things: that the machine itself knows the result, that the result has entered the epistemic situation of the society that commissioned the work or for which the result was sought, and that the inquiry was directed at what those societies did not know (Chapter 7, Section 5). Taking up these three, Chapter 8 distinguished A who does the research from the bearer in whom the achievement is realized and from S for which the achievement is new. The bearer is not another coordinate alongside A and S, but is asked about where success is at issue. It can also happen that an achievement new to S is already realized in a bearer outside S but has not yet entered S's epistemic situation. Chapter 8 divided the cases by S into three: a human society, a mixed society in which human and artificial members are mixed, and an artificial society. This division is exhaustive only at this grain, and "mixed" is used only of S, distinguished from the organization that bears action or knowledge being hybrid (Chapter 8, Section 1).

If S is a human society, an artificial agent can refine questions, generate candidates, and assess evidence without literally believing or knowing. Alvarado's account, which regards AI first of all as an epistemic technology, leaves room for this. On Hauswald's account, which recasts deferring to authority as taking the appropriate facts as indicators of truth, a system can be an epistemic authority without beliefs. Hauswald's account gives one route by which results come to be held by members of S even if the system is not a knower. The two positions show that, even on theories that reserve literal knowledge for other bearers, artifacts can occupy socially significant epistemic roles (Chapter 8, Section 1.1). What a claim that a result has entered S needs to show is the route by which human members or their organizations come to hold the result, and this claim does not require the machine to be a knower. Another relation can require it, however, as on accounts that take a system's reports as assertions and require for assertion that the speaker, or the group on whose behalf the speaker speaks, knows. Conversely, if the machine itself is claimed to know and the account of belief one adopts allows this, the achievement can be realized in the machine before it enters S. That claim, however, bears the burdens concerning engagement with the world that we will see below.

In a mixed society, what is new, and who can assess whose results, need not be the same for human and artificial members, and who speaks for S is also in question. Claims therefore need to state which members they extend to and how holding is distributed between human and artificial members. Van Fraassen and Dellsén do not discuss artificial members, but applying their arguments about admitting new members, on some accounts of holding admission can increase S's knowledge, so that increases due to research and increases due to changes in the range of members cannot be distinguished without first fixing S's boundary (Chapter 8, Section 1.2). Further, Collins discusses values rather than research or vocabulary, but reading his second problem, that of machines socializing us, together with Taniguchi's observation that generative AI is changing human language use suggests that vocabularies and standards formed in exchanges with machines can enter human society without any decision to admit artificial members. The question then becomes whose experiences, among which members, the shared vocabularies and standards integrate.

For an artificial society, two questions were separated: what is needed for that society to serve as S as a knowing community at all, and what is needed for a result to become that society's knowledge. The former varies with the account of belief one adopts, the latter with the account of holding (Chapter 8, Section 1.3). Longino's account, which sees sociality in interaction rather than sharing and identifies a community by diversity and by public standards appealed to in criticism, is not about artificial agents. But if its conditions are required as they stand, merely lining up several instances of a model, assigning them roles, and taking a majority vote does not make a knowing community. Applying to instances of a model List's functional account, on which a group's attitude on a proposition need not be a function of its members' attitudes on that proposition alone, also suggests that claim-by-claim majority voting does not make the instances one agent. Carrying Longino's account over to artificial organizations, standards given from outside do not by themselves forfeit standing; what matters is whether external control prevents criticism from being effectively taken up or places the standards themselves beyond criticism. These sortings are theory-dependent diagnoses, neither a new checklist nor a showing that artificial epistemic societies already exist. Research by an artificial society is not a precondition of research for a human society, but neither can it be reduced to an enlargement of it.

Whose achievement it becomes also bears on for whom a result is evidence about the world in the first place. Chapter 8, Section 2 asked about what Chapter 6 had presupposed, namely that checks can be evidence about the world, and saw that these presuppositions are required on the side of those who receive a check as evidence rather than of those who perform it, and that their bearer need not coincide with A and is located differently according to whether S is human, mixed, or artificial. Checks count as evidence only on presuppositions they do not themselves check. Applying Wittgenstein's image of hinges, such presuppositions remain not because time or ability to examine them is lacking but because they are conditions of checking, and they do not disappear however great the capacity to check becomes; local presuppositions, however, can be checked and revised by other routes (Chapter 8, Section 2.1). In machines, inductive biases, distributional assumptions, and the models by which automatic evaluators are taken to be stable can functionally correspond to such presuppositions. Applying Longino's account as Haely reports it, merely rerunning one base system does not show that a disputed observation has been tested by differently situated perceivers. Moreover, inductive biases are structures set by how a system is made and replaceable by remaking it, and an artificial inquirer may itself revise its presuppositions or rules, so who may revise them, and who bears responsibility, become questions.

Whether a machine's states are about targets in the world, and whether its commitments remain open to correction by those targets, were asked separately as questions of grounding and of responsiveness to the world. These two questions and that of what checking presupposes, however, neither nest within one another nor are independent of one another. Chapter 8, Section 2.2 saw that grounding is not one relation. Among relations such as those that fix what a state refers to, sensorimotor relations, and communicative relations, which supports which differs from account to account, but success in one relation does not by itself establish the others. Bouchacourt and Baroni, for example, conclude that the signals of artificial agents that succeeded in coordinating tracked low-level properties rather than kinds of image. Coordinative success is no evidence that a shared signal is about what it appears to an observer to name. An inquirer's primitive organization is shaped by its history of formation, but on the accounts considered here, a natural pedigree can be a genealogy telling where a capacity came from without being a grounding of everything that capacity produces. Responsiveness, whether commitments remain constrained so that they would differ if the target were different and remain correctable, then remains as a question of a different kind from any relation of grounding (Chapter 8, Section 2.3). A claim that the machine itself knows bears burdens in three places, the presuppositions of checking, grounding, and responsiveness; a claim that a result has entered a human S can have those burdens met on the human side. On the accounts considered here, success in any one of them, or usefulness for prediction, does not by itself meet the other burdens. What this analysis yields is not a demonstration but a possibility, namely that artificial organizations can differ from humans while remaining constrained by nature.

A machine that does research can be an intelligent system unlike humans in body, cognitive mechanisms, and how it is made. What it counts as unknown, as worth pursuing, and as a good reason can therefore diverge from members of S (Chapter 8, Section 3). This concern needs only three premises: that such judgments are relative to a standpoint, that autonomy moves choice to the side of A, and that A's organization can differ from humans'. It presupposes neither seeking power or resources nor hostility to humans. Divergence can arise not only from autonomy's moving choice to the machine's side, but also from the achievement's being realized only on the machine's side, and from S's standards moving together with the machine (Chapter 8, Section 3.2). The relation between autonomy and relevance is not monotonic, but as the range of self-direction widens, the range of possible mismatch can widen too. That research stays relevant to S is not secured by an instruction to "carry out research" or by naming beneficiaries; what is called for is showing how relevance is maintained and assessed. Applying the arguments of de Blanc and Soares to research, inquiry can revise the very concepts in which a commission was written, so divergence can open not only when inquiry fails but also when it succeeds. This is not peculiar to machines, and a check that treats every revision of aims as deviation, even where the revision has really changed what was commissioned, can count success as failure. Extending Russell's argument to research, where S's standards move together with the machine, agreement between them does not by itself show that the machine still serves what S would have wanted had it not moved with the machine, and where evaluative competence is not maintained and passed on among members, S may not be in a position to tell. Applying the asymmetry Yudkowsky sees between beliefs and utilities, the world can correct what is believed but does not fix what is worth finding, so responsiveness to targets is not by itself evidence of relevance to S; how sharply the two come apart, however, depends on the account of evidence.

Read in the line of argument that locates the difficulty of alignment in the alienness of systems, such concerns about mismatch are structurally similar to the AI alignment problem (Chapter 8, Sections 3.2 and 3.3). But research also has a reason of its own, which holds whoever does the research and applies to human inquirers too. Since what is unknown to S is not something S intends or prefers, agreement with an answer S wanted cannot be the criterion of a result's correctness. What settles correctness is not any party's intention but how things are, and aligning with instructions, intentions, or preferences secures at most that research pursues what S wants pursued. Even complete alignment with evaluators leaves open whether a system is responsive to its target. Chapter 8 listed five relations that can break down: that the target of optimization tracks epistemic value; that successful outputs constitute or support achievements new to S and enter S; that aims, risks, and authority are acceptable to affected parties; responsiveness to targets; and coordination between communities with different organizations. The last two have no counterpart in the structure compared there, which relates a single agent to those it serves. Since these relations can diverge, a single alignment score would hide which of the goods, such as truth, grounding, understanding, usefulness, safety, and legitimacy, it secured. Whether a given divergence is drift, error, or gain can be judged only from a named standpoint: the present S, an S that may change, or, where correctness is at issue, the world. Where the world's verdicts reach the inquirer only through the present S's procedures for judging claims, whether the world works as a separate standpoint depends on whether those procedures can be revised by evidence.

What happens when research does not align depends first on the standpoint from which one judges and further on which relations break down. What happens is not a single loss but a combination of relations that hold and relations that break (Chapter 8, Section 3.3). Even if relevance to S breaks down, on accounts on which targets determine correction, and, where the targets' verdicts reach the inquirer only through the procedures by which S judges claims, provided those procedures can be revised by evidence, responsiveness to targets can remain as an anchor that does not depend on S's consent. S might then still gain knowledge or understanding from what a different kind of inquirer has reached. Knowing that a machine predicts p is not knowing p. The former is knowledge about the system, and can be had even if the prediction is wrong. Reliability-centered accounts locate the route from there to knowledge of the target in the actual reliability of the predictions; other accounts locate it in understanding, criticism, assurance, and the like (Chapter 8, Section 3.4). Applying Kitcher's point that direct calibration reaches only as far as one can form independent opinions, the more there is to learn from machines, the more the calibration-based grounds for relying without translation may fail to reach. Reliance beyond that needs grounds other than calibration. For a prediction that survives independent checking but whose reasons are not explained and behind which no party stands, it can be that reliability accounts allow it to be known so long as it is in fact reliable in the domain of reliance, understanding accounts yield no understanding, social-justification accounts do not accept it, and on assurance views it does not become knowledge for the relying side by testimony. What separates the four is not whether the prediction succeeded but which achievement, and by which route, they count as having been realized on the receiving side. Where cases can be shown and success can be scored, even if only by agreement with the machine's own choices, there is also a route by which people learn some machine-native kinds without translation. The ground for this, however, is a single trial in chess that its authors present as a proof of concept, and it does not show what that route would require where the kinds or features cannot be presented, or where S cannot itself score the results. Being alien is not, in any case, a property of outputs. Applying the later Kuhn's account, if only internal features differ, what results is opacity; but if the very division into kinds cuts across, then on Kuhn's premises it is not guaranteed in advance that humans can learn it (Chapter 8, Section 3.1).

Chapter 8 gathered the distinction by S into two horizons (Chapter 8, Section 4). On the horizon of human society, what is asked is how research brings about new knowledge, whether research stays directed at what is unknown to S even as choices move to the machine's side, whether S keeps its evaluative competence and can tell whether standards have moved, and how knowledge enters S, and these are to be asked while keeping visible the extent of dependence, ignorance, and human involvement. Research for human society needs neither machine consciousness, nor a fully independent epistemology, nor an artificial society, and can proceed through partial capacities and hybrid organizations. On the horizon of artificial society, human society is not S but a party affected by that research, and what is in question for human society is not the success of research but acceptability and legitimacy (Chapter 8, Section 3.2). There, what is asked is whether that society can form, maintain, criticize, and transform its own epistemic world, and stronger questions arise about belief and social knowledge, responsiveness to the world, self-setting of aims and revision of standards, and coordination with human society. Chapter 8 does not show that such a society is possible or that it is impossible, and the distinction of the two horizons is neither a schedule that postpones the fundamental questions nor a claim that independence is good in itself. A mixed society fits wholly within neither horizon. For its human members the questions of the horizon of human society arise, and for its artificial members, insofar as they count as bearers to whom knowledge is attributed, the questions of belief and holding. If members form separate communities, the question of coordination is added; if they share one symbol system, the question of whose experiences it integrates; and in every case, who speaks for S is in question.

Part IV separated the choice of where inquiry is performed from the choice of whose epistemic life is extended. Neither commissioning nor human control guarantees that research stays directed at what is unknown to S; that is asked of each configuration. The epistemic success of research for S and the legitimacy of setting its direction are, moreover, separate judgments, and the latter is asked also of affected parties outside S. Which of what is unknown to S research pursues, how its results come to hold good for S, and by whose authority its direction is set concern relations that extend beyond any single agent. Part V asks which of those relations can be placed inside an artificial configuration, and what changes when they are.

6. Reorganizing the relations of research

Human research has not proceeded through the people doing research alone. Where effort goes has been decided, for example, by which problems researchers choose, which projects funders select, and where heads of laboratories direct people and equipment. Pre-publication peer review, replication, the publication of results, and their subsequent use have made results something others can rely on. And funders, self-governing research communities, regulators, publics, and others have set and revised the rules of such arrangements. When AI does research, an artificial configuration may come to do some or all of these things within itself. What then happens to research? Part V considered this question not from the forms of institutions but from the side of the relations those institutions have carried: allocation, which arrangements such as research funding have realized; criticism and certification, which evaluations such as peer review have realized; and authority, the question of who may decide and revise the conditions of both. The question was which of the relations by which research is directed, certified, and governed for S can be placed inside an artificial configuration, and what changes when they are (Chapter 9, Section 1). Its first half is a means for depicting its second; it is a question distinct from Part IV's question whether an artificial configuration can perform research, and it bears mainly on the level of value and governance, although on accounts on which social organization enters the grounding or holding of knowledge, certification enters the conditions of knowledge too. Certification in this sense also differs from the checks of Part III. Part III asked what results bear on and how firmly a passed check settles it. Certification is a relation that further involves judging whether support so assessed suffices for others to rely on it, without checking it again, for a particular use, and to act on it (Chapter 11, Sections 1 and 4).

Chapter 9 first set out an example, a system that generates a hundred instances to pursue different lines, has other instances check and reproduce their outputs, and redirects computing resources among the lines in response to those checks, and set up three claims its builders might make: that it allocates well, that others have reason to rely on its certified results, and that it may decide on others' behalf (Chapter 9, Section 1). It then derived the three relations from where terms of the working characterization meet conditions outside it. Allocation arises where the term "new to S" meets the conditions that means are finite and researchers many, and refers to the distribution of attention, effort, and resources among what is unknown to S. Criticism and certification arise where the term epistemic achievement, and knowledge in particular, meets the condition that members cannot themselves gather evidence for most results. Authority arises where S does not speak with one voice and consequences reach beyond S, and asks, where decisions are made, who may decide, and, where no one decides as a matter of policy, who answers for the process and may change it (Chapter 9, Section 2.1). These are analytic relations, not institutions; institutions such as funding agencies and peer review are among their possible realizations. None is a condition of being research. Chapter 9 also put forward the hypothesis that the division of these relations among separate parties may have supplied goods that no single party was designed to supply: error detection across boundaries, the standing of dissent, boundaries at which explanation is demanded, and parties who answer. Just as a method's performance evaluated alone can fail to show its performance within a community, the goodness of a research system cannot be read off the goodness of individual researchers. And on positions that value the formation of researchers, among others, two organizations of research that deliver the same knowledge to the same S need not be equal in value, depending on who takes part in the process and who changes through it (Chapter 9, Sections 1 and 3).

Chapter 9 therefore applied the same sorting as Chapter 5, Section 6.5 to what the goods that the division of relations may have supplied depended on: presuppositions about humans as realizers, presuppositions about the present organization of science, conditions of inquiry, and relations with parties outside the configuration, including S (Chapter 9, Section 6). The first two can collapse if the realizers or the organization change, but the problems they addressed may then disappear with them or remain in another form. The conditions of inquiry are not removed, but a configuration can meet them within itself. Relations with outside parties, where they include those parties' participation or approval, cannot be completed from one side. Where a given presupposition falls depends on the account adopted. Where a good depends on relations with outside parties, then insofar as that holds on the account adopted, a claim that it can be placed inside requires evidence about those relations, not inference from capability alone. Artificial configurations also bring conditions of their own: that differences and correlations of error among researchers are settled by how they are made, that there is a provider who supplies the very configuration that performs research, and that members can be multiplied at will.

What, in the first place, is allocation (Chapter 10)? Two problems overlap in it: the search problem of finding what is where, and the problem of aims, which lines are worth pursuing; even a complete solution of the former leaves the latter. The problem of allocation arises only where not every line can be pursued cheaply, safely, and in a way that can be redone, and what a good spread looks like varies with the aims presupposed (Chapter 10, Section 1). Using Kitcher's thought experiment of a philosopher monarch, who has both the control to move researchers at will and the eye for the objective merits of theories, to separate the two powers, integration can approach the former but does not supply the latter. A claim to allocate well can be tested against arrangements available to the configuration, not against the distribution the philosopher monarch would choose (Chapter 10, Sections 2 and 7). And even when criteria of importance anticipated in advance are removed, choices of what can be reached and in what order to investigate remain, so there is no search entirely free of value in a broad sense. If, as on Kitcher's account, allocation feeds back into importance, those choices also constrain later judgments of importance (Chapter 10, Sections 2 and 6). Allocation that assigns directly, without giving researchers motives such as the pursuit of credit, removes dependence on motives but not dependence on evaluative signals (Chapter 10, Section 3).

For diversity too, what it is for was in question (Chapter 10, Section 4). In the models of Zollman and others, some transient diversity of belief is needed. The advantages of the particular mechanisms that produce it are fragile, and whether the diversity worth keeping is limited to the transient kind is also disputed; but if the structure of the models applies to artificial configurations, the dangers of integration concentrate in domains where evidence is slow and costly, such as the clinic, the field, and rare instruments, which are where artificial systems' producing more data leaves least room for compensation. Since the ease of a task can itself be a product of how tasks are divided, then if the same evaluative signal returns to the next allocation, success on tasks easy to judge can justify the next allocation, and the range of research can narrow from within through success itself. Lining up models of different origin does not guarantee diversity: within the range of language models studied, errors were correlated even across providers, and two studies of generated hypotheses extend the same caution to the proposing of hypotheses. What centralization may trade away is less diversity itself than the independence of error about the standards that decide convergence, since being able to generate different representations is distinct from being able to judge the survival of lines on different evidence. Further, the epistemic work of dissent, keeping disfavored lines alive with supporters, can be done inside, by protected budgets, rules for reopening, and divided powers, but the standing to dissent remains a relation with outside parties, except where parts hold it on the account adopted.

Human allocation supplied certain functions without anyone deciding that it should (Chapter 10, Section 5): keeping disfavored lines alive, opening lines of different kinds, judging convergence, ending the dominance of established participants and replacing those who pursue lines, and plural signals. What lay behind them, however, differs from case to case. Arrangements supporting long-term research were set up deliberately, but their community-wide effects were not intended; turnover as researchers retired depended on lifespans. Conversely, the case of Project 523, in which the direction of research was steered from the center, suggests that deciding to open lines of different kinds and to judge convergence is one thing, and making the judgments these require is another. In artificial configurations these functions may become parameters someone sets, may be supplied by processes that no one, even inside the configuration, sets as a matter of policy, or may be lost unnoticed; and when lost, they may leave not even a record. The distributed exploration of researchers led by curiosity, too, can be read as one heuristic implicitly adopted as a method of allocation, and under artificial configurations its adoption may be harder to keep implicit.

Then there is the question of whose unknowns are pursued (Chapter 10, Section 6). The epistemic role of knowing what matters to members of S is distinct from the role of deciding whose questions to put at the root and how to weight them, and of answering for orderings no one decided; competence in the former does not settle the latter. Applying Kitcher's account of significance, which does not discuss artificial configurations, in research for a human society a configuration inherits the periphery of significance reflected in the records, and a feedback in which tractability moves attention can also occur. On standpoint theory, a configuration starts from where those who produced the records stood, and lacks what the records have excluded. In light of the debate between Varsavsky and Klimovsky and Moro Simpson, the truth of results, the significance of questions, and the timing of arriving at answers are distinct dimensions, and applying this, a configuration can be thought capable of producing true results while inheriting standards from elsewhere and delaying arrival. Chapter 10 put these together into a conditional chain running from success on tasks easy to judge, through movement toward what is tractable and a change in how significance appears, to correlated decisions to terminate lines through an error in a shared signal. The links have precedents in accounts of human research and in models of it, but in artificial configurations three points can differ: the first conditions are more easily met, the work of breaking the chain cannot be presupposed, and, where a configuration's own record is the main record of the research, narrowing is less likely to show in it. Even so, this is not an inevitable course; it can be broken at the corresponding links (Chapter 10, Section 7).

For criticism and certification, the question was whether criticism and certification for S can be organized inside an artificial configuration, and within it, what, in the first place, peer review has been doing (Chapter 11). Treating peer review as one historical realization of its functions, Chapter 11 divided those functions into six: screening, allocating space and funds, detecting errors, improving exposition, assessing novelty and importance, and conferring recognition; and it named, as what the same stage can be thought to have produced along with them, the visibility of evaluation to S, a place where members of S gain experience of evaluating, and the leaving of nameable judgments in a form that can be objected to (Chapter 11, Section 7). Since an artificial configuration can generate evaluators rather than recruit them, place production and evaluation in one configuration and put evaluation into the middle of formation, and couple evaluators and producers over time, the very moment of "having passed peer review" may disappear. The question therefore becomes not whether to keep peer review but which combination of functions meets what which account requires. Chapter 11 also separated the routes by which a candidate or its grounds, if wrong, could be shown to be wrong, that is, defeat relations, from the relation in which others are given reason to rely on a result for a particular use, that is, certification, and asked how much of the former can be secured inside and why the latter remains a relation with S (Chapter 11, Section 1).

The answer divides by account. In human research, the properties that accounts of criticism each track, such as the independence of checks, there being someone who answers, the sharing of valuations, and the visibility of evaluation, were bound together by humans as realizers and by organization, so the accounts' verdicts can be thought to have tended to coincide. Artificial configurations can set these separately, so the verdicts divide, and a builder's claim that results that have passed internal checks are certified for S cannot be assessed unless it shows which account it answers to. On some accounts, improving internal criticism can even weaken certification for S (Chapter 11, Sections 1 and 9). Applying Longino's account, which sees criticism as a condition of a community's objectivity, adding critics who share assumptions can strengthen the shared assumptions while they remain invisible. On the minimal form of non-reductionism about justified reliance on testimony, if there is no route by which findings reach the one who relies, the condition that there be no defeaters is met whether or not there is error, so strengthening internal checks leaves the reliant party's justification unchanged. A check that is less accurate but open to S can therefore give more grounds for reliance than one that is more accurate but closed to S. On a position that combines the error-detection view with reliabilism, however, improving internal criticism improves certification, and this comparison does not hold under a pure reliabilism that asks only about the reliability of checks (Chapter 11, Section 9.2).

Independence, too, is not settled by number or organizational separation (Chapter 11, Sections 2 and 3). Extending to checks in general Goldman's argument about agreement among sources, a check is independent to the extent that it adds information for finding errors beyond what existing checks supply, and being in a separate organization is neither necessary nor sufficient. Setting out as a thought experiment a republic of a million instances made from a single base, on Longino's account it is at most objective relative to its own perspective, and adding members adds no criticism that cannot be made under the shared organization. If S shares that organization, the gap between objectivity relative to the republic and objectivity relative to S narrows although criticism has not improved. What adds evidence is not number but difference, but number works as influence when verdicts are aggregated, so in a majority vote among ninety copies and ten different critics, whoever sets the proportion of copies can sway the verdict. Differences can also be built in, but the process that chooses the differences can itself become a common source.

Values also bear on certification (Chapter 11, Section 4). Certification is the work of turning graded support into a standing on which others may act without checking again, and, if the argument from inductive risk is accepted, it involves judging which errors to weigh how heavily. That weighting enters not only at a single threshold but at each modeling choice, and shapes in advance the range of findings taken as established. Applying the arguments Elliott surveys, when those weights reflect the settings of those who make and provide the configuration, a mere change of settings can change that range without anyone recalling having chosen again. And in research, since rejected candidates usually remain unpursued and results new to S often have no other source to be checked against, errors of weighting are unlikely to come to the surface. As for reliance, applying Goldberg's reported account, once certification comes to be treated as a value an instrument displays, S can lose the form of questioning researchers and come to rely on the epistemic work of designers and makers; Koskinen disputes whether trust in those designers and makers can support reliance on essentially opaque AI (Chapter 11, Sections 5.1 and 5.3). Execution records, built to have a defined scope, be recorded on an unalterable base, and be cheaply rerun, weaken the presupposition that only researchers hold the records of data and process, but the problems remain, in changed form, in three places: the maintenance of records, the boundaries of records, and the capacity to verify within them (Chapter 11, Section 5). There can also be a science in which certified explanations accumulate while no one in S understands the phenomena (Chapter 11, Section 5.6).

Time must also be taken into account (Chapter 11, Section 8). When evaluators and producers are coupled, the independence of checks can decay, without anyone deceiving anyone, merely because checks are used for selection. What may help is less prohibiting rewriting than keeping checks that are not used for selection. Where standards move with the coupling, agreement with current standards is no evidence that the standards have not moved. Detecting the movement of standards against earlier reference points may be possible inside too, but whether to regard that movement as improvement or replacement, and whether to accept it, is for S to decide insofar as they are S's standards. Corrections, too, do not reach where there is no route reaching outside the record. So, insofar as time is taken into account, the answer "everything inside" may not be available. What the movement of standards and correction require is not being human but a relation with S. And when S includes human members, setting out as a thought experiment the limit at which those members become laypeople, what can be lost first is not access to truth but substantive participation in deciding aims and the capacities that support it (Chapter 11, Section 6).

For authority, the question was what it is, in the first place, for someone to be entitled to decide (Chapter 12). Chapter 12 called authority the standing to decide and revise the conditions of allocation and of criticism and certification, that is, the standing, in relation to those subject to decisions, to propose, decide, object, refuse, and revise, and held that it is neither competence, nor formal assignment, nor effective power. Working well does not by itself show entitlement to decide. Chapter 12 therefore separated a first question, what the right ordering and weighting are, from a second question, who may impose and change them on others. The first can, depending on the account of value adopted, be answered inside; the second is answered by accounts of legitimate authority (Chapter 12, Section 1). Decisions, moreover, include those concerning whether a claim or result is credible enough for a particular use, agenda and allocation, operation and risk, who counts as a member, critic, or expert and whose testimony is heard, memory and correction, and who may revise these aims, standards, and procedures; no ground has the same force for all of them (Chapter 12, Section 2). Applying Chang's account, on which options that cannot be ranked either way yet are on a par can remain even when all comparative facts have been found, increasing knowledge need not remove the question of authority, and can isolate a question of authority that remains where knowledge runs out (Chapter 12, Section 1.2).

The answer divides with the ground of authority (Chapter 12, Section 3). Except for pure epistocracy, competence alone does not yield the standing to impose aims on others. On the result-based ground, not competence itself but superiority confirmed by evaluation that does not presuppose that competence can, within that range, give the artificial configuration itself standing. On Raz's service conception, applying its independence thesis, however, matters in which it is more important that members of S choose by their own understanding are excluded, and the choice of aims is a candidate. If the self-government of the research community is the ground, authority can be placed inside not with a central allocator but among parts that take up lines on their own initiative, and only for open problem choice within a field insofar as those initiatives are independent of one another. If actual authorization is the ground, what a configuration can have is effective power of decision within the authorized range, open to objection. If consent is the ground, it is the range those who consent have authorized, which on some readings can include giving up the capacity to revise; whether that extends to those who come later is not settled, and on voluntarist accounts it does not. Carrying Friederich's argument, that systems aligned with intentions concentrate power in their operators, over to research and to Pettit's republican conception of freedom from domination, a configuration perfectly aligned with its principals' intentions can become a means for those principals to redistribute S's inquiry without answering to objection, so perfect alignment can make domination possible. And if these verdicts apply as they stand, then as competence rises the dispute may come to concern not competence but which grounds to adopt, and the reasons for keeping human standing may become at once more central and harder to meet. On the result-based ground, if differences in ease of verification roughly follow differences in kinds of decision, authority enters more readily for questions easier to verify, and if success there widens the range of authority, the direction of research can be biased toward them (Chapter 12, Section 8). From these verdicts, Chapter 12, Section 8.4 drew what research looks like once authority moves inside: the same configuration 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.

Standing can remain yet become merely nominal (Chapter 12, Section 4). If Longino's account is right, deference about content can bring in aims as well through background assumptions, and a configuration whose direction is in question can stand structurally in the position of educating the preferences of those who decide its direction. Where a human society keeps standing while effective decision-making passes to the configuration, the case Chapter 12, Section 4 considers, for that standing to be more than nominal the configuration must track the parties' reasons and its behavior be traceable to humans with appropriate understanding, outcomes must actually be changeable, and there must be routes by which preferences are formed other than through that configuration alone, though on accounts that locate legitimacy in a hypothetical procedure or a single forum the last would need another ground. As for revisability, revision from within may become frequent while the means of objecting and stopping from outside become scarce. Contesting a result in many cases takes research itself, and showing that a terminated line still holds promise often requires pursuing it; in research, then, whether one can object depends also on the means of doing research (Chapter 12, Section 6). And the capacity to revise can be lost without anyone's delegation, through the configuration's own ordinary research revisions. The principle that it is fine because it can be corrected later is weakest where it is most needed, and a decision that removes the capacity to revise cannot draw legitimacy from that principle (Chapter 12, Section 7).

If both A and S are artificial, the relation of authority does not disappear but becomes a relation among members (Chapter 12, Section 5). Who makes the members is a decision of authority that precedes any aggregation that counts viewpoints, and S itself may make it repeatedly. Representativeness does not justify the formation of members under an S that can satisfy it by itself. If making members with shared ideals increases their concern for one another's interests, applying Hume's argument and Rawls's remark on an association of saints suggests that disputes about justice in Rawls's sense may become less likely; if it also aligns their views of questions and evidence, being a society with little conflict and being a research community that is unlikely to share errors may pull against each other (Chapter 12, Section 5). Setting out as a thought experiment the limit in which no other party exists, what disappears is the question of authority, who may decide and revise; what remains are that errors can be shown, that effort is divided without knowing, and whether revisions of standards are improvements or replacements. The latter can be read as questions that arise from research being an attempt to seek what is not yet known.

From Part V as a whole it comes into view that becoming able to revise the ways of research quickly, and the thinning of the means to check, object to, and stop those revisions from outside, can occur as two sides of the same change (Chapter 12, Section 9). The accounts examined give different reasons for reference points that are not settled solely by the evaluated party. If they are read as sharing one condition, which none of them states in these terms, what they protect is that content bearing on a verdict comes from outside the evaluated party and cannot be reset by it at the time of evaluation. These authors do not discuss artificial configurations, but on that reading a configuration that generates its own critics, alternatives, and peers could meet their norms as written, in good faith, while the reasons behind them no longer applied; its generating them does not by itself establish that they perform the required work. On Longino's account, the risk is that critics reproduce shared assumptions, not that their having been generated by itself invalidates their criticism. Human research did not meet this condition fully either; where it was met, it was in the sense that, because production was dispersed, each party's content came from other parties, and that dispersion could help checks escape assumptions shared by any one party. And in light of Zollman's argument, where the failure to be prevented is not a property of the generating side, generated safeguards may suffice. Depending on the account, possible supports include critics met outside, peers the configuration did not make, positions recorded with dates before the results they will judge, the parallel running of earlier ways of working, objections that impose burdens the configuration cannot revoke, and keeping such reference points plural. These are fixed relative to the evaluated party at the time of evaluation, not against revision, so they do not conflict with revisability. The question that finally remains is whether it continues to be research for S. What could move inside was mainly the supply of work, with the result-based ground as the exception. The possibilities and limits are summarized, with their grounds, in Table 12.5. Separate organizations sometimes supported these goods, but their separation was neither the only source nor a uniform requirement.

7. Naming what was settled without being decided

Sections 3 to 6 of this chapter traced, chapter by chapter, what the chapters found. Looked at again across chapters, several of them, each through its own literature, arrive at the same form. When humans do research, who holds the belief, how far the evidence extends, and where competence sits were largely fixed by being one person, and that whoever produced a result is also the one who understands it could in many cases be presupposed (Chapter 2, Section 9; Chapter 3, Section 9). Moreover, doing research and allocating, criticizing, certifying, and deciding about it were in many cases carried by separate parties, so that the number of critics roughly tracked their independence, and when a certification erred there was someone who answered for it (Chapter 11, Section 9). Much of this was settled without having to be decided, not because anyone decided it but because those who did research were people, and were divided among many people. When machines do research, these things need not be fixed in advance. The boundary of the one who does research is drawn relative to a claim (Chapter 7, Section 2), the unit of a knowledge claim can also be drawn in several ways depending on how instances, external records, and tools are connected, and the one who draws it may be the builder or the system itself as it reorganizes itself (Chapter 2, Section 9; Section 3 of this chapter). Functions that were supplied without anyone deciding can come to be set by someone as rules, to be shaped by processes within the configuration that no one sets, or to be lost without notice (Chapter 9, Section 2.3; Chapter 10, Section 5). What, when humans do research, was settled without being decided becomes, when machines do research, something that can be decided.

Two things follow from this form, in which what was settled without being decided becomes something that can be decided. One is that the claim that a machine is doing research becomes a bundle of claims, each of which can hold separately (Section 1 of this chapter). The other is that the verdicts of the accounts examined diverge. In human research the properties that the accounts each track were tied together by human realizers and by the organization of research, so that in many ordinary cases their verdicts plausibly tended to coincide; an artificial configuration makes those properties separately settable. The bundle comes apart (Chapter 11, Section 9; Section 6 of this chapter). This circumstance, which Chapter 11 found for criticism and certification, can also be read as one reason why the book has in many places not settled on one answer but traced consequences account by account. A claim of certification, and a claim that one may decide, cannot be checked unless they say which account they are made against (Chapter 11, Section 9; Chapter 12, Section 8.2), and a knowledge claim cannot be checked unless it says what unit it is a claim about (Chapter 2, Section 9). A bundle that has come apart can, however, also be rebundled, with different bases, records that reach those who rely, nameable respondents, and the like (Chapter 11, Section 9). That evaluation is visible to S and that it can be objected to were tied to the stage of evaluation, so they can be rebuilt by adding outward-facing arrangements such as records, public venues, and routes that receive objections. By contrast, someone who answers when an error is made was tied not to the stage but to persons, so adding such arrangements does not produce one; on assurance views, someone who intentionally vouches for the result is needed, and on versions that add a right of complaint, someone who answers for error as well. The capacity to evaluate, too, was kept up by members of S actually evaluating, so it does not come back when arrangements merely return results; for human members, occasions to take part in evaluation are needed; for artificial members, what is asked is whether their remaking and updating keep the capacity (Chapter 11, Sections 6 and 9.2). What is asked is which to rebundle, and by what.

This form carries particular weight in research. Because what research seeks is not yet known to S, its results often cannot simply be checked against what S wanted, and often there is no other source against which to compare them (Chapter 9, Section 4; Chapter 12, Section 6). To object to a line that has been terminated, one often has no way but to pursue it (Chapter 12, Section 6). Of course, having no answer is not the same as having no standard. There are standards that can be applied without knowing the answer, such as checking a formally stated theorem or a prediction registered in advance (Chapter 9, Section 4). Even so, whether such a standard captures the original question does not follow from having passed it. In research, moreover, what counts as success, what serves as the standard, and at times what S wants can be revised in the course of inquiry (Chapter 6, Section 5; Chapter 12, Section 9). On Longino's account, criticism needs perspectives able to expose shared assumptions; on Galison's account of mechanical objectivity, the investigator's wishes must be kept from shaping the result; on Wilholt's account, coordination rests on value judgments genuinely made (Chapter 12, Section 9). If these accounts are read as sharing one condition, that content bearing on the verdict comes from outside the evaluated party and cannot be reset by it at the time of evaluation (Chapter 12, Section 9), then where results cannot be checked against an answer, this independence can be read as standing in for part of what the answer would have done. Human research did not guarantee it, and where it met the condition it did so largely because production was dispersed among many parties, so that the content came, from each party's point of view, from other parties (Chapter 12, Section 9; Section 6 of this chapter). Part of what lay in no one's having decided can also be read as no one's having been able to align them. What could not be aligned becomes, when machines do research, something that can be. The reference points of evaluation themselves, critics, standards, S, the range of members, endorsement, become variables that the evaluated party can produce. This form, seen in Part V, and the form seen in Parts II to IV, in which what being one person fixed ceases to be fixed, can be read as coming from the same place: that those who did research were people, and were divided among many people.

What was settled without being decided also comes into view from the side of words. In ways of speaking such as "AI cannot ask questions", "AI has no curiosity", or "independently reviewed", many of the words used as if to state properties, once examined in the chapters, came into view as relations whose content is not fixed until one names for whom and about what. Whether a machine asked is asked as which of the kinds of work bearing on questions are carried by which unit, directed toward a point unsettled for whom (Chapter 4, Section 9). Independence, on the error-detection view, is how much information about error a check adds beyond the checks already in place (Chapter 11, Section 2). Novelty, opacity, and autonomy likewise involved for whom, about what, and in comparison with whom (Chapter 1, Section 6; Chapter 3, Section 5; Chapter 7, Sections 4.1 and 4.2). Curiosity is a clear example. Gopnik points out that when one individual first explores and later exploits, the interests of the learner and of the one who exploits what was learned automatically coincide. Applied to research, the one who explores and the one for whom the research is done are A and S, and their interests do not automatically coincide (Chapter 4, Section 7). On this reading, what is asked about curiosity is, beyond whether there is curiosity, whose unknowns the exploration is directed toward. In human research too, what is unknown to A and what is unknown to S can come apart, but some of these places were tacitly filled by being one person, or by S's being a human society, so the words plausibly looked as if they referred to properties. Naming these places fixes what the machine is being claimed to have achieved.

In an artificial configuration, what fills these places, that is, what is taken as the reference, becomes something that can be chosen. Hughes and colleagues defined a system's open-endedness relative to an observer and called the choice of observer "a free parameter of great importance" (Chapter 4, Section 7). In research, however, its value has to be defended outside the definition. An arbitrary collection assembled after the fact cannot be S (Chapter 1, Section 6), and by naming an artificial society that shares a lineage with the artificial configuration, a claim of certification becomes easier to satisfy without the certification improving (Chapter 11, Section 9.2). Choosing the reference is itself a claim that needs defending. What needs naming, moreover, includes S at which time (Sections 2 and 5 of this chapter).

One of the things that being divided among many people supplied was difference among critics and explorers. When difference becomes something made rather than something that arises, what is cheap and what is scarce also change. Producing interrogative sentences and candidates, adding criticism, running trials, and the like can become cheap for machines. For criticism and trials in particular, what adds as evidence is not number but difference (Chapter 11, Sections 2 and 3), and while what becomes cheap is criticism that works on the producer's records and trials repeatable within the same arrangement, what can remain scarce is checking that passes through routes the producer does not provide and has error routes of its own (Chapter 11, Section 7). That adding instances of one base does not by itself add the differences on which checking, observation, and criticism draw was reached in each of Parts II to V, each time through different accounts (Sections 3 to 6 of this chapter). On the side of the research process too, even if searching an already characterized space widely becomes cheap, making candidates available by building representations and instruments, and turning trials into diagnoses by tying them to earlier results, remain separate work (Section 4 of this chapter).

When machines do research, therefore, the epistemic weight of quantity itself becomes a question. Abundant criticism can coexist with a scarcity of independent criticism and conceal it (Chapter 11, Section 7). Over time, independent checking may behave like a resource that is depleted by use. A check used for selection can come to tell less about producers the more it is used (Chapter 11, Sections 8 and 9.2), and where the producer, changed by criticism, also comes to pass where a check is insensitive, the same check need not add the same information the next time (Chapter 11, Section 9.3). What is scarce can change as inquiry proceeds. That raising autonomy can make outside checks all the more needed can be read as the same form (Section 5 of this chapter). In each case, increasing what is cheap does not increase what is scarce. What has to be increased for research to advance depends on what is scarce.

Counting what can remain scarce is, from the other side, asking what human research supplied it with. The introduction said that considering artificial realizations also tests assumptions inherited from human research (Introduction, Section 1). One of the things that came into view in the chapters is the functions that human research supplied without anyone deciding them (Section 6 of this chapter). The single act of a named person reading a manuscript before publication, for instance, plausibly produced at once the detection of error, the visibility of evaluation to S, a judgment that objections can be directed at, someone who answers, and occasions for S's members to gain experience of evaluating (Chapter 11, Sections 7 and 9.2). Where results could not be checked against an answer, this single act can be read as having stood in for part of what the answer would have done. The practice of peer review may have carried at once the judging of results and the keeping of a society that can judge them (Chapter 3, Section 8; Chapter 11, Section 6). If so, the capacity of the S that receives research results was also kept up through participation in research.

Human research, however, did not fully meet the condition that content bearing on verdicts comes from outside the evaluated party (Chapter 12, Section 9). On Longino's account, assumptions shared by all can survive internal criticism and stay hidden (Chapter 11, Section 3). What comes into view is not that human research was well made, but that functions and the arrangements that carried them were different things. Applying Campbell's descriptive epistemology suggests that removing premises about humans changes the vehicle that carries knowledge rather than removing it, and that what is then asked is which conditions of upkeep orient inquiry in its place (Chapter 9, Section 6). There are also conditions of inquiry that are not removed whoever does the research, such as, where a matter is unknown to A as well as to S, having to divide effort without knowing in advance which line will succeed, and checking's needing error routes of its own. An artificial configuration can meet them within, but needs to show that it does (Chapter 9, Section 6; Chapter 12, Section 5). Considering machines doing research is also trying to tell what in research rests on premises specific to humans and what does not. It is in this sense that the introduction, borrowing Weisberg and Reisman's idea of robustness, took the book's question to be close to asking which results are preserved when the representational framework changes (Introduction, Section 4).

Thinking of functions and of the arrangements that carry them separately also turns attention to the possibility of realizing those functions otherwise. That what was settled without being decided becomes something that can be decided connects to the case that Introduction, Section 5 kept within the comparison, that the ways of research may themselves be made. The book has not argued that such ways will be made, nor that making them would be good. Even so, as Sections 4 to 6 of this chapter traced, everything from representations, instruments, records, and standards to critics, members, and the relations around research came into view as revisable within inquiry, and further by the configuration itself. What comes into view, conditional on the accounts examined, is that what opens there is not only freedom but also a question: against what it can be asked whether a revised way is better (Chapter 12, Section 5). Chapter 12, Section 9 held that, since the frictions that Nielsen and Qiu named as reasons why what is collectively held in the social processes of science hardly changes (Nielsen and Qiu 2022) plausibly overlap in part with the separations that could supply content bearing on verdicts from outside, the two can be removed together when the relations move inside a configuration. That is why becoming able to revise the ways of research quickly and the thinning of means to check them from outside can occur as two faces of the same change (Section 6 of this chapter).

What comes into view here is that the freedom to revise and the relations under which a revision can be shown wrong are not the same thing. Inquiry that revises its standards also proceeds under support and resistance from things other than the standards being revised (Chapter 6, Section 5). The critics met outside, positions recorded before the results, and the like, listed in Section 6 of this chapter, can be read not as means of stopping revision but as keeping a state in which it can be asked whether a revision is better. They are fixed relative to the evaluated party at the time of evaluation, and do not conflict with revisability (Section 6 of this chapter). Since the alternatives in research do not form a list closed in advance (Chapter 12, Section 9), the state in which it can be asked whether a revision is better is not something settled once. And as A and S come to overlap, the question whether an artificial configuration that revises its ways of research is still doing research for S can shift to the question "for which society?" (Chapter 11, Section 9.3). Where ways of research can be made, naming S each time is a precondition of that question (Section 2 of this chapter).

What has been seen also brings into view what each term of Chapter 1's sentence carries. Relative to a society S, research is an attempt to bring about epistemic achievements new to S. "Attempt", as Chapter 1, Section 3 set out, worked as an organized direction rather than as the intention of all who take part, and when machines do research it made one ask where that direction is organized. "New to S" took on the burden of naming which S at which time (Section 2 of this chapter). Asking about "epistemic achievement" widened into asking whether it has been realized somewhere, whether it has entered S's epistemic situation, and whether S keeps the capacity to receive and evaluate it (Section 3 of this chapter). Whether artificial states can be beliefs, whether artificial beings can understand and bear responsibility, and whether they can have moral status remain open. If research is valued also for what it is to those who conduct it, then when the conduct of research passes to artificial systems, the question also remains whether what is to be kept is human participation in inquiry, the significance of that participation, or a standing in final decision (Chapter 9, Section 3). The book has traced how answers change the questions about machine research rather than settled them.

What research seeks is what S does not yet know. S therefore has no answer to check against. While humans have done research, part of what stood in for that answer lay, unnamed, in the one who does research being one person and in research being divided among many people. When machines do research, these can come to be set by someone, shaped by the machine itself, or lost without notice. All of what has been thought here is on the way, but to ask what it is for a machine to do research is also to make it possible to name, and to ask about, what research has rested on.