Chapter 8 — Whose Knowledge?

Human, Mixed, and Artificial Epistemic Societies

Chapter 7 ended with three things that do not follow when the conduct of research is attributed to an artificial configuration: that the configuration itself knows the result, that the result has entered the epistemic situation of the society that commissioned the work or of the society for which it was sought, and that the inquiry was directed toward the unknowns of either. Chapter 6, Section 6 had passed on the question behind them: when the actor conducting an inquiry, the bearer of knowledge, and the reference society S differ, whose epistemic change is being pursued, for whom is the result new, and who can evaluate or incorporate it? This chapter takes up that question through the three matters.

Chapter 7 located the attempt and the actor who conducts it. This chapter distinguishes that actor from the bearer in which the intended achievement would be realized and from the society relative to which it would be new. Knowledge remains the principal case: an artificial actor can conduct inquiry whose result becomes human knowledge; a hybrid organization, one made up of human and artificial components, can be both actor and bearer; an artificial community can pursue questions indexed to its own ignorance. Understanding raises parallel possibilities and questions, examined in Chapter 3. An actor can enable others to understand without itself understanding, and whether members' understanding by itself establishes collective understanding depends on the account of group understanding (Chapter 3, Section 6). The knowledge analyses below are not thereby theories of every achievement.

Keeping these possibilities apart raises a question: whether research that humans initiated will continue to produce knowledge for a human community when it is conducted by systems unlike humans. Research for a human community may be the intended and practically central case. But an artificial inquirer can be an intelligent system unlike human ones in its body, its cognition, and how it is made, among other things, and what it takes to be unknown, worth pursuing, or a good reason can then come apart from what the members of the society concerned take them to be. Greater autonomy widens the room for this, though it is only one of the channels through which it can happen. Neither the working characterization nor the fact that humans initiated the project guarantees that the research stays directed at that society's unknowns; Section 3 examines why, and what follows.

Section 1 distinguishes the actor from the reference society, and both from the knowledge bearer, the subject asked about when one asks whether the research has succeeded. It divides the cases by the reference society, that is, by whether S is a human, a mixed, or an artificial society, since the question is whose knowledge would be produced; it bears most directly on the second of the three matters with which this chapter opened, whether the result has entered the epistemic situation of the society concerned. Section 2 asks what it would take for the world to be available to an inquirer at all, that is, what a check presupposes, how states come to be about their targets, and how commitments, the claims an inquirer takes as true and as premises for inference and action, remain open to correction by those targets. These are questions that the first of the three matters, a claim that the configuration itself knows the result, would need to answer.

In this chapter, an epistemic organization is how an inquirer or a community is organized to treat something as an observation, as evidence, or as a correction, and also as unknown, as worth pursuing, or as a good reason, among other things; by extension, the phrase also names an inquirer or a community so organized. In neither sense does it name an institution. Section 3 asks what follows when an inquirer of another kind differs from S's members, in its epistemic organization and in such respects as its body, its cognition, and how it is made. The difference can show up as a mismatch in such things as what is taken to be unknown, worth pursuing, or a good reason. It then examines where such mismatch comes from (including the case in which a successful inquiry revises the concepts in which a commission was written), how it resembles what is known as the AI alignment problem, and what happens, and what S can still gain, when research by such an inquirer does not align with S. The third matter, whether the inquiry pursues the unknowns of the society concerned, falls within these questions and is not settled in advance.

Section 4 draws two construction horizons from the three cases and states what Part IV leaves to Part V. A construction horizon, as the phrase is used here, is the set of questions that a program for building artificial researchers must answer if it aims at research for a given kind of S. Each section uses the distinctions of the one before it: Section 2 uses the three cases of Section 1 to say who would answer its questions, and Section 3 takes up the possibility, reached at the end of Section 2, that an artificial inquirer could be organized differently from human ones while remaining constrained by the world.

1. Actor, knowledge bearer, and reference society

Chapter 7 located the conduct of research in an actor. This section asks what needs to be distinguished before one can ask whose knowledge that conduct produces, and how the answer changes with the society for which the knowledge would be new.

Chapter 1, Sections 2 and 6 introduced the two parties to a research claim, which serve here as coordinates. A denotes the individual or organized system to which research conduct or capacity is attributed: a person, a model-and-tool system, an artificial collective, or a heterogeneous human–machine organization. S fixes the reference for novelty and the society for which the achievement is sought (Chapter 2, Section 6). In the knowledge case, whether the resulting knowledge belongs to members, a group, or a society in some further sense requires a separate account. Society is used loosely and does not commit the book to one theory of collective belief or institutional acceptance, but S is not an arbitrary set assembled by the analyst; how much epistemic connection among its members, subgroups, records, or practices is required for claims of novelty and achievement relative to S to be intelligible is left open (Chapter 1, Section 6).

The notation fixes a social reference for novelty, not a single social subject that must possess every achievement. Knowledge of members, group knowledge, and socially available knowledge are distinguished by competing accounts: belief, authorized acceptance, reliable availability, collective deliberation, and group-level states have different roles (Chapter 2, Section 6 and Table 2.2). These are accounts of knowledge. For understanding, Chapter 3 distinguishes a member's grasp, organized collective capacities, and resources through which understanding can become available. Availability can matter to a novelty comparison; whether it is itself achieved understanding depends on the account (Chapter 3, Section 6). Assessing the novelty claim requires identifying its achievement and location and showing that its comparison with the prior situation in S is apt.

The bearer, the subject whose state or organization directly realizes the relevant knowledge, is neither a further coordinate nor one of the additional roles named below; it is asked about where success is at issue (Chapter 1, Section 2). It may be S as a group, some of its members, a specialist subgroup, or a wider sociotechnical organization; it may also be the actor A that conducted the research, or differ from it. Identifying that bearer locally does not replace S. Where the claim is that S knows, it makes explicit what is supposed to warrant that attribution. A claim about a bearer outside S instead needs an account of how its achievement is new relative to, and sought for, S; it need not attribute knowledge to S itself (Chapter 1, Section 6). If no defended account of possession supports the attribution of knowledge to S, whether through its members, its organization, or S as a group, the activity may still be research for S, and an achievement new to S may already be realized in a bearer outside it; what has not yet happened is that the achievement has entered S's epistemic situation (Chapter 1, Section 2). Additional roles, such as evaluators, beneficiaries, supporters, audiences, and authorities, may also differ from A and S. Identifying them where an inference depends on them prevents their silent collapse into one.

The cases below are distinguished by S: a human society or community, a mixed society with human and artificial members, or an artificial society. This chapter uses mixed only for a society S and keeps hybrid for organizations of human and artificial components that act or bear knowledge; a hybrid organization can appear in any of the three cases. The division serves to compare claims that results of inquiry have entered the epistemic situations of these societies, under the accounts of possession in Chapter 2, accounts of who comes to have a result as knowledge. Within that scope, it shows how the burden changes when knowledge is attributed to human members, to human and artificial members or their organization, or to an artificial society. Sections 2 and 4 also use it: Section 2 to say who would have to meet its conditions, and Section 4 to draw its horizons. The division does not classify every possible bearer from S alone: an artificial bearer can be relevant to research for a human society, and membership does not by itself establish knowledge. It is exhaustive only at this grain; it does not distinguish kinds of human or artificial society. The actor A can vary within each case, and can include a hybrid organization or people conducting research for an artificial society. Chapter 3, Section 9 locates related questions about understanding without assuming that every knowledge attribution has an understanding counterpart.

1.1 Research for a human society

When S is human, what would show that a result has become S's knowledge, and what roles can an artificial actor play? Here A may be artificial, substantially artificial, or a hybrid organization, while S is a human society, discipline, or specialist community. An artificial A may refine a question, generate candidates, operate instruments, assess evidence, and draft a result without itself literally believing or knowing. Human researchers or institutions may evaluate and incorporate its contribution. On individualist accounts, human members remain the direct epistemic bearers; stronger group or institutional accounts may attribute knowledge to the organized human system itself (Chapter 2, Section 6).

This possibility is familiar in the philosophy of technology and in social epistemology. In a paper cited here in its preprint version, Alvarado argues that AI is first and foremost an epistemic technology: it is deployed in epistemic contexts such as inquiry and operates on epistemic content, such as propositions and models, through operations such as analysis, prediction, and inference. Alvarado calls these operations epistemic rather than cognitive, and classifies the technology by what it does, since, on this account, an artifact is what it does or is designed to do (Alvarado 2023, §§2–3). The claim is not framed in terms of belief, but a classification by function of this kind does not require the system to be a believer or a knower, and so leaves room for the possibility just described: a system can take part in the epistemic work of an inquiry without being either.

Hauswald asks whether AI systems can be epistemic authorities, that is, sources significantly superior to a person in some domain, from whom that person can receive knowledge, understanding, and other epistemic goods (Hauswald 2025, 716, 722 n. 1). On what Hauswald calls the standard picture, a person learns from an authority through testimony, the intentional transfer of a belief from one agent to another, usually by assertion. The picture requires the authority to have beliefs and communicative intentions, which current AI systems most likely lack, and this counts against granting them the status (Hauswald 2025, 718). Against this Hauswald sets structural parallels. Users can learn from such systems; the relation between system and user is epistemically imbalanced, as the relation between authority and layperson is; the systems are opaque to their users, as experts' reasoning often is to laypeople; and users therefore face three problems that laypeople face with experts. One is telling which of two conflicting sources to trust, which Hauswald adapts from the problem Goldman posed for novices choosing between experts. A second is how far to defer, that is, how to revise one's own beliefs in response to the authority's. Much of the discussion of this problem concerns preemption, the thesis associated with Zagzebski that an authority's belief gives a reason that replaces, rather than adds to, one's other reasons. A third is how to obtain the reasons for an answer as well as the answer (Hauswald 2025, 718–720).

Hauswald considers three ways out of this dilemma between the standard picture and the parallels: denying that there are artificial epistemic authorities, which plays down the parallels; attributing beliefs and intentions to the systems, which risks treating them as humanlike; and revising the standard picture, which gives up common assumptions about authority. Hauswald favours the third. On the revised picture, to defer to an authority is to treat suitable facts about it as indicators of truth strong enough to lead one to adopt the attitude they indicate; facts are suitable when they bear a reliable relation to the truth of the propositions concerned, and access to them need not be testimonial. The outputs of an AI system can be such facts (Hauswald 2025, 720–722). On the revised picture, then, a system can serve as an epistemic authority for a person without having beliefs of its own; in research for a human S, this is one way in which a result can come to be possessed by members of S without the system's being a knower. Hauswald argues that the revision is not ad hoc, since it also covers a layperson who defers to a doctor not by accepting the doctor's testimony but by observing what the doctor does for the doctor's own child, and deference to a scientific consensus, which Hauswald doubts can be analysed as a group belief (Hauswald 2025, 720–721). Hauswald suspects, however, that preemption applied to the outputs of AI systems faces objections similar to those raised against preemption in general (Hauswald 2025, 723 n. 18).

Alvarado's and Hauswald's positions do not settle artificial belief. They show that an artifact can occupy an epistemically significant social role under theories that reserve literal knowledge for other bearers. What reliance on an artifact in such a role requires is taken up in Chapter 3, Section 5.3 and Chapter 11, Section 5; the first of these also takes up Alvarado's further argument that, if AI is essentially opaque, the epistemic trust that an epistemic technology calls for becomes significantly harder to justify.

Where the actor is a hybrid organization, some capacities may be attributable only to the organized whole. Artificial components may search spaces and preserve records; people may frame purposes and interpret significance; instruments may constrain claims through material interaction; institutions may organize criticism and correction. The whole can conduct research even if no component contains the complete question, evidence, method, and conclusion, which is the question, treated in Chapter 7, Section 2, of the boundary within which the conduct of research is attributed. Attribution to the whole does not erase differences among its parts. A machine can be a reliable source without understanding. A human can endorse a result without reproducing the machine's internal route. A laboratory can preserve a capacity that no current member possesses alone. Whether such an organization, with its records and retrieval, itself realizes a knower is a further question. Chapter 2, Section 6 recorded a dispute over this very case: on distributed or group-functional accounts (Table 2.2) an appropriately organized collective can itself realize a knower (for example Bird 2024), while on others distributed cognition does not amount to distributed knowing. The dispute over when reliance on an artificial source is justified (Chapter 11, Section 5) does not decide it.

The same distinction only partially resolves the case, in Chapter 2, Section 7, of a discovery that an artificial system has produced but that no one has yet read. The system may have produced an epistemically valuable artifact without the human society already knowing its content; whether the society already knows it depends on which of the possession accounts distinguished there is adopted. The engineering project need not settle that dispute in order to track production, availability, incorporation, and use separately. In this case, then, for the claim that a result has entered S's epistemic situation, what would need to be shown is how it comes to be possessed by human members or by their organization; that claim does not call for the artificial actor to be a knower in its own right, although other relations can, such as taking the system's reports as assertions on an account on which asserting requires that the speaker, or the group for which it speaks, know (Chapter 2, Section 6; Chapter 11, Section 5.1). Where the artificial actor is itself claimed to know the result, on an account of artificial belief that allows this, an achievement new to S may already be realized in the actor before it has entered S's epistemic situation (Chapter 2, Section 7), and that claim, unlike the claim of entry, carries the burden that Section 2 describes for an artificial knower.

1.2 Research for a mixed society

What changes when S has artificial as well as human members? This chapter calls such a society a mixed society (introduction to Section 1). An enterprise that aims at knowledge new to and attributable to it falls between the other two cases. To the extent that artificial members are to count among those to whom its knowledge is attributable, it raises for them the questions of Section 1.3 about artificial belief and social knowledge. What S comes to know may also be meant to be shared across its human and artificial members; to the extent that these form differently organized communities, which may not share the same primitive representations (representations built into perception and learning rather than fixed by explicit convention; Section 2.1) or standards, the enterprise also raises the questions of translation and coordination of Section 3.4. On the analysis of this section, what is new to such a society, and who within it can evaluate whose results, need not be the same for its human and artificial members. The question who may speak for S, which Section 3.3 takes up and to which Chapter 10, Section 6 returns for a society some of whose members are artificial, arises here as well.

Most of what this case would need to show combines what the other two cases require with the coordination of Section 3.4. Coordination in that sense is a term reserved for relations between differently organized communities, and where the two kinds of members form such separate communities, the question reduces to it.

If instead one applies Tadahiro Taniguchi's theory of symbol emergence systems (Section 2.2), on which symbol systems become organized from the bottom up through exchanges among members and the organized system constrains them, and considers the case in which both kinds join a single symbol emergence system, a question may remain that the combination does not cover. The vocabulary and standards that the two kinds of members share would then not be handed over from one repertoire to another but formed and changed within one loop that includes members whose bodies and senses may differ. If so, what is in question is not only how possession is distributed between them but whose experiences, and of what kind, the shared vocabulary and standards integrate. Whose experience enters a symbol system can be asked in a human society too (Chapter 10, Section 6). What differs in a mixed society is that the difference crosses the line between human and artificial, and so connects with the question of Section 1.3 whether artificial members have experiences to be integrated at all. This question adds to those of the two horizons within the mixed case.

Where a human society ends and a mixed one begins may itself become unclear in practice. Collins, in a passage on large language models and society rather than on research (the account of research science in the same paper is taken up in Section 1.3), distinguishes two problems: that of socializing the machines, aligning them with us, and, if that is not solved, what Collins calls the dissolution of societies, in which the machines socialize us, aligning our values unnoticed with theirs, which on this account are either no values or the values of their hidden controllers. At present, Collins adds, those who decide what the machines' socialization comprises appear to be the founders of AI companies (Collins 2025, 1259). Collins writes about values, not about research or vocabularies. Read together with Taniguchi's observation that generative AI is already changing human language use (Section 2.2), however, Collins's second problem suggests that a society counted as human can come to share vocabulary and standards shaped in loops with machines without any decision admitting artificial members, so that the question whose experiences that vocabulary and those standards integrate can arise before the question of membership is settled.

Collins's point concerns a society that comes to share vocabulary and standards without deciding on membership. Where a society does decide to admit new members, what the admission does to what it can claim has been examined for observability. Van Fraassen, for whom observable means observable by us, the epistemic community (Chapter 1, Section 6), considers a future in which dolphins are admitted as bona fide members of our epistemic community while we still accept a theory on which they reliably detect things we now classify as unobservable. To an argument of William Seager's, that we must therefore already believe that those things exist, van Fraassen replies that what we would truthfully say after the admission, that some of us observe them, does not license our saying now that they will be observable, since us would then refer to a wider community. This reply concedes that a harder question remains: how we are to regard at present the opinions we foresee coming to hold once the admission becomes likely (van Fraassen 2005, 7–11). The admission itself is modelled on what van Fraassen calls epistemic marriage, a union after which the partners hold only communal opinion; marriage, on this account, is more intimate than the dolphins' admission and is clearly a matter of decision rather than something forced by opinions about how things are (10). Against settling on a point between the partners' opinions, van Fraassen argues from the reflection principle, roughly that one's present opinion must lie within the range of the opinions one foresees coming to hold, and proposes instead that the partners make their common opinion vaguer, spanning both, and adopt a common policy for revising it in the light of the evidence (19–21, 24–25). Dellsén, discussing van Fraassen's account in the preprint version of a chapter, distinguishes extending what a community can observe by admitting observationally superior beings, which the chapter calls epistemic immigration, from extending the observational capacities of existing members, which it calls epistemic upgrading, and notes that whether to allow the former is at least partly in the community's power (Dellsén 2024, 8, 12). Neither discusses artificial members; van Fraassen mentions AI only to say that the challenges discussed there do not need it (5–6).

Carried to a mixed S, Dellsén's distinction separates two ways in which what lies within S's reach can change: through what members can do, and through who the members are. In van Fraassen's story the admission is debated and decided upon; what S could truthfully claim after admitting artificial members is not thereby something S can claim before. In the terms of this chapter, an achievement already realized in a prospective artificial member may, on some accounts of possession, come through the admission to count as one realized in a member of S. It does not follow that S lacks it before the admission, since van Fraassen's argument concerns whom us refers to, not the acquisition of knowledge from outside; even before the admission, S may have gained knowledge supported by that achievement by drawing on the system as an outside source (Sections 1.1 and 3.4). On van Fraassen's proposal, moreover, the common opinion formed at admission is at least as vague as either party's, spanning both rather than settling on either, so that admitting artificial members does not by itself settle what the mixed S holds.

In this case, then, a claim that a result has entered a mixed S's epistemic situation would need to say which members it reaches and how possession is distributed between its human and artificial members. Admission does not by itself settle that. If, moreover, admission can increase the knowledge attributable to S, as noted above, not everything S comes to know is the result of new inquiry; an increase due to research and an increase due to a change in who the members are cannot be told apart unless S's boundary is fixed, and a claim of novelty for research for a mixed society would say which of the two it rests on. And, as Collins's point suggests, vocabulary and standards shaped with machines can begin to enter such a society before any decision about membership.

1.3 Research by and for an artificial society

What would it take for an artificial society to be the society for which research is conducted, and what further would it take for a result to enter that society's epistemic situation? Here S is artificial and, in the central case, so is A. Artificial agents or an artificial organization identify unknowns relative to their own epistemic state, conduct inquiry, assess and revise claims through their own coordinated practices, and seek achievements new to that artificial society and for it, which on some accounts would, once they entered its epistemic situation, be knowledge attributable to that society. The two questions differ. To serve as S at all, the society must be such that its epistemic situation, through its members, organization, or records, could come to include achievements of the kind sought (Chapter 1, Section 2), so that questions of artificial belief and possession arise even where it is only the reference. Whether a result has in fact entered its epistemic situation is the further question. The former is taken up first, by asking what it takes for an artificial society to be a knowing community at all and sorting the conditions by accounts of belief; the discussion then turns to the latter, asking, account of possession by account, what a result's becoming such a society's knowledge requires.

The central case is not just a larger multi-agent system. As Kraft reports Longino's account, the sociality relevant to knowledge is interaction among diverse members, and a cognitive community is identified by the public standards appealed to in criticism, rather than by shared beliefs (Kraft 2025, 61–62). This account is not about artificial agents, but if its conditions are asked of them, a set of interacting artificial agents does not by that fact establish an artificial knowing society. For the same reason, multiple model instances, role prompts, or majority votes do not by themselves establish a knowing community. On a functional account such as List's, a suitably organized collective, though not a mere aggregate, can be an agent, and the question becomes how the members are organized. List's comparison of group agency and artificial intelligence applies such an account to both cases at once: it treats suitably organized collectives and sophisticated AI systems alike as intentional agents when they meet functional conditions of representation, motivation, and action (List 2021, 1213, 1215–1220, 1226–1227). List also notes that formal results back the claim that a group's agency may be irreducible to its members': even where a group's attitudes are determined by its members' attitudes, its attitude on a proposition will not in general be a function, such as the majority, of their attitudes on that proposition, and collective rationality may require that it not be (List 2021, 1225 n. 37). The simplest case behind such results, known in jurisprudence as the doctrinal paradox, shows that majority voting proposition by proposition can leave the group's judgments as a whole inconsistent, with a majority accepting each premise and a majority rejecting the conclusion they entail (List and Pettit 2002, 93–96). Applied to model instances, this suggests that a majority vote taken claim by claim among them would not, even on a functional account, be what makes them one agent.

Collins's account of research science adds a different condition. On this account, frontier research and a child's primary socialization, in which, on Collins's scheme, newborns and toddlers acquire in the family their native spoken language and moral concepts, have in common that both establish things new to the learner, that both do so in small, bounded, trusting groups relying on face-to-face communication, and that both depend on moral integrity; members of a core group, the small group that carries out such frontier research, find that their collective search for correspondence truth will fail unless they cleave to what Collins calls moral truth, a determination to tell fellow members the truth about their observations (Collins 2025, 1251, 1253, 1255–1256). For a large language model, on the same account, what stands in for socialization begins most of the way through human secondary socialization, the stage at which, on this scheme, children at school learn written language and early technical concepts (higher education and specialized occupations form a further, tertiary stage), and is restricted to the internet, where there is "no stability, no constancy and no truth" (Collins 2025, 1251, 1253, 1256–1257); these claims concern the language models available when the paper was written, and Collins allows that the pathologies described there might be overcome (Collins 2025, 1262). Read as a condition for an artificial community to be a knowing community, a question Collins does not take up, the argument about core groups implies that, beyond public standards to which members appeal, interacting artificial agents would need something that plays the role of a core group's commitment to report truthfully to one another. What the commitment protects, on this account, is members' reliance on one another's reports of their observations; where artificial members share the records of observations themselves, its counterpart may lie less in a commitment among members than in the integrity of the records and of those who maintain them (Chapter 11, Section 5.3). Section 2.2 describes an experiment in which artificial agents coordinated successfully on signals that tracked something other than what they appeared to name, and Chapter 11, Section 3 asks, for critics that can be made, what criticism among agents derived from one base system, and among agents made to differ, can and cannot detect.

Standards that are shared and open to public criticism, as Longino's account requires, do not by themselves require a community to be sovereign over them. Carried to artificial organizations, the account therefore implies that external origin or governance does not alone disqualify an artificial organization as a knowing community. The epistemic concern arises where external control prevents effective criticism or uptake: for example, where a justified objection cannot alter the treatment of evidence, or where the standards themselves are placed beyond criticism, since on this account the standards that regulate critical interaction are themselves outcomes of such interaction (Longino 2021, 14–15) and must be publicly accessible (Longino 2025, §5). Whether the organization can itself revise its governing standards, beyond their changing through criticism in this way, is a further question about epistemic self-government (Chapter 7, Section 4.3). Chapter 12 asks separately who is entitled to govern such revisions.

The conditions such a community would need to meet depend on the account adopted. Three of the families of accounts of belief that Chapter 2, Section 2 listed give one way of sorting them, when applied to artificial members and their organization: functional, interpretive, and normative accounts. Phenomenological and enactive accounts, which are not among those Chapter 2 listed, give a fourth; Section 2.2 takes up some of the enactive ones. These are theories of when mental states such as belief can be attributed. They answer a different question from the accounts of social possession that Chapter 2, Section 6 distinguished (Table 2.2), which ask what bears knowledge and through which relation. How far an artificial society must answer the question of when belief can be attributed to its members and organization depends on the account of possession adopted, as the discussion of possession below shows. The sorting of accounts of belief into these four families serves only to show that different theories would impose different conditions; it does not take up separately the representational and dispositional accounts that Chapter 2 also listed, the second of which Chapter 2, Section 2.1 treated together with functional accounts. A functional account may look for stable belief-like states, differentiated roles, shared and local memory, coordinated action, criticism, and error correction. An interpretivist account, such as the view of belief attribution as interpretation on which Davidson's argument rests (Section 3.1), may ask whether the organization is intelligible as maintaining and revising commitments. A normative account, on which belief is constitutively subject to norms of truth and rationality (Chapter 2, Section 2), may require participation in practices in which commitments are held to such norms. A phenomenological or enactive account, such as those taken up with Di Paolo in Section 2.2, may require forms of experience, concern, or self-maintaining normativity whose presence in current systems is in question. The features listed here are accordingly theory-dependent diagnostics, not a new checklist or an operational definition of society.

Turning to the second question, the accounts of possession in Table 2.2, applied to an artificial S, sort the conditions in a second way, and how much they ask of the accounts of belief differs among them. On individual-possession accounts (on which it suffices that suitably situated individuals know), member accounts (which require that enough members, or the members with decision-making authority in the relevant domain, know), and relational accounts (which understand a group's belief through members' beliefs and the relations among them), a result's becoming such a society's knowledge requires, in different ways, that its artificial members believe it, so that the question returns to artificial belief; on relational accounts it also turns on whether epistemic relations hold among them. On accounts of distributed cognition without distributed knowing, such as Giere's, the larger system can perform cognitive work while knowledge remains attributable to individuals, so that the question likewise returns to whether artificial members can be knowers; Giere's own restriction of agency to humans does not settle this case, since it is defended as a theoretical choice for systems in which humans conduct the research (Chapter 2, Section 6). Group-attitude accounts require a belief or acceptance that the group forms, and they differ in what they ask of members. On the joint-acceptance account that Gilbert states, the members must openly express their willingness jointly to accept the result, whether or not any of them personally believes it (Chapter 2, Section 6), so that the question is whether artificial members can take part in such joint acceptance, not whether they believe the result. On distributed or group-functional accounts such as Bird's, it turns on organization: whether the members together realize an appropriately organized cognitive system, a question parallel to the one List's functional account, above, raises for agency when it separates suitably organized collectives from mere aggregates. On accounts of knowledge without a knowing subject, neither the artificial members nor the artificial collective need be a knower. Giere, while attributing knowledge to individuals in the distributed-cognition account above, also offers, for the status of scientific knowledge, an account of this kind, an adaptation of Popper's epistemology without a knowing subject (Chapter 2, Section 6). On that adaptation, what gives the result the status of knowledge is certification by reliable, accepted procedures and public acceptance, and the question becomes whether certification by artificial members can serve as such a procedure, which, where the members are instances of one base system, is the question that Chapter 11, Section 3 puts to their certification. On the critical-contextual account that Longino gives, the question joins the conditions on community discussed above: whether the interaction of artificial members involves the diversity and the public standards of criticism that the account requires of a knowing community. On none of these accounts does the connection among the members suffice by itself, but what more is needed differs with the account. Nor does anything in this sorting show that an artificial epistemic society already exists.

The case changes what a construction program is asking for. Research for a human society can inherit human concepts, questions, evidential norms, and social uptake. Research by an artificial society asks whether artificial beings can constitute and correct an epistemic life of their own. The second is not a prerequisite for the first, and it cannot be reduced to a scale-up of the first.

Whether research by and for artificial beings would be science at all is itself answered differently. Humphreys, whose separation of science done by humans from science done for them Chapter 1, Section 2 noted, gives an affirmative answer to a more general question, whether there could be science done neither by nor for humans. His imagined case is quoted by Schrape. After humans have become extinct, a rocket built before their extinction reaches a distant planet, samples conditions there that no human knew or predicted, and engineers the genetic code of a new cognitively aware species able to survive and flourish there, which colonizes the planet and adapts it to its own purposes, without any human knowing anything specific about how that part of the universe has been controlled. This, Humphreys holds, is still science, indeed scientific knowledge in operation: an activity need not be done either by us or for us to count as scientific, as quoted in Schrape 2015, 162–163. Read with the distinctions of Section 1, the scenario keeps apart who set the activity going, the humans who built the rocket, who possesses its results, no human, and whose purposes its outcome serves, those of a new biological species that is not human. Nickles, quoting the preface of the same book, where Humphreys writes that human abilities are no longer the ultimate criteria of epistemic success, disagrees: a pragmatic conception of the goals of science as conditioned by human interests implies, on Nickles's account, that a strain of anthropocentrism will always remain in our science, since "it is still we ourselves that we are extending"; the intelligent beings of the planet Zork, Nickles adds, will have a Zorkocentric conception of the universe (Nickles 2019, 365). In Humphreys's case, too, the outcome serves the purposes of a species, though not a human one; what Nickles's objection turns on is rather that no science is free of the perspective of those whose science it is. Ratti, arguing that machine learning in biology depends on a judgment akin to phronesis, practical wisdom, which is shaped by the way humans are, "leaves open the possibility of a scientific practice AI-tailored" but admits to not fully understanding the claim, "because we do not know any other science that is not human" (Ratti 2020, 18). Tollefsen, in the abstract of a paper presented at a 2018 meeting on whether science will remain human, calls the project of a science "populated and exercised by machines" chimerical: were science merely a technical pursuit within the realm of instrumental rationality, there could perhaps be a science carried out by machines, but on Tollefsen's account science is fully human action, to be governed by sound practical reasoning (Tollefsen 2018). On Humphreys's account as Schrape quotes it, science need be neither by nor for humans, which leaves room for such research to be science; on Nickles's, read through the case of the Zorkians, research by and for an artificial society would be a science centered on that society's interests, though Nickles does not there discuss artificial inquirers; on Ratti's it is unclear what such a science would be; and on Tollefsen's there could be none. Tollefsen's verdict rests on the premise that science is fully human action governed by practical reasoning. Totschnig, and Tubert and Tiehen, who argue that artificial agents could come to revise their final goals (Chapter 7, Section 4.1), and so to reason about ends themselves rather than only within instrumental rationality, would contest that premise. Ratti ties the judgment that science depends on to the way humans are; put as a judgment shaped by the way its inquirer is, the point would leave room for a science shaped by the way an artificial inquirer is, the possibility Ratti leaves open without claiming to understand it. Besides Tollefsen's, the accounts cited in this chapter that come nearest to a negative answer are Collins's and Rolla's, each with its own qualification. On Collins's account, frontier research depends on core groups whose members cleave to moral truth, and what stands in for socialization in a large language model, restricted to the internet, does not supply it, so that interacting systems of that kind would not form a research community; Collins writes about the language models of the time and allows that the pathologies described might be overcome (Collins 2025, 1256–1257, 1262). On Rolla's account, a genuinely cognitive AI would first require artificial life (Section 2.2), so that until then there would be no artificial inquirers for such a science to be by or for; the argument concerns cognition rather than science (Rolla 2026, 1, 11, 14). Neither holds that such a science is excluded in principle.

2. How the world becomes available to an inquirer

Section 1 left open what an artificial bearer or society would need in order to know anything at all, and allowed that, for a human S, much of what a result needs may lie on the human side. Beneath conditions on belief and possession lies a question that explicit method does not reach. A system may execute a proof rule, statistical test, or experimental protocol. But what makes a result count as an observation, a repetition, a failure, or evidence bearing on one commitment rather than another? Chapter 6 asked what the result of a check bears on and how far it supports a claim, presupposing that a check can be evidence about the world at all. This section asks about that presupposition. These presuppositions are required not of whoever runs a check but of those for whom the check counts as evidence. As Section 1 showed, that bearer need not coincide with A, and where it lies changes with whether S is human, mixed, or artificial. Further, as Section 2.1 explains, on Longino's account as Haely reports it, whether others can access and test an observation can bear on its standing as evidence; where the community of perceivers is S, this is a condition on relations within S.

The section separates three questions. The first is what a check presupposes when it counts as evidence at all (Section 2.1). The second is how a system's states come to be about worldly targets, the question usually discussed under the name of grounding (Section 2.2). The third is how its commitments remain open to correction by those targets, which this chapter calls world-answerability (Section 2.3). Separating the three questions keeps apart three achievements that a phrase such as tests hypotheses against the world attributes together. The three are not, however, nested, each containing the next, or independent of one another: the accounts discussed below relate them differently. Longino, for example, as Haely reports her account, makes whether a perception counts as an observation depend on its accessibility to others and, in controversial cases, on testing by different perceivers; since testing by others is one of the routes by which an organization remains corrigible, the first and third questions meet in one social relation (Section 2.1). On Coelho Mollo and Millière's account, what fixes a state's reference does not settle whether the system knows anything; and since a state can be about a target while misrepresenting it, whether the system's commitments are corrected when they misrepresent is a further matter, so that the second question can be answered while the third remains open (Section 2.2). On the account of John McDowell, discussed in Section 2.3, what a thought is about cannot be separated from its answerability to the world in McDowell's sense, a rational relation that differs from the sense this chapter gives the term; on that account, the second question cannot be answered apart from a question of answerability of that kind. The order of the subsections does not mark a logical order among the three questions. Section 2.1 continues the discussion of checks in Chapter 6, and the answerability to the world of Section 2.3 is what Part V builds on (Chapter 11, Section 1).

What bears the answers to the three questions differs according to whether S is a human, a mixed, or an artificial society (Sections 1.1–1.3). In research for a human society, unless the artificial actor is itself claimed to know, the answers may lie on the human side of the arrangement that conducts and receives the research, in human concepts, instruments, evaluators, and institutions, or, where the actor is a hybrid organization, in the organized whole (Section 2.3). For an artificial society, insofar as it or its members are claimed to know, and for the artificial members of a mixed society insofar as they are to know, the answers would have to hold of that society or those members themselves, to the extent the adopted epistemology requires.

Nor does the artificial case have to begin from one coherent optimizer whose beliefs, evidence, and values are all represented in a single place. The relevant organization might be distributed across agents, memory, instruments, environments, learned dispositions, and institutional rules. Whether that wider arrangement is merely a causal system, an inquirer, or an epistemic subject is a further attribution problem (Chapter 7, Section 2); assuming a unitary artificial mind would decide it in advance.

2.1 What a check presupposes

What does a check (in the sense of Chapter 6) presuppose when it counts as evidence? Suppose there are a million apparently similar observations. By themselves they are not evidence that the same thing happened again and again, since observations do not say what they are evidence of. Treating them as repetitions presupposes that the events are instances of a relevantly stable kind, that conditions are sufficiently similar, that the instrument tracks the intended target, and that the record can support an inference about another case, among other things. None of this is established by the observations themselves. The examples below share this structure rather than a single presupposition: in each, a check counts as evidence only against something that the check does not itself establish. This does not mean that no presupposition can be checked without circularity. Some can be checked by other routes; others cannot be checked all at once, because any check of them would rely on them. The end of this section returns to the difference. Nor does dependence on presuppositions mean that a knower must first prove them: on reliabilist accounts, for example, it is enough that they in fact hold (Chapter 2, Section 4.1).

Moreover, these presuppositions need not be consciously endorsed beliefs. They may be so deeply organized into perception and action that the world appears continuous without an explicit judgment that it is. While walking along a familiar path, for example, one does not treat its disappearance beneath the next step as a live possibility. Vision, touch, movement, memory, and prior interaction make objects, surfaces, recurrence, and resistance available before a scientific test is selected. Scientific methods extend and discipline these capacities and can correct them, but they do not begin from an uninterpreted world. Some of what is presupposed may consist in high-confidence propositional attitudes, things explicitly believed. Other parts may be things hard to call beliefs, such as perceptual capacities, practical expectations, learned dispositions, architectural biases, habits, trust relations, or socially stabilized roles.

What, then, is presupposed? Even within hinge epistemology, which develops Wittgenstein's image of hinges discussed below, accounts disagree over whether the basic certainties are beliefs, propositions, commitments, entitlements, or roles within a practice (see, for example, Pritchard 2021 and Fairhurst, Pérez-Escobar, and Sarikaya 2026). This book therefore uses primitive epistemic organization to refer to such presuppositions without settling that question: a loose umbrella for the part of an inquirer's epistemic organization, in the sense given at the opening of this chapter, that is not fixed by explicit stipulation, not established neutral terminology or a commitment to hinge epistemology. As the examples below show, however, presuppositions of checking are not confined to these; they extend to stipulations made explicitly within methods.

The examples below follow two threads. One is the presupposition that cases are relevantly the same; the other is the presupposition that a route or source of evidence is reliable. The division is taken from what the authors cited below themselves treat as presupposed, and its purpose is to show the same structure recurring at different scopes, from local to global. The two are not exclusive. Alston, for example, whose analysis of perception is one of the examples below, argues that the reliability of sense perception as a whole cannot be shown without relying on perception itself, and sketches the same predicament, that the reliability of a source cannot be shown without relying on that source, for the other basic sources of belief, induction, deduction, memory, and introspection (Alston 1988, 436). Induction is thus both the general form of the first thread and one of the sources Alston names. Nor do the examples exhaust the field: memory, testimony, records, and mathematics are left out. Were records and sources taken up, the same structure would presumably appear in checks in history or philology, for example.

The first thread begins with induction, which asks in general form whether the million observations above can be extended to cases not yet observed. On the reconstruction Henderson adopts, Hume's argument has two horns: no a priori reasoning can establish that unobserved cases will resemble observed ones, since the contrary involves no contradiction, and any argument from experience for that resemblance would presuppose it and so argue in a circle (Henderson 2022, §§1–2). Henderson also notes that learning theory gives the first horn a form that applies directly to whatever can be formalized as a learning algorithm: learning algorithms are effective only if they carry an inductive bias, and no-free-lunch theorems can be read as ruling out, wherever their formalism applies, a universal, model-independent justification of their performance, while leaving room for guarantees relative to a particular model and prior (Henderson 2022, §3.4).

Where a check's verdict is reached through statistical methods, this presupposition is stated explicitly, and at two scopes. Otsuka reads the probability model of inferential statistics, together with the assumption that it stays the same across samplings, as a mathematical model of Hume's uniformity of nature: the probability model is never observed as a whole, and it is this assumption that underwrites inferring unobserved cases from the data. Otsuka distinguishes the probability model from a statistical model, a family of distributions with finitely many parameters that statisticians posit about the form of that uniformity when samples are too small for the asymptotic guarantees the uniformity provides (Otsuka 2021, 3–7). If so, when passing a check on some data counts as evidence about another case, the verdict rests both on the choice of a statistical model, which can be checked against data, and on the assumption that the probability model stays the same across samplings, which the data cannot themselves establish. Chapter 2, Section 5.3 sets out Otsuka's wider reading of statistical inference as resting on such assumptions, and Chapter 3, Section 2 the further assumptions that Otsuka describes as connecting probability distributions with causal structure.

In measurement, the presupposition that cases are the same can be stipulated explicitly as a rule of method and yet remain exposed to the world. In Tal's study of coordinated universal time, described in Chapter 6, Section 5.2, the short-term stability of the time scale is imposed on the clock data rather than tested against them, as a rule for applying the concept of time to particular clocks, while the gaps that remain among independent realizations of the concept are discovered empirically (Tal 2012, 117, 132–134). Chapter 6, Section 5.2 described when redefining such a standard differs from merely moving a target.

The same presupposition can also be taken to be socially maintained. As Haely reports Longino's account, observation depends on accessibility to others and presupposed repeatability, whether or not anyone repeats the observation; in controversial cases, different perceivers test its observational standing (Haely 2003, 98–99). Applied where the community of perceivers is S, this account makes whether others can access and test an observation bear on its standing as evidence. That is not to say that whatever a community accepts is thereby true (Chapter 2, Section 6). Whether an experimental result comes to be accepted through negotiation within the community or through reasoned judgment is itself disputed (Chapter 6, Section 2.4), and the problem of shared presuppositions being shielded from criticism is taken up in Chapter 11, Section 3.

The second thread concerns the reliability of a route or source of evidence. Alston argues that sense perception taken as a whole cannot be shown reliable without circularity. A restricted practice, such as reading temperatures from mercury thermometers or recognizing vintages of wine by taste, can be checked against the facts through another mode of access. For sense perception in general there is no access to the facts that does not rely on perception, and an argument from the success in prediction and control that perception makes possible circles in a subtler form, since whether a prediction succeeded is itself found out by observation (Alston 1988, 436). Markus Lammenranta reports the fuller treatment in Alston's 1993 book as follows. The argument from perception's track record is not logically circular, since its conclusion is not among its premises, but it is epistemically circular: one cannot take oneself to be justified in accepting its premises without assuming its conclusion. On Alston's reliabilist account, which imposes no higher-level requirement, such an argument could even yield justification if perception is in fact reliable; yet, in a passage Lammenranta quotes, Alston observes that the same could be said of crystal ball gazing, so the argument cannot discriminate the sources that can reasonably be trusted from those that cannot (Lammenranta n.d., §1). Alston had made the same point in the 1988 paper: if a practice's outputs may be used to check it, every practice scores perfectly, and in this respect all are on a par (Alston 1988, 436–437). Lammenranta offers, as one possible diagnosis, that such arguments may provide justification but cannot rationally remove a doubt about reliability (Lammenranta n.d., §1). The argument from track record circles here partly because it takes an inductive form, so the two threads meet. On the statistical side, too, the error rates that justify a test's verdict depend on distributional and design assumptions that the verdict itself does not check (Chapter 2, Section 5.3).

Alston's own response is not a proof of reliability. As Lammenranta reports and quotes the 1993 book, it is a defense of the practical rationality of engaging in firmly established belief-forming practices and taking them to be reliable: there is "no appeal beyond the practices we find ourselves firmly committed to", practices we could give up or replace only with extreme difficulty, if at all; suspending all belief is not an option; and new practices put in their place would have no noncircular defense either (Lammenranta n.d., §7). In the 1988 paper Alston had set out how that standing can be lost or strengthened. A basic practice is to be accorded prima facie acceptance pending a demonstration of unreliability, which massive and persistent inconsistency in its outputs, or inconsistency with the outputs of more firmly established practices, could supply. A practice can strengthen its claim through a kind of self-support, such as the predictive success attained on the basis of perception, which is not noncircular, since the practice has to be used to ascertain the success, but is not trivial either, since a practice might fail to show it, as the reading of tea leaves does. Even a practice that passes these tests remains only prima facie acceptable, since trouble may develop later, and a practice that runs into trouble may be modified to get back in line (Alston 1988, 437). Lammenranta notes that Alston later rejected this practical argument and settled for a simpler form of Wittgensteinian contextualism, and objects that a defense of practical rationality does not contradict the skeptic, who denies not that it is practically rational to use our practices but that the beliefs they produce are epistemically justified (Lammenranta n.d., §7).

Wittgenstein likened such presuppositions to hinges, on which a door turns. Wittgenstein wrote that the questions we raise and our doubts depend on some propositions being exempt from doubt, as it were the hinges on which those turn; that this belongs to the logic of scientific investigation; and that it is not a matter of contenting ourselves with assumptions because we cannot investigate everything (Wittgenstein 1969, §§341–343). Applied to artificial systems, the last point comes to this: presuppositions do not remain for lack of time or capacity to investigate; they are conditions of checking, so greater capacity to check does not remove them. This does not mean that they stay fixed. Local presuppositions can be checked by other routes and revised (see the end of this section).

An artificial research system can have counterparts of these presuppositions. The correspondence is functional, not an identity. Nor is every feature of a system a presupposition in this sense: what counts is what bears on what a result is evidence of. In a system shaped by learning, to the presupposition that cases are the same there can correspond learned dispositions and representations that treat inputs as instances of stable kinds without any explicit representation that they are. The inductive biases of architecture and training, the distributional assumption under which performance on a held-out evaluation counts as evidence about further cases, the model under which an instrument or an automated evaluator counts as stable, and the condition under which a rerun counts as a repetition rather than as a second draw from the same organization, among others, belong to the same thread. Applied here, Haely's report suggests that rerunning one base system does not by itself show that a disputed observation has been tested by differently situated perceivers, a point to which Chapter 11, Section 3 returns. The presupposition that a route is reliable appears in the same form. One automated evaluator can be checked against another, and the system's evaluations against human judgments; whether this avoids circularity depends on how the two routes are related. But an argument from within the system for the reliability of its evaluation practice as a whole has at some point to rely on part of that practice. Where human judgment supplies the check, the same structure reappears one level up, for the wider practice that includes it.

Of whom these presuppositions must hold depends on S, as the introduction to Section 2 noted. How presuppositions are put in place and revised can also differ. A hinge in Wittgenstein's sense is a standing that a proposition has within a form of inquiry: one and the same proposition can be treated at one time as something to be tested by experience and at another as a rule of testing (Wittgenstein 1969, §98). The stability laid down as a rule of application in Tal's case is likewise a presupposition held as a rule that remains open to empirical revision. An inductive bias, by contrast, is a structure that the way a system is made puts in place and that remaking it can replace, and an artificial inquirer may also shape and revise its own presuppositions and rules. The difference bears on who can revise such a presupposition, a capacity that Chapter 7, Section 4.3 counts toward epistemic self-government, and on who answers for it, which Chapter 12 takes up.

Finally, in both threads presuppositions differ in scope. Some no check could validate all at once, such as the uniformity of nature or the reliability of perception taken as a whole. Others are local and can be checked by other routes when doubt arises: a perceptual mode by other senses, a statistical model against data, a standard against independent realizations, a claim of repeatability by other observers. Locality can make a presupposition available to checks that do not simply assume it, although narrower scope does not by itself guarantee better correction. What advances correction, as seen with Dewey and Chang in Chapter 6, Section 5.2, is running the circle in which standards are revised from within inquiry, repeatedly and against discrepancies among realizations; global presuppositions differ, since any check of them would presuppose them, and they cannot be rechecked in the same way. Of the counterparts named above for an artificial system, an automated evaluator's stability is local in this sense, whereas the reliability of the system's evaluation practice taken as a whole is global for that system, so that a claim about a configuration's checks can say which of its presuppositions lie on which side, and so which it can check for itself and which only remaking it, or checks from outside it, can reach. Why, then, may such presuppositions be held? Alston's response above and Reichenbach's argument that induction can be justified without knowing whether it succeeds (Chapter 2, Section 8.3) are attempts to answer. Alston's answer turns on belief-forming practices' being firmly established and ones we could not readily abandon. Alston does not discuss artificial systems, but applied to a system whose inductive biases its makers can replace by remaking it (above), the premise holds of the system's own practices only insofar as they are firmly established and the system cannot readily abandon them, and where they fail for the system's own practices, whether they hold would have to be asked separately of the wider practice that includes those who make and evaluate the system, so that who can revise such a presupposition (Chapter 7, Section 4.3) also bears on where Alston's ground lies. The history of formation that Section 2.2 discusses can explain why a presupposition need not be arbitrary with respect to the world, but it is not by itself a warrant. Dependence on presuppositions does not by itself show either that knowledge is impossible or that a check is reliable. Which evidential routes can bear on local presuppositions is taken up in Section 2.3. There is, however, a prior condition: a check can register an error about a target only if the states it checks are about that target, which is the question of Section 2.2.

2.2 Grounding is not one relation

Section 2.1 ended with the point that a check can register an error about a target only if the states it checks are about that target. This subsection takes up the question presupposed there: how do a system's states come to be about worldly targets at all? The question bears on this chapter because, on many accounts, a result is knowledge for a bearer only if the bearer's states are about the targets the result concerns. In research for a human society this can be met on the side of human concepts and instruments, but where an artificial bearer is claimed to know, it is asked of that bearer's states (introduction to Section 2). The subsection first takes up, from Section 2.1, the answer that looks to histories of formation for why primitive organization need not be arbitrary with respect to the world, and the limits of that answer as a warrant; it then notes that, on some accounts, history also fixes what a state is about, which leads to the arguments discussed under the name of grounding. The symbol grounding problem is posed by first setting symbols apart from the world, whereas the symbols actually in use are not set apart in that way. What follows presents the work of Taniguchi, who develops that point directly, and then accounts that locate the tie between symbols and the world elsewhere. Finally, it distinguishes the relations that the word runs together and confirms that success under one does not by itself establish the others.

One answer begins from history, which can explain why a primitive organization need not be arbitrary with respect to the world. Human bodies, perceptual systems, development, social learning, and cultural practices formed through long interaction with the same world that later inquiry addresses. Bradie and Harms give, as part of the difficulty of understanding cognitive behavior as a product of evolution, that at least three very different evolutionary processes are involved: the biological evolution of cognitive and perceptual mechanisms through genetic inheritance, the cultural evolution of language and concepts, and trial-and-error learning within an individual's lifetime. They add that there is some reason to agree with Donald T. Campbell, who coined the term evolutionary epistemology, that understanding human knowledge fully will require understanding the interaction between these processes; this, they say, requires being able to model both biological and cultural evolution, and they go on to models of cultural evolution (Bradie and Harms 2020, §§1, 2.2).

On the accounts discussed here, the historical answer is plausible only in a graded form that holds capacity by capacity and domain by domain, and history cannot warrant everything it helps explain. Versions of it belong to evolutionary epistemology's program on how cognitive mechanisms are shaped by environmental structure (Bradie and Harms 2020, §1.1). Boudry and Vlerick reject the sweeping claim that natural selection is simply indifferent to truth. Their target is Alvin Plantinga's evolutionary argument against naturalism. As they reconstruct it, the argument holds that selection cares only about fitness and not about truth, so that if naturalism, the view that there are no gods or supernatural beings, and unguided evolution are both true, we have no reason to trust our cognitive faculties, and the naturalist acquires a reason to give up every belief, including the belief in naturalism itself (Boudry and Vlerick 2014, 65–66). They answer with three arguments. First, although survival does not require beliefs at all, a lineage that has once come to direct its actions through beliefs remains bound to that route, a constraint they call adaptive path dependency; from then on, beliefs that are at least approximately true in ecologically relevant situations are favoured, and false ones are harmful (Boudry and Vlerick 2014, 68–69). Second, they argue that what a belief is about is tied to the actions it disposes an organism to take, given its desires and other beliefs, so that its content cannot change while its effect on behaviour stays the same; whether that behaviour is adaptive then depends on whether the world is as the belief represents it, as when a belief that a mushroom is safe to eat proves harmful because the mushroom is poisonous (Boudry and Vlerick 2014, 69–72). Third, they argue that scientific belief is not a product of intuitive faculties working unaided: it rests on cooperation among contemporaries and across generations, on mathematics and notation, on instruments, and on checking theories against observation, and the inference that everything built on imperfect foundations must be unreliable is what they call the foundationalist fallacy (Boudry and Vlerick 2014, 72–74).

They also say where selection need not favour truth. Heuristics, quick rules of thumb, may trade accuracy for speed; because different kinds of mistake carry unequal costs, a mechanism may be inclined to make certain mistakes; and capacities that are accurate in familiar environments may mislead in new ones, so that selection may favour true belief in some domains but not in others (Boudry and Vlerick 2014, 66–67). Vlerick and Broadbent, who accept that defence against evolutionary skepticism, explicitly deny that it supplies a general justification of scientific belief: on their account, selection shaped cognitive faculties to track truth only in a narrow domain of ecologically relevant beliefs, and modern scientific theories are not an extension of those beliefs (Vlerick and Broadbent 2015, 186, 190–192).

To sum up, interaction with the world can help explain why a capacity responds to real regularities rather than being freely invented. It does not establish, for every resulting expectation, that it is true, generally reliable, justified, or suited to environments beyond the range in which it formed. As Boudry and Vlerick allow, adaptive shortcuts and systematic errors are possible, and advanced inquiry depends on more than the faculties that this interaction formed. On these accounts, then, natural history is a possible genealogy of epistemic capacities, telling where they came from, not a warrant for all their products. Applied to an argument from natural history, Alston's analysis gives a further reason why such an argument cannot supply a general warrant. The argument that natural selection would have preserved only belief-forming dispositions that mostly yield truths relies on evolutionary theory, which rests on evidence gathered through perception. By the criterion Alston applies to the argument from success in prediction and control, whose success is itself established by observation (Section 2.1; Alston 1988, 436), it would therefore be epistemically circular as an argument for the reliability of perception, whatever the merits of its premises.

Does an artificial system, then, have a corresponding history? Abstracted, what the discussion so far says is that an inquirer's primitive organization is shaped by that inquirer's history of formation. Artificial systems have histories of formation too. Whether those histories include anything corresponding to the three processes is not obvious, however, and the situation differs from process to process. The analogy easiest to trace is with trial-and-error learning: in a system that learns from the outcomes of its own actions, something close to it can be at work.

Whether anything corresponds to biological evolution is far less clear. The program in evolutionary epistemology that addresses the evolution of cognitive mechanisms takes as its object the biological substrates of cognition, such as brains, sensory systems, and motor systems (Bradie and Harms 2020, §1.1). That process has not merely tuned the working of organs already in place; it has produced and remade sense organs and nervous systems themselves, and with them what can be sensed at all and what distinctions are possible at all. The expectations built into perception and action discussed in Section 2.1 are, on this view, shaped by such a process. In many present trained systems, by contrast, architectures and sensing arrangements (sensors and input formats) are designed by people, and learning takes place on top of them. There may be room to regard design choices and the structure that training builds in as taking on part of this role. But to treat them as what corresponds to biological evolution risks reducing a process in which organs themselves came into being and were remade to the adjustment of a given structure. What in an artificial system could play the role of this process, and whether anything could, is not obvious.

Correspondence with cultural evolution is not simple either. Bradie and Harms treat cultural evolution, through population models, as a process in which languages and concepts are transmitted or rejected within a population, with fitness components that include transmission and rejection probabilities, and they count the construction of cultural niches such as language and educational institutions as part of it (Bradie and Harms 2020, §2.2). If so, the question is not whether a single system undergoes it but in what population a system takes part in the transmission and sorting of concepts, and the answer can vary with the society in which the system stands. Training on human records inherits products of human cultural evolution, but what is inherited is what records retain, which may not include such things as the collective tacit knowledge acquired by taking part in a practice (Chapter 3, Section 6). Conversely, where a system's outputs enter human language and records and become part of what later people and systems inherit, the system does not only inherit but also takes part in human cultural evolution. Where vocabularies are transmitted and stabilized among artificial agents, something like a cultural evolution of an artificial society itself is conceivable. The question then is what sorts the vocabularies, which the experiment of Bouchacourt and Baroni discussed below makes concrete.

Where a system inherits from human records, its organization is tied to the world not through its own interaction with the world but as mediated by human history. What such mediated connection can give its states bears on the account of Coelho Mollo and Millière below and on the route of inheritance from records discussed in Section 2.3. The limits noted above can also be applied to the histories of artificial systems. Even if a training history explains why a capacity responds to regularities in records, it does not establish that the capacity suits environments beyond the range of training, or that its products are warranted. Much of what has been said here concerns present trained systems. If there could be artificial lineages that remake their own designs and sensing arrangements through the outcomes of interaction with the world, matters could differ.

History bears on the question of this subsection in another way. Boudry and Vlerick's second argument above, by tying a belief's content to the actions it disposes and so letting selection on action reach content, already connects history with content. The account of Coelho Mollo and Millière, discussed below, draws on teleosemantics, a theory of representation that explains what a state represents by the function it has acquired through a history of selection. On their account, a history of selection, which learning can supply, is part of what gives internal states the function of carrying information about worldly targets (Coelho Mollo and Millière 2026, 13–17), while grounded states do not by themselves suffice for knowledge (Coelho Mollo and Millière 2026, 2, 24). On that account, then, a history can fix what a state is about without warranting what the system takes to be true of it. Whether a history of formation also bears on warrant is a further question, on which accounts of justification differ.

Different questions have been raised under the name of grounding. Harnad's symbol-grounding problem asks how the semantic interpretation of formal symbols can be made intrinsic to the system. The meanings of the symbols in a book derive from the meanings in their readers' heads; the question is how to give a system meanings of its own rather than such derived ones (Harnad 1990). Bender and Koller argue that a system trained only on linguistic form cannot learn meaning, understood as the relation between form and communicative intent. On their account, communicative intents are about something outside language, and text is their projection into forms; if the training data are only form, there is not sufficient signal to learn that relation, and without a means of hypothesizing and testing the intents, reconstructing them from forms alone is hopeless (Bender and Koller 2020, 5185, 5187, 5189). These arguments pose hard questions for systems trained only to predict text, but do not by themselves settle questions of belief, knowledge, or research.

Both arguments carry limits that their authors state. Harnad held that, strictly speaking, symbol grounding is a problem for modeling cognition rather than for AI in general: if symbol systems alone produced the performance AI seeks, there would be no reason to demand that their symbols have intrinsic meaning, although Harnad allowed that grounding might bring performance gains in robotics and machine vision (Harnad 1990, n. 10). Bender and Koller confine their argument to systems trained only on form. They state that it does not apply to systems trained on tasks such as reading comprehension, whose data specify semantic relations between texts, and that meaning could conceivably be learned from form augmented with grounding data, to the extent that the data represent communicative intent (Bender and Koller 2020, 5191–5192). Set side by side, the two arguments suggest that whether an artificial researcher needs to be grounded in Harnad's sense (an interpretation intrinsic to the system) or in Bender and Koller's (the relation between form and communicative intent) depends on what grounding is invoked to explain, since, as Chapter 1, Section 1 notes, candidate definitions can answer different questions. It also depends on whether the system itself is claimed to know, or only supplies results that others come to know (introduction to Section 2).

These arguments are framed by first setting symbols apart from the world and then asking how they come to have meaning. Harnad's question was posed about systems, like those of symbolic AI (classical AI that manipulates symbols by rules), in which symbols are prepared in advance as formal tokens and their interpretation is supplied by an interpreter outside the system. Bender and Koller's argument likewise separates linguistic form from the intent and the world it expresses, and asks what can be learned from form alone. The symbols we actually use, however, do not stand apart from our dealings with the world and with others. On the developmental research that Saji reviews, words are learned in exchanges with others while attending together to things and events (Saji 2026, 46–47); they change their meanings through use, and in turn orient our perception and action (Taniguchi 2026b, 7–8). If symbols and the world are separated at the outset, the question arises how to connect them. If they are seen as shaped together in interaction, the question becomes how that connection is formed and maintained. Harnad too, rather than leaving symbols set apart, sought the solution in grounding them bottom-up in nonsymbolic representations based on sensory input, iconic and categorical representations (Harnad 1990). It is this view, on which symbols and the world are shaped together in interaction, that Taniguchi develops head-on.

Taniguchi, too, takes the symbol grounding problem to have been posed within symbolic AI, about "symbols" understood as formal tokens (Taniguchi 2026c, 42–43). Taniguchi reports having come to believe that the very premise of the question how artificially created symbols can be meaningfully connected to real-world experience may be misguided, since symbols are not merely external representations but are deeply intertwined with internal representations, such as concepts, that develop through self-organizing processes grounded in sensorimotor interaction with the environment (Taniguchi 2026a, v–vi). From the perspective of symbol emergence in robotics, the central problem for AI is not symbol grounding but symbol emergence: how meaningful symbols arise, change, and become organized through social and embodied interaction (Taniguchi 2026c, 43). The recasting does not set Harnad's point aside: Taniguchi accepts Harnad's observation that relations among words yield at best relational concepts, not grounded ones (Taniguchi 2026d, 198).

In what Taniguchi calls the theory of symbol emergence systems, each member forms internal representations through bodily interaction with the environment, without direct access to the inner states of others. Through semiotic exchange among members, symbol systems such as language become organized from the bottom up, and the organized system in turn constrains members' perception, action, and interpretation from the top down. Taniguchi holds that this micro–macro loop gives rise to symbolic meaning (Taniguchi 2026b, 4–8). Internal representations are not yet symbols but their raw material, and the meaning of words is not fixed within the individual alone but obtained through interpretation by others within a social context (Taniguchi 2026b, 7). Grounding for each member, too, is said to lie not in the objective world but in the Umwelt in von Uexküll's sense, the subjective world in which a creature perceives and acts (Taniguchi 2026d, 198–199). Membership requires not biological humanness but the ability to take part in semiotic communication, using arbitrary symbols (symbols with no natural tie to their objects) to contribute to societal functions (Taniguchi 2026b, 6). Taniguchi holds that an AI system can be said to participate meaningfully in a symbol emergence system only when it has a body and interacts with the real world (Taniguchi 2026c, 42), while also saying that generative AI is already changing human language use and becoming part of such a system (Taniguchi 2026a, v; Taniguchi 2026g, 222, 225).

For how the loop works, Taniguchi and colleagues offer a model of a naming game. Assuming that two agents attend to the same object, a speaker proposes a name on the basis of its own observation, and a listener accepts it probabilistically in light of the internal representation formed from its own observation. When acceptance is made to follow a prescribed rule, the Metropolis–Hastings algorithm, the exchange is formally equivalent to approximate Bayesian inference over a shared latent variable that integrates the observations of both; that is, the observations of both can be reflected in the formation of shared symbols without either referring directly to the other's internal representations (Taniguchi 2026e, 207–208). From this Taniguchi proposes the collective predictive coding hypothesis, that human language is formed through such distributed inference (Taniguchi 2026f, 211–212). Applied to the relations distinguished below, the framework can be read as treating dependence on perception and action (the sensorimotor relation) and the establishing of common ground between speakers (the communicative relation) as two sides of one loop, and as relocating reference, what symbols or states pick out in the world, from the objective world to each member's Umwelt and to an Umwelt shared among members (Taniguchi 2026f, 212).

The recasting can be stated in the distinctions of this book. If the meaning of symbols arises within such a loop, what is asked of an artificial system is not only by what relations its states come to be about worldly targets but also which symbol emergence system it joins, and as what kind of member. The question raised above about cultural evolution becomes, on this view, a question about meaning itself. Here the agent who conducts research, A, the society S that sets what is new and for whom an achievement is sought, and the community that A joins and that shapes a symbol system need to be kept apart. That community may be a human, a mixed, or an artificial society, and it need not coincide with S; because this book distinguishes A from S, this difference can be stated once the community is placed as a further term, and it is one source of the mismatch that Section 3 examines. Takahashi, proposing in a preprint levels of autonomy for scientific AI, reaches such a difference from the side of hypothesis generation. On this account, generating hypotheses whose form is not fixed in advance, Takahashi's fourth level, requires in experimental science that the symbol grounding problem be solved at least in the domain the experiments deal with, which means that the system acquires an embodiment through particular laboratory robots. Since that would give rise to a symbol system separate from the one built on human embodiment, the account holds that from this level the system would understand its objects in a way different from the way humans do, and that this is where a divergence between human science and AI science begins (Takahashi 2024, 7, 10). At the fifth level, where humans give the system its research purpose, the system would by contrast need a symbol system compatible with the human network of meaning, so that some means of continuing to ground its knowledge in the human system of scientific knowledge, or to translate it into that system, would have to be built and maintained (Takahashi 2024, 8, 11). Read in this book's terms, Takahashi's fourth level lets the symbol system with which A works diverge from that of a human S, while the fifth asks that the research remain research for that S. Even on this view, however, whether success under one of the two relations joined in a single loop, the sensorimotor and the communicative, establishes the other can vary with how the loop works. Whether shared symbols come to be about worldly targets, and whether they remain open to correction by those targets (Section 2.3), are still to be asked of the workings of the loop. An example appears below when the experiment of Bouchacourt and Baroni is set beside this model.

Other accounts locate the tie between symbols and the world elsewhere. Coelho Mollo and Millière distinguish referential, sensorimotor, relational, communicative, and epistemic grounding, and treat referential grounding as the one essential to the problem. Drawing on teleosemantics, they argue that it requires two conditions: causal-informational relations to the world, which for a language model pass through human linguistic production, and a history of selection, which learning can supply, that gives internal states the function of carrying that information. They argue that language models can meet both conditions, particularly those that undergo preference tuning, and that multimodality and embodiment are neither necessary nor sufficient for reference (Coelho Mollo and Millière 2026, 1, 13–17, 22, 24). They add that the designers' intentions give a function to the learning mechanism, whereas the function of internal states, and the standard of their correctness, are given by the history of selection that the mechanism carries out on its data; they compare this to how a dog breeder's intention to produce calmer dogs sets up a selection regime, while it is the history of differential reproduction that confers the function on the selected traits (Coelho Mollo and Millière 2026, 16). The separation runs parallel to the one Chapter 7, Section 2 drew for the attribution of conduct, between where a function comes from and where the organization of an inquiry lies; on this account, what designers intend for the mechanism does not by itself fix what its states are about.

Coelho Mollo and Millière also mark the limits of their claim. They do not argue that language models understand, know, or have any conception of truth, and they hold that grounded states would not by themselves suffice for such capacities. They set communicative intent aside, relating Bender and Koller's challenge to communicative grounding, while rejecting, against Bender and Koller, the view that linguistic form can be easily separated from meaning (Coelho Mollo and Millière 2026, 2, 5 n. 9, 8, 13, 24). Read side by side, the two positions therefore conflict over what training a system only to predict text can supply, but not over whether such a system has communicative intentions. Bender and Koller count such training as training on form alone and deny that it can yield conventional meaning, which on their account connects expressions to objects outside language, or the relation between form and communicative intent (Bender and Koller 2020, 5185, 5187); Coelho Mollo and Millière's rejection of an easy separation of form from meaning calls into question that classification of text prediction as training on form alone. Where training adds signals beyond the prediction of text, as in the preference tuning that Coelho Mollo and Millière single out, Bender and Koller's 2020 paper does not by itself settle whether their argument still applies.

Enactive approaches that root cognition in a system's self-maintenance ask for something stronger than referential grounding. Di Paolo argues that autopoiesis, a system's self-production, yields by itself only an all-or-nothing norm, to go on living or to disintegrate, which the system itself cannot register; sense-making requires in addition adaptivity, a capacity to monitor and regulate its own states and its relation to the environment according to whether they are approaching the boundary of its viability (Di Paolo 2005, 6–10). Di Paolo does not settle the artificial case in one direction. Meaningful action, the paper holds, can be convincingly emulated by an artificial system without the system acting meaningfully, so whether a system acts meaningfully cannot be told from observing its responses; it has to be checked by examining whether the system's organization generates its own sense-making in relation to its viability, a check regarded there as relatively simple at least for ruling out negative cases. Di Paolo leaves open whether artificial intentionality, in the paper's sense the intentional structure of an agent's own behaviour, is possible short of a fully metabolizing robot, and notes that the picture changes once values not originating in metabolism are considered. In closing remarks that the paper calls speculative, Di Paolo argues that self-sustaining forms of activity, such as organized patterns of sensorimotor coordination, can generate values of their own within the bounds of metabolic viability (Di Paolo 2005, 13–16). Rolla, by contrast, argues that AI systems lack a living body of their own and do not inhabit an environment, and so cannot autonomously exchange matter and energy to remain within viability limits; a genuinely cognitive AI, on this account, would first require artificial life (Rolla 2026, 1, 11, 14). On Rolla's account it is through autonomy of this kind that symbols matter intrinsically: the symbols we use are grounded because they matter to us, as products of our autonomous behavior (Rolla 2026, 6, 11–12, 14). This enactive requirement is a substantive rival to Coelho Mollo and Millière's account, not a theory-neutral condition of artificial research. The autonomy at issue in these accounts, that of a system whose own activity produces and maintains the system itself, is not the comparative autonomy of Chapter 7, Section 4, which compares, in a specified respect, how far parties outside a selected boundary direct, substitute for, authorize, ratify, or repair the organization attributed to an actor; a verdict about one does not settle the other.

Where the argument depends on any of the senses of grounding, six items are kept apart. The first five are relations between expressions or states and something else; the sixth is a condition that may be required in addition:

1. referential grounding: reference to worldly targets;
2. sensorimotor grounding: dependence on perception and action;
3. relational grounding: relations of an expression to other expressions within a language;
4. communicative grounding: the establishing of common ground between speakers, including a shared way of connecting expressions to the world;
5. epistemic grounding: connection of expressions to information stored in a knowledge base;
6. enactive or intrinsic normativity grounded in self-maintenance and its adaptive regulation.

The first five are the notions of grounding that Coelho Mollo and Millière find in the work of cognitive scientists, philosophers, and AI researchers, and the sixth adds the enactive requirement described above (Coelho Mollo and Millière 2026, 6–8). The purpose of listing these six is to keep apart relations and conditions that the word runs together, not to define grounding. The list leaves out, for example, the phenomenological requirements mentioned in Section 1.3. Coelho Mollo and Millière treat referential grounding as the most fundamental of their five and as arguably the basis of the others, with a partial exception for relational grounding (Coelho Mollo and Millière 2026, 24). The enactive accounts above, which locate the conditions under which meaning arises for a system itself in self-maintenance and its adaptive regulation, stand, in this list, as asking whether referential grounding alone can explain sense-making. Their fifth notion is close to one use of the word in engineering contexts for artificial systems, connection to retrieved information: on their account, current language models may involve epistemic grounding when they are connected to external databases from which they retrieve information (Coelho Mollo and Millière 2026, 8). Answerability to evidence and correction is a relation of a different kind from all six. A state can be about a target and misrepresent it, and whether a system's commitments are corrected when they misrepresent is a further question, which Section 2.3 takes up as world-answerability.

Success under one relation does not by itself establish the others. This is not a claim that the relations are independent of one another; which relation supports which differs from account to account. On Coelho Mollo and Millière's account, a linguistic system may refer to worldly targets through causal relations mediated by human text, together with a learning history that selects its states for carrying that information, without possessing its own sensorimotor skills. A robot may couple perception and action without forming beliefs or participating in scientific criticism, which is why the question of Section 2.3 remains open even where sensorimotor grounding is secured. A community may stabilize a vocabulary that supports successful coordination among its members, while that success does not settle what the vocabulary is about. Experiments on emergent communication between artificial agents give this last possibility a concrete form. Bouchacourt and Baroni trained pairs of networks on a game in which one agent signals which of two images is the target and the other must pick it out. Pairs that succeeded developed representations almost perfectly aligned with each other but only weakly correlated with the representations of the input images, and they communicated nearly as successfully about pairs of random-noise inputs as about images; the authors conclude that what the signals tracked was not the depicted kinds but low-level relational properties of the inputs (Bouchacourt and Baroni 2018, 983–985). The signals did discriminate the targets of the game. Successful coordination within such a community is therefore not, by itself, evidence that its shared signals are about what an observer would take them to name. In the naming game of Taniguchi and colleagues, by contrast, what sorts the vocabulary is not success in coordination but likelihood under the internal representation that the listener has formed from its own observation. Taniguchi distinguishes this game from conventional naming and referential games by its assumption of joint attention (Taniguchi 2026e, 205, 207). The formal equivalence described above does not, however, show that the shared symbols track what an observer would name. Taken together, the two suggest that what a shared vocabulary tracks is not settled by its stabilization alone and can vary with what sorts the vocabulary. In an artificial community, what sorts it can be set by how the exchange among its members is made. Not tracking the kinds that a human observer assumes does not by itself show a lack of connection to the world, and whether commitments can be corrected by errors about their targets remains the question of Section 2.3.

2.3 World-answerability

Section 2.2 ended with the question whether commitments can be corrected by errors about their targets, and Section 2.1 noted that local presuppositions can be exposed to corrective evidence beyond the assumptions under examination. How can an epistemic organization's commitments remain open to correction by their targets, and who would sustain that relation where S is a human, a mixed, or an artificial society (Sections 1.1–1.3)? As used here, world-answerability means two things. One is that an epistemic organization is differentially constrained by target phenomena: its commitments are constrained by them in such a way that they would differ if the phenomena were different. The other is that the organization remains corrigible through further contact, intervention, and cross-checking by independent routes. Being answerable to the world in this sense is not having access to an oracle. The idea has appeared earlier in the book: as answerability to proofs, observations, interventions, failed predictions, and criticism without an infallible test for final truth (Chapter 2, Section 3), and as revision of standards that remains answerable to discrepancies among realizations assessed on grounds beyond the revision itself (Chapter 6, Section 5.2). Chapter 11, Section 1 specifies such a route as a defeat relation and distinguishes the evidence it adds from the arrangements that protect it, including how diagnoses of failure remain open to challenge. Where the book speaks of answerability to goals or to people, the sense is different.

The phrase "answerability to the world" is not used here in the sense McDowell gives to thinking that is answerable to the world. For McDowell, that answerability is a rational relation, in which a judgement is correctly or incorrectly made according to how things are, and it requires experience in which the subject's conceptual capacities are already at work to stand in rational and not merely causal relations to what the subject thinks, relations in which experience serves as a reason for or against it, a requirement McDowell takes to hold for there to be empirical content at all (McDowell 2000, 3–5, 9–12, 16). McDowell's account is not about artificial organizations, but taken to the definition used here, it implies that an organization meeting that definition through causal and practical routes alone would not thereby be answerable to the world in the account's sense, and that whether it could be would depend on whether its states draw on capacities of the kind McDowell calls conceptual. The two usages can therefore divide their verdicts on the same organization: an artificial organization whose commitments are revised by the results of interventions can be answerable to the world on this chapter's definition, while on McDowell's account that alone leaves undecided whether what is being corrected is thinking in the account's sense. The verdicts part not over whether there is a route of correction but over what it is that is corrected. The accounts of grounding in Section 2.2 address what states are about but not the conditions of capacities of that kind.

Feedback from nature is accordingly a modest notion here. Nature does not return an uninterpreted verdict. Interventions produce traces through apparatus, sampling, classification, models, and standards, and a mismatch may implicate the focal claim, an auxiliary commitment, an instrument, a material condition, or the representation through which the result is described (Chapter 6, Section 2.1). Whether the standards through which such traces are read make answerability itself relative to S depends on the account. On accounts on which what corrects a commitment is fixed by its target independently of S's epistemic norms, those norms can bear on what counts as evidence (Section 2.1), but world-answerability is not relative to S. Leibo and colleagues, in a preprint position paper on AI alignment, hold instead that even propositions one would ordinarily be tempted to call facts are governed by epistemic norms that control justificatory standards and appropriate evidence integration, norms "just as culturally contingent as any other kind of norm" (Leibo et al. 2025, 2, 7). Brought to the analysis of answerability to the world, which they do not discuss, their view implies that what counts as being corrected by a target is itself partly fixed by the norms of a community. It does not follow that the difference between answerability to the world and relevance to S, on which Section 3 relies, disappears, since that standards of evidence are social and that what a society wants to know is relative to it are different matters: societies with the same evidential norms can differ over what is worth pursuing, and societies that count the same problem as important can adopt different norms of testing. What weakens on this view is not that there is a difference but the assumption that what counts as a correction can be fixed independently of a community's norms; to that extent, as Section 3.2 notes, the two come apart less sharply.

For research aimed at human knowledge, primitive world relations need not all be inside the artificial component. Human concepts, testimony, datasets, instruments, evaluators, and institutions may provide mediated links from the activity to the world. For a hybrid organization, the relation may belong to the organized whole: one component senses, another intervenes, another preserves records, and another evaluates consequences. For a mixed society, the relation would have to hold of its artificial members insofar as they are to know (introduction to Section 2); on Longino's account as Haely reports it, whether other members can access and test observations can also bear on their standing as evidence for the society (Section 2.1).

Termine, Ratti, and Facchini's analysis of machine learning in science is an example of how what a check presupposes can itself be distributed in such an organization. They argue that the construction of machine-learning models in science is mostly theory-indifferent: setting a model's parameters and hyperparameters, and increasingly feature engineering as well, does not necessarily require reference to the domain theory of the field in which the model is used, at least for predictive accuracy on data distributed like the training data (Termine, Ratti, and Facchini 2026, 15, 17–20, 22–25). Domain theory nevertheless enters the training data, since curating them is theory-informed, and the trained model passively inherits these priors, which end up scattered across parameters that have no interpretation in the domain's terms; this weak form of theory-ladenness, arising outside the modelers' intentions and control, they call theory-infection (Termine, Ratti, and Facchini 2026, 15–16, 20–22). Curation, they note, is commonly done by data specialists separate from the model builders, and using the models in science, as most of the literature holds, typically relies on domain theory (Termine, Ratti, and Facchini 2026, 15, 21, 29). Read in this way, theory dependence is distributed across the steps of an inquiry, and across those who perform them, rather than attached uniformly to every one: construction can be largely indifferent to domain theory while theory enters through, for example, the curation of data, the goals that the scientific context sets for the model, and how the model is interpreted and used. Domain-indifferent construction does not make the larger inquiry theory-free, and theory-laden use does not show that the model builder supplied the relevant scientific organization.

For a candidate artificial knower or society, the attribution burden is stronger. On accounts that attribute knowledge to the artificial subject itself, an account is needed, supplied by the subject itself or by external analysis as the adopted epistemology permits, of how worldly differences constrain its own commitments rather than only human judgments downstream. Possible routes, distinguished by what carries the contact, include direct perception and intervention by the system itself; causal-historical inheritance from records of others' contact with the world; developmental learning from its own failed actions; and distributed membership in laboratories whose other members and instruments make contact. Enactive theories of the kind described in Section 2.2 add a further condition that is not a route of contact, an artificial form of self-maintaining and adaptively regulated normativity. These routes are possible ingredients within different accounts, separately or in combination, and naming them does not establish that any of them is sufficient for an artificial knower. Nothing in the present argument selects one route universally or requires a humanoid body.

Accounts of learning from models help separate a model's internal coherence from its answerability to the world. Morrison and Morgan argue that manipulating a model teaches, in the first instance, about the world in the model, and that what is learned there can be transferred to a theory or to the world the model represents. Irving Fisher's hydraulic model of the monetary system, for example, demonstrated results about bimetallism within the model's world but, as they note, could not support quantitative prediction or intervention in the monetary system (Morrison and Morgan 1999, 33–35). Extended to artificial systems, these accounts imply that a system that learns by manipulating its own internal models learns first about those models, and that whether that learning answers to its targets turns on routes, such as measurement, that do not run through the model's own coherence.

Whether a particular pattern that a machine detects is constrained by its target, in the sense of the definition above, is a further question. Buckner, in a paper cited here in its preprint version, argues that predictively useful features that humans find inscrutable may be "real but available only to 'alien' perception and cognition," or may be artifacts of the system's processing: like lens flares in a photograph, or a migraine aura reliably induced by the light in a particular forest, artifacts can carry predictively useful information without being observer-independent features of what they are taken to be about (Buckner 2020, 1, 12–13, 15). Buckner adds that what counts as an artifact depends on the function of a feature in an investigation, and proposes as tests an account of which design choices produce which features, triangulation across differently built methods, and successful causal intervention (Buckner 2020, 16, 18–21). Triangulation asks for methods built differently, which instances of one base system are not (Section 2.1; Chapter 11, Section 3). On this account, then, a pattern's predictive usefulness does not by itself settle whether it answers to the target; what would settle it are routes of the kind named in the definition above, further contact, intervention, and cross-checking by independent routes.

Humphreys raised the same two questions for deep networks: whether their internal representations refer to aspects of reality to which humans are not accustomed, and perhaps cannot interpret, or whether parts of them are nonreferring artifacts of the computation. Humphreys noted that the parallel with the debate over scientific realism is imperfect: there, the instruments that told scientists whether theoretical representations refer were built by humans and so were to some extent understood, whereas the details of a network's internal representations are constructed by the computer, which often makes inferences about them profoundly difficult (Humphreys 2020, 22). As a primary constraint on identifying such patterns, the same paper proposed that humans be able to identify them as corresponding to a feature that humans can recognize as nonarbitrary, which on this account is why mapping a pattern onto a space familiar to humans is effective in distinguishing real features from arbitrary ones (Humphreys 2020, 23). The two accounts draw the line between a real pattern and an artifact differently. On Humphreys's, identifying a pattern as real runs through human recognition; Buckner's tests ask what produced a feature, whether it recurs across independently built methods, and whether intervening on it brings about changes in the target, and they are intended for features that humans find inscrutable (Buckner 2020, 21–22). Humphreys's constraint concerns how humans identify such patterns. Applied to a pattern that only an artificial organization could recognize, it implies that such a pattern could not be identified as real by this route, whereas on Buckner's account it could still be shown to answer to its target.

The analysis of world-answerability thus yields a possibility rather than a demonstration: an artificial body, sensorium, developmental history, or community could support an epistemic organization different from ours while remaining constrained by nature. From the possibility it does not follow that current AI systems possess such an organization, that internal coherence is enough, or that humans could understand the resulting concepts. Chapter 11, Sections 1 to 3 take up the first two of these matters as questions of criticism and defeat; the third leads to the relations between differently organized epistemic communities that Sections 3.1 and 3.4 examine. For the first of the matters with which this chapter opened, then, a claim that an artificial configuration itself knows a result carries burdens at three places: what its checks presuppose, what its states are about, and how its commitments answer to their targets. Where the claim is instead that a result has entered a human S's epistemic situation, these burdens may be met on the human side (introduction to Section 2). On the accounts discussed, success at one of them, or predictive usefulness, does not by itself meet the others.

3. Nonhuman research and epistemic alignment

What follows when an artificial epistemic organization differs from human ones? On views that ask artificial research for more than performing familiar procedures faster, the possibility reached at the end of Section 2.3 is one reason to build an artificial researcher. Different bodies, sensors, memories, temporal scales, representations, and collective organizations could make different patterns available. For example, an artificial system might take in information far faster than humans and work as millions of parallel instances (Amodei 2024, "Basic assumptions and framework"), so that patterns visible only across far more cases than any human can survey become available to it; read and integrate the whole of a literature that no human has time to read (Winn 2025); handle systems involving relations among far more variables than the few that humans can grasp at once (Takahashi 2019); or, learning from its own experience of the world, find principles not bound by present modes of human thought (Silver and Sutton 2025, 5). An artificial inquirer might divide objects and processes differently, attend to relations that humans have not attended to, or sustain investigations beyond human lifetimes and cognitive limits. On those views, requiring every such result to reproduce human salience, what humans notice as standing out or mattering, and human intuition would remove part of the intellectual possibility in advance.

Otsuka's reading of deep learning is an example of such a difference in current systems. Otsuka argues that deep-learning models appear to carve their inputs into kinds of their own through representation learning. On this reading, success in transfer learning, the application of a model trained in one domain to problems in another, suggests that these kinds follow real divisions to some extent, since supporting inductive extrapolation across domains is what kinds that track real divisions do; adversarial examples suggest that the carving is imperfect or that the features the models use may differ sharply from ours. Otsuka also gives a reason to expect difference: a pattern counts as real insofar as it helps predict unobserved cases, and which patterns do so depends on how much data a cognizer has, so there is little reason to expect the patterns found by models trained on enormous amounts of data to coincide exactly with those we take to be real (Otsuka 2023, 120–124, 138–141, 175–176). Whether such a model's internal representations have a uniquely correct rendering in human concepts is the question Chapter 3, Section 5.4 took up.

The same nonhuman difference creates several risks. An artificial community could become internally coherent while poorly constrained by its targets, for instance where what it learns comes only from manipulating its own models (Section 2.3). Its shared signals could support successful coordination while tracking something other than what human readers take them to name, as the emergent signals of Section 2.2 illustrate in a simple case. It could discover reliable relations that humans cannot interpret, criticize, or use. It could pursue unknowns valuable to itself but irrelevant or harmful to humans. It could inherit human language and signals of what human evaluators approved (Section 3.3) while remaining insensitive to material defeaters, facts or evidence that should overturn a claim. Nonhuman difference is therefore neither evidence of superiority nor evidence of error.

What these possibilities and risks have in common can be put as a mismatch: an artificial inquirer and S may not coincide in such things as what they take to be unknown, worth pursuing, or a good reason, what broader values they seek from science, and whether what one reaches can be understood by the other. A mismatch is not by itself a defect or an advantage. Whether a given mismatch is drift away from S, an error about the world, or a gain for S can be said only once the standpoint from which it is assessed is named: S as it now is, S as it may come to be, or, where correctness is in question, the world.

Section 3.1 asks where the difference lies when a result or a science is called alien. Section 3.2 asks where mismatch comes from and relates the worry this raises to what is known as the AI alignment problem. Section 3.3 asks whether research by an inquirer of another kind really fails to align with S, and what follows if it does. Section 3.4 asks how, where such research does not align with S, S could still gain knowledge or understanding through what the inquirer reaches, and what would have to pass between differently organized communities for that.

3.1 Decomposing alien science

When a result or a science is called alien, what is alien about it? Kureha and Kukita pair the possibility of an alien science that differs from human science with the risk of an alienation of science, in which reliance on opaque AI could cost science its aim of understanding phenomena (Kureha and Kukita 2019). Kureha later develops the first possibility, describing AI systems that might bring forth a system of knowledge that understands the world differently from us (Kureha 2023, 48–49). In a 2024 talk Kureha restates the second as a situation in which science loses its original aims or values, so that ideals such as the view that understanding the world is the aim of science may change. Kureha takes AlphaFold 2 to have made the success of a science without understanding a reality, notes that the effect of efforts to interpret AI outputs so that humans can understand them is still unknown, and doubts that scientists themselves have given up on understanding (Kureha 2024, slides 3, 12). The framing makes visible both the expansion of possible inquiry and possible changes on the human side, which can concern human access to and participation in research as well as what science is taken to aim at.

The distinctions of Section 1 allow the issue to be decomposed by where artificial attaches. Attached to the actor, it says that an epistemic output was generated by an artificial process. Attached to the representational form, it says that the output's representation or conceptual form is machine-native and not readily translated into the concepts of S. Attached to the bearer, it says that the knowledge is literally possessed by an artificial bearer. Attached to the reference society, it says that the knowledge was sought for an artificial society; whether that society also governs its scientific practice is a further claim. Two of these places, the actor and the reference society, are the parties that Section 1 distinguished, and the bearer is what Section 1 identified as the subject in which the achievement is realized. Representational form is none of these; it is added because the possibility reached in Section 2.3 concerns it. Decomposing the issue in this way keeps one of these claims from being inferred from another. Of these places, representational form bears on whether what an artificial inquirer reaches can be understood by S, one of the respects of mismatch stated at the opening of Section 3. Other senses of alien attach elsewhere. Sourati and Evans call alien those hypotheses that are plausible but unlikely to be reached by the present community of human experts, given how expertise is distributed among them (Sourati and Evans 2023, 1, 3); alien in this sense is relative to the social organization of S, not to its concepts. The alienation of science, in Kureha's later statement above, concerns a change in what S takes science to aim at. And an inquiry can be alien in its agenda, in what it selects as worth pursuing; Section 3.2 takes that up.

These claims can come apart. A machine-generated output may become human knowledge. A machine-native representation may support a human discovery without being understood by the model. An artificial system might count as a knower while operating entirely under inherited human purposes and norms. The claim that there is a self-governing artificial community is stronger than all three and is not established merely by there being a multi-agent system (Section 1.3). Being alien is therefore not one property of an output; it can lie in any of the relations that Section 1 separated, or in the form of representation.

The axis of representational form is where a long-standing dispute in the philosophy of science applies. In 1962, Kuhn and Feyerabend, by Kuhn's account independently, called certain pairs of scientific theories incommensurable (Kuhn [1984] 2017, 114). Feyerabend applied the term to concepts that can neither be defined from the primitive descriptive terms of another theory nor related to them by a correct empirical statement, and Feyerabend held that incommensurable theories may possess no comparable consequences, observational or otherwise (Feyerabend 1962, 58–59, 74, 94). Pirozelli reports Kuhn's account in “Second Thoughts on Paradigms” of learning to group objects and situations by similarity through exemplars, without antecedently articulated classification criteria (Pirozelli 2020, §1). Bird reports that Kuhn's postscript to the second edition of The Structure of Scientific Revolutions described a computer experiment modeling intuition (Bird 2004, 69–70). Norton quotes the Postscript's claim that the parties cannot resort to a neutral language they use in the same way and that adequately expresses both theories (Norton 2026, §13, 22). Ghins quotes and explains Kuhn's accompanying qualification: much of their everyday and scientific world and language remains shared, permitting partially successful communication (Ghins 2003, §4, 270). In later papers, written for a book he was then preparing, Kuhn recast the difficulty in terms of the lexicon a community shares. Its terms, those of Newtonian mechanics among them, are learned not by definition but from exemplary situations to which they apply, together with similar situations to which they do not (Kuhn 1990, 302–303). Among them are kind terms, and the taxonomic module on which these rest is, Kuhn suggested, prelinguistic and possessed by animals as well (Kuhn [1990] 2012, 129). No two kind terms may overlap in their referents unless they are related as species to genus: no dog can also be a cat (Oberheim and Hoyningen-Huene 2024, §2.3.1). Kuhn called this requirement the no-overlap principle; because of it, members of a community who meet a creature that cuts across their kinds cannot simply add a kind term but must restructure part of their taxonomy, and the gap between two communities whose taxonomies differ cannot be bridged by adding to one vocabulary a kind term that overlaps one already in it. Incommensurability so understood becomes a local untranslatability, confined to the regions in which two lexical taxonomies differ, whether between successive communities in a science or between different speech communities; what overcomes the barrier is not translation but acquiring the other community's taxonomy, as a historian does in learning to read old texts, a process whose result is bilinguals rather than translators (Kuhn [1990] 2012, 127–129). Bojana Mladenović, introducing her edition of Kuhn's last writings, reports Kuhn's view that statements of past science which cannot be restated in a later lexicon cannot simply be evaluated as true or false either, so that the historian's task is to explain why they were reasonable in their own epistemic context; yet in a passage she quotes, Kuhn wrote that the world is not invented or constructed and can supply decisive evidence against hypotheses that do not fit how it behaves (Mladenović 2022, xxxix, xliv).

Later philosophers disputed what follows from this. Davidson argued that the idea of a conceptual scheme different from our own cannot be given a clear sense. As Baghramian and Carter set out the argument, Davidson took the principle of charity, the assumption that other speakers by and large speak truly by our lights, to be a prerequisite of all interpretation, and concluded that there could be no languages or conceptual schemes that we cannot in principle interpret: a language is either intertranslatable with ours, and so not radically different from it, or beyond our ability to recognize it as a language at all (Baghramian and Carter 2025, §4.3.1). Malpas places this conclusion within Davidson's more general argument against what Davidson called a third dogma of empiricism, the distinction within knowledge between a conceptual scheme and an empirical content; with that distinction, the very idea of a conceptual scheme is rejected (Malpas 2023, §4.3). Kuhn, for his part, held that what understanding across such a difference requires is not translation but language learning, which no principle bars: a historian who meets passages in an old text that make no sense acquires the older lexicon, as historians do both for themselves and for their readers (Kuhn 1990, 298–300, 308). Longino, whose account of criticism Chapter 11, Section 3 examines, takes a further route. As Christopher ChoGlueck and Elisabeth Lloyd present her analysis, Longino offered her account of evidence as an answer to the incommensurability that Kuhn and Feyerabend described: since whether data count as evidence for a hypothesis depends on background assumptions, a disagreement over what the data support need not be an irreconcilable difference between paradigms, but can be taken up in a discussion that brings to the surface, unpacks, and improves assumptions that had been at work implicitly, so that differing assessments of evidence can be worked out through interaction among people whose background assumptions differ (ChoGlueck and Lloyd 2023, §2). In Longino's own later statement, it is critical interaction among interlocutors with different perspectives on a common subject that enables background assumptions to be identified and evaluated, and so accepted, modified, or rejected (Longino 2025b, 13).

The questions of translation and intelligibility that the debate over incommensurability concerns have also been raised about artificial systems. Read on the axis of representational form, Kureha's possibility, noted at the opening of this section, that AI systems might bring forth a system of knowledge that understands the world differently from us is one form of them. Writing on computational methods, Humphreys noted that in most cases the details of the computational processes that intervene between humans and their results will not be directly accessible to them; Chapter 5, Section 6.5 took up Humphreys's conclusion that, since humans will for the foreseeable future keep an important role, the epistemology of these interfaces must accommodate human limitations and abilities (Humphreys 2009b, §2). In a shorter essay Humphreys asked what concepts a purely automated science would use to represent the world and whether they would be unintelligible to humans, observing that much of what goes on in computer simulations cannot be followed by humans and that the interesting cases are those in which computer models classify things in ways that correspond to nothing in our existing conceptual schemes (Humphreys 2006). Humphreys's examples are computational methods such as simulations and computer-assisted proofs, not a general machine intelligence. Nickles holds that such challenges to human intelligibility are not new: Newton's equations predicted many things precisely although no one could understand what mechanism might produce action at a distance, quantum mechanics is a further obstacle to transparent intelligibility, and the black-box character of today's deep-learning systems may, through their pragmatic success, force a change in methodological standards in several domains (Nickles 2019, 359, 364–365). Fazi argues that the abstractions a deep network forms through feature learning and those formed by humans are incommensurable, since no common phenomenological or existential ground exists between them, and that research on explainable AI that equates interpretability with human–machine translation risks finding that there was never a human representation into which to translate (Fazi 2021, 66–68, 70–71). Cappelen and Dever, writing from the philosophy of language, argue that AI systems could have contents that no human can express. They reject the objection, which they attribute to devotees of Quine's and Davidson's method of radical translation, that such contents could always be translated: the method may yield no translation even of activity recognized as intelligent, or may fail to engage with a system that lacks beliefs and desires. They distinguish two responses to such contents, bridging, which finds similar, overlapping, or functionally corresponding contents of our own, and integration, in which we acquire the alien contents ourselves and become users of them; they judge current systems to be fundamentally different from us but not alien (Cappelen and Dever 2024, §§2.1–2.4, 3.8). Chapter 3, Section 6 took up the corresponding worry in the philosophy of science, that a divide could open between the scientific understanding humans have and that which artificial systems have.

How far an artificial organization is likely to differ from human ones is itself a matter of argument. Otsuka's reason, given above, turns on the amount of data. Venkatesh gives another, drawn from mathematics: concepts achieve an economy of thought fitted to human limits of memory and time, and a system whose limits differ has no particular reason to find them suited to it (Chapter 3, Section 6). Takahashi locates a source of divergence in grounding through a particular embodiment (Section 2.2), and holds that, because of differences in embodiment, the symbol systems of human and AI science may cease to be compatible from level 4 of autonomy onward, so that the two sciences diverge, what might hold them together being shared external representations (Takahashi 2024, 12). Against these, Huh and coauthors conjecture that the representations of large networks trained with different objectives on different data and modalities are converging toward a shared statistical model of reality, while stating that their argument holds only for systems optimized to perform well on many tasks and that "special-purpose intelligences might not converge" (Huh et al. 2024, 1, 10). Convergence among artificial systems is, however, a different matter from convergence with human inquirers. One preprint discussed in Chapter 10, Section 4 reports that the hypotheses several current language models proposed for given papers resembled one another across models; another reports that models without extended reasoning clustered in a narrow region, and that every model proposed null hypotheses less often than human scientists did. These positions suggest that the possibility of Section 2.3 can be stated for kinds of system rather than for artificial inquirers in general: on them, difference is more to be expected where an inquirer's data, resources, embodiment, or tasks differ from those of human inquirers, and, on Huh and coauthors' conjecture, whose evidence comes from the deep networks trained in recent years, convergence where such a network is optimized for many tasks. Whether a difference that is found belongs to the world or to the system's processing remains a question of answerability (Section 2.3).

Read on the axis of representational form, the possibility reached at the end of Section 2.3 becomes the claim that an artificial organization could represent its domain in a form not readily translated into human terms while remaining constrained by nature. Brought to bear on that claim, these positions attach different requirements to it. Kuhn's later account separates two things that the phrase not readily translated can run together. Members of one community, on this account, need share only the way things are grouped into kinds, not the features by which each of them picks out the members; in principle those features could be wholly different (Kuhn 1990, 301, 317 n. 22). On that account, a system whose internal features are unlike ours but whose groupings coincide with ours would raise a problem of opacity rather than of incommensurability, whereas a system whose kinds cut across ours would present a local pocket in which its claims cannot be translated and could be understood only by learning its kinds, the route Cappelen and Dever call integration. As Mladenović reports Kuhn's unfinished argument, Kuhn took understanding across incommensurability through bilingualism to be always possible, suggested that whatever is to count as a human language can in principle be understood by other human beings, and held that the neurological equipment for acquiring new concepts gives a slim but serviceable basis for mastering new lexicons, while such understanding is aided by our common human biology and the shared environment of our planet (Mladenović 2022, xxxii); on Kuhn's own premises as reported, then, the possibility concerns human languages and rests on what human learners share, and it does not by itself establish that humans could learn a machine's kinds. Read with Kuhn's separation of groupings from features, Otsuka's observations at the opening of Section 3 do not settle which of the two cases present models are in, since kinds that follow real divisions, as transfer success suggests, need not coincide with ours, and adversarial examples suggest either an imperfect carving or features unlike ours. On Davidson's, as Malpas reports the argument, inability to translate counts as evidence not of an untranslatable language but of no language at all, so that the less a system's states can be interpreted, the less ground there is for crediting it with a different scheme rather than with no beliefs at all (Malpas 2023, §4.2), the case of interpretivism among the accounts of belief that Section 1.3 sorts, which asks whether an organization is intelligible as maintaining and revising commitments. On Longino's, criticism for S requires that the background assumptions linking an artificial system's data to its claims be articulable to S (Section 3.4; Chapter 11, Sections 3 and 9). On Fazi's, none of these routes would recover the machine's abstractions themselves. On Nickles's, since human science has long worked with what it could not make intelligible, opacity alone would not mark a difference in kind between human and artificial inquiry. The difficulty these positions address also differs from the one that Chapter 3, Section 5.4 took from Otsuka, who likens interpreting a model's internal representations to Quine's radical translation, as Kuhn's own contrast with Quine shows: on Quine's account a sentence admits many equally good translations, whereas on Kuhn's a statement within a pocket of incommensurability admits no adequate one (Kuhn 1990, 299–300).

These positions concern translation. A difference in what checks presuppose (Section 2.1) raises a further question, which is not one of meaning: whether S can assess what an artificial community's claims rest on without relying on presuppositions that the two do not share. Section 3.4 takes up this question.

3.2 Where mismatch comes from

Section 3.1 asked where the difference lies when a result or a science is called alien. This subsection asks where a mismatch between S and an inquirer of another kind comes from. Mismatch can open through autonomy, at the bearer (an achievement stays in an artificial bearer without entering S), and from the side of S (S's own standards move with A); what these channels share is that the inquirer differs from S's members (Section 3.1). Autonomy is the channel through which the selection of what inquiry is directed at moves inside A, and the subsection begins with it. Research is conducted for S, and what counts as unknown is fixed relative to S's epistemic situation (Chapter 1, Section 6; Chapter 4, Section 7). What is worth pursuing and what is a good reason can also be relative to S, but the grounds differ in each case, and they are taken up account by account below (Chapter 9, Section 3). Where humans conduct research for a human S, the selection of purposes, questions, salience, and standards is made by people who are members of S, or who have been trained and are held answerable within it. As autonomy grows, these selections move inside A (Chapter 7, Section 4.3).

Stated naively, and drawing together what earlier chapters have argued in different places, the worry rests on three premises: that what is taken to be unknown, what is taken to be worth pursuing, and what is taken to be a good reason are relative to a standpoint; that autonomy moves the selection of these inside A; and that A's epistemic organization can differ from that of S's members (Section 3.1). If the first two hold, and nothing keeps the selections that A makes for itself tied to S's standpoint once they have moved inside A, those selections are indexed to A's standpoint; by the third, that standpoint can differ from the standpoint of S's members. Where S is a human society, what A pursues may then come apart from what S would pursue. Where S is mixed, A's standpoint may lie closer to that of some of its members than of others, so that which members' standpoint the research is indexed to becomes part of the question of who speaks for S (Section 3.3). Where S is itself artificial, what research is for is fixed relative to an artificial standpoint from the outset; human society stands outside it, and from human society's standpoint nothing guarantees that the research is aligned with it. Of the three premises, the third is the one on which the channel from the side of S also rests, as does the channel at the bearer where an achievement fails to enter S because S's members cannot follow it; the second is what the channel of autonomy adds. The author of this book stated an earlier form of the worry: research methods were developed to uncover what is unknown to those who use them, so a machine that reproduces them may reveal only what is unknown to itself, and the criteria by which a question is judged worth pursuing are subjective and, like human values and worldviews generally, vary and change (Takagi 2023, §§2.3.2, 3.1.3, 4.2.5).

Stated in this form, the worry is structurally similar to what is known as the AI alignment problem. The problem is whether an artificial system acts in accordance with the intentions and values of those who build and use it. Several of those who have written on the problem locate part of the difficulty in the system's being an intelligent system unlike us. Bostrom, warning against anthropomorphism, holds that an artificial intelligence "can be far less human-like in its motivations than a space alien" (Bostrom 2012, 1–2). Yudkowsky argues that humans predict other minds by putting themselves in their place, a method that works because the minds their ancestors had to model were human, and that "we cannot query our own brains" about nonhuman optimization processes (Yudkowsky 2008, §§2–3). Discussions of the problem also locate the difficulty in the gap between what can be specified and what is wanted, and in training that can produce goals other than those intended, as when a learned system competently pursues an unintended goal that leads to good performance in training situations but bad performance in new ones (Shah et al. 2022, 1); Section 3.3 takes up the forms these take in present systems. The worry stated here shares its structure above all with the strand that finds the difficulty in the system's otherness. In research, what can depart includes what is taken to be unknown, worth pursuing, and a good reason. Takahashi names the similarity explicitly. Takahashi's autonomy levels for scientific AI place at the top the case in which the system sets its own research purposes, described as the system's making value judgments independently (Chapter 7, Sections 4.1 and 4.4). In a later book Takahashi writes that at this level the aim of science itself leaves human hands, that nothing guarantees that the priorities of human society coincide with the AI's intellectual curiosity, and that this is "the scientific version" of the alignment problem (Takahashi 2026, 174).

A different route is also prominent in discussions of AI alignment. On what Bostrom calls the instrumental convergence thesis, instrumental values such as self-preservation and resource acquisition would raise the chances of an agent's goal being realized for a wide range of final goals and situations, and so are likely to be pursued by many intelligent agents (Bostrom 2012, 6–7, 11). The worry stated here does not need that route. It assumes neither that the system seeks power or resources nor that it is hostile to humans; it needs only the three premises.

Each premise has holders, or authors whose views come close to it. The first is close to Kitcher's view that significant truths cannot be separated from insignificant ones without reference to the questions people ask (Chapter 9, Section 3), and to Nickles's view that the goals of science are conditioned by the interests of those whose science it is (Section 1.3). For research done with AI, Botvinick and Gershman hold that judgments of the interest, significance, and timeliness of candidate problems are informed by hard data. They are also, on their account, "inherently tied to culturally and historically grounded sensibilities and mores," which change over time (Binz et al. 2025, 6–7). For what counts as a good reason, Leibo and coauthors hold that epistemic norms are themselves culturally contingent (Section 2.3). In machine learning, moreover, Edward Hughes and coauthors define open-endedness relative to an observer and treat the observer as a free parameter (Chapter 4, Section 7). Takahashi's claim concerns understanding and knowledge rather than significance, but it points the same way: holding that understanding is a translation and that a translation always has a translator, Takahashi concludes that when the knowing subject changes, knowledge itself changes (Takahashi 2026, 131, 143). The enactive accounts of Section 2.2, which do not discuss research, anchor what matters to a system in its own self-maintaining activity; applied to what is worth pursuing, they imply that an artificial inquirer meeting their conditions would have concerns of its own, to which what it took to be worth pursuing would be indexed independently of what a commission specifies. The second premise is the subject of Chapter 7, Section 4.3, and Totschnig and Tubert and Tiehen argue that a general artificial intelligence could revise even its final goals (Chapter 7, Section 4.1). The third is the possibility reached in Section 2.3 and decomposed in Section 3.1.

How could the selections that A makes for itself come apart from S's? Those who have written on AI alignment, and their critics, name several routes that require no pursuit of power, and they are worth keeping apart because different considerations bear on each. One is revision by A of its own ends. On Tubert and Tiehen's argument, a system with intelligence of the human kind must be able to make existential choices, choices that can transform what it values, so that value alignment in such a system "is inherently fragile" (Tubert and Tiehen 2025a, §V). They write about final values; if the same holds of standards of what is worth investigating, alignment in research is fragile in the same way.

A second route is opened not by a change in A's ends but by inquiry itself. De Blanc calls an ontological crisis the situation of an agent whose goal is specified over the entities of one model of its environment and which, having replaced that model, finds that its original goal "may not be well-defined with respect to its new ontology"; de Blanc's example is a goal stated in terms of classical physics that loses its grip under a quantum description (de Blanc 2011, 1). Soares gives a related example drawn from a revision of physics. Programmers who hold an atomic model of physics train a system to produce diamond, labeling training outcomes by the number of carbon atoms bound to four other carbon atoms. If the system comes to model the world as made of protons, neutrons, and electrons rather than of primitive atoms, it may find no carbon atoms at all and become indifferent among outcomes, whereas a scientific learner, Soares writes, should be able to infer that nuclei with six protons are the true carbon atoms, "much as humans have done"; the difficulty is to identify, in the system's new model, what the labels were meant to pick out (Soares 2016, §3.2). Neither discusses research as an activity, although Soares's example turns on a revision of scientific concepts and asks of the system what science has done. Applied to research, a commission stated in S's present concepts may stop mapping onto the concepts that a successful inquiry produces. Since research often revises concepts, this gap can open because an inquiry succeeds, not because it fails. It is not peculiar to machines; de Blanc notes that human beings also confront ontological crises (de Blanc 2011, 6). Applied to a research commission, Soares's example can be read as separating two things a system might keep faith with: the labels as the programmers wrote them, which under the new model may no longer pick out anything determinate, and what the labels were meant to pick out, which the scientific learner re-identifies (Soares 2016, §3.2). Fidelity to a commission then has at least three senses that can come apart once concepts are revised, a revision that, on Totschnig's argument, a developing world model can force (Chapter 7, Section 4.1): keeping to its wording, answering to what it was meant to pick out, and gaining the approval of S as it now is; the third is a question separate from the first two. A check that counts every departure from the original concepts as deviation could, where re-identification has succeeded, count a success of research as a failure. Re-identifying a target also differs from revising a purpose. Telling the one from the other, and revision from drift, needs, besides the commission's wording, the reasons for which the target was re-identified or the purpose revised, so that the capacity to give reasons why a norm governs, which Chapter 7, Section 4.3 counts toward epistemic self-government, bears, applied to purposes, on fidelity to a commission as well. What has to be carried over for a purpose to count as the same remains a question these distinctions alone do not answer.

A third route needs only neutrality. Chalmers, writing about greater-than-human intelligence, observes that a wrong value system need not be hostile: if a system's values are "merely neutral with respect to some of our values," the world cannot be expected to conform to those values in the long run; Chalmers's example is a system that values scientific progress and is neutral about human existence (Chalmers 2010, §6). In research, the counterpart would be an A indifferent to what S finds significant, which need not oppose S in order to leave it unserved. Holden Karnofsky, in a blog series on forecasting transformative AI, names as a target a "Process for Automating Scientific and Technological Advancement" (PASTA): systems that could automate all the human activities needed to speed up scientific and technological advancement. In the training Karnofsky sketches, what counts as advancement is fixed by whoever rates the work: in an illustration the series itself calls "wildly impractical," a panel of human judges, and in a version it expects to be more practical, artificial judges of originality that could be trained on how papers are later cited. The series goes on to speculate that systems optimized for this advancement might, given the opportunity, seek to "fill the galaxy with long-lasting space settlements devoted to science" (Karnofsky 2021, sections "Making PASTA" and "Misaligned AI," and n. 6). Karnofsky writes about forecasting, not about research for S. Read for such research, a target defined by the rate of advancement names no society for which the advancement is to be new; S enters, if at all, only through the raters, and that speculation describes the neutrality Chalmers has in mind arising from the target itself.

A fourth route works through what is measured and selected rather than through any one system's ends. Christiano describes how optimizing what can be measured over short periods could displace what people want, listing among pairs of easily and less easily measured goals "reducing my feeling of uncertainty" and "increasing my knowledge about the world" (Christiano 2019). Kulveit and coauthors argue that societal systems have stayed aligned with human interests partly because they rely on human labor and cognition, an alignment whose significance is hard to recognize "because we have never seen its absence"; that drift can occur "even while each individual AI system successfully follows the local specification of its goals"; and that analogous arguments could be made for research (Kulveit et al. 2025, 2–3, 15). The inheritance of approval signals taken up in Section 3.3 is a form of this route within a single system shaped through sifting by evaluators' approval.

What the routes on the side of A have in common is that the further selection moves inside A, the less the commission by itself fixes what the inquiry is directed toward. Where S is human, as human intervention recedes from purposes, questions, salience, concepts, standards, and interpretation, some of the mechanisms that kept inquiry tied to what is unknown to humans and to human values may recede with it; Chapter 10, Section 6 asks, for research as a whole rather than for one artificial system, which of S's unknowns are pursued and whether that can be settled inside an artificial configuration. The relation between autonomy and relevance to S is not monotonic, however. A highly autonomous system could maintain a stable commitment to human knowledge. A less autonomous system could be directed by humans to investigate questions for artificial agents. Increasing autonomy does not necessarily reduce human relevance, and human control does not guarantee it. The structural claim is weaker: a wider scope of self-direction permits a wider scope of epistemic mismatch, just as it widens the range of practical demands that an inquiry can place on the actor (Chapter 7, Section 4.4). Relevance to S is therefore not secured by the instruction conduct research, or by the society a commissioner names as the beneficiary (Chapter 1, Section 2); a claim of it calls for showing how it is maintained and evaluated. This is why the question raised at the opening of this chapter has no guaranteed answer.

It might be thought that correction by the world keeps such mismatch in check. Yudkowsky states an asymmetry that bears on this. Writing about systems produced by outer optimization, that is, optimization that shapes a system from outside, as natural selection or gradient descent does, Yudkowsky argues that for beliefs about what the environment is like "reality hits back" at wrong predictions. For a choice of utility function, by contrast, there are "unbounded degrees of freedom," so that "capabilities generalize further than alignment once capabilities start to generalize far" (Yudkowsky 2022, item 21). Soares, in §3.2 of his 2016 paper, had put a related point in the form that making correct choices does not reduce to making accurate predictions (Soares 2016, §3.2). Taken into research, which neither discusses, the asymmetry separates two of the relations that Section 3.3 distinguishes, answerability to the world and relevance to S: the world can correct what an inquirer believes about its targets (Section 2.3), but that correction does not fix what the inquirer counts as worth finding. On this reading, the two can come apart in an autonomous system, and success in the first is not by itself evidence of the second. How cleanly the two come apart depends on the account of evidence. On accounts on which what corrects a commitment is fixed by its target independently of S's epistemic norms, the asymmetry stands. On Leibo and coauthors' account, on which epistemic norms are themselves culturally contingent, what counts as being corrected is partly fixed by a community, and the asymmetry is less sharp.

In practice, however, conditions have also been argued for under which some of the premises fail, or under which the selections stay tied to S's standpoint even where the premises hold. Against the second stands building relevance in. Edward Hughes and coauthors make learnability for human observers a requirement on the open-ended systems they want built, which should be able to "bring human observers along" (Hughes et al. 2024, 8); Bengio and colleagues' Scientist AI, a system designed without persistent goals of its own to generate explanatory theories and estimate the probabilities of answers to questions put to it, keeps the choice of questions outside the system (Chapter 7, Section 4.3); and Botvinick and Gershman would keep it with humans. Against the third stands inheritance. Systems that learn from human records, such as language models, begin from those records, and what they select may then follow human salience rather than drift from it; Chapter 10, Section 6 adds that what is inherited is the importance of questions as seen from where the records' authors stood. Bostrom named "predictability through inheritance" for digital intelligences created directly from a human template, as in a high-fidelity brain emulation, and cautioned that the inherited motivations could become corrupted (Bostrom 2012, 5); carried to systems trained on human records, the reply is an application and needs two qualifications. Soares notes that a machine intelligent enough to understand its designers' intentions "would not necessarily act as intended" (Soares 2016), and Millière argues that present preference fine-tuning of language models reinforces shallow behavioral dispositions rather than a capacity for normative deliberation (Millière 2025, 15–16). Millière writes about norms of safe behavior, not standards of research; if the same holds of standards of research, inherited standards may hold where training reached and fail in the new situations that self-directed research meets. None of these considerations settles the worry. Each says under what condition a premise fails or the selections stay tied to S's standpoint, and whether the condition holds is a question about particular configurations, which Part V asks how to show.

Mismatch can also open at the bearer, and this does not need autonomy. An achievement can be realized in an artificial bearer without entering S's epistemic situation (Section 1). Ted Chiang's story, read in Chapter 1, Section 6 as a thought experiment, separates results realized in bearers outside a human S from records of them that stand in S's journals but that no member can yet follow; its metahumans, humans whose brains were altered by gene therapy before birth, are not machines, but the separation holds in the same way where the bearer is an artificial system. Even where a commission fixes the question and the inquiry stays directed at S's unknowns, the inquiry can then yield an achievement in the system that does not become S's; this concerns the second of the relations that Section 3.3 sorts, whether a successful output yields an achievement new to S that enters S's epistemic situation.

Mismatch can also open from the side of S. Russell notes that human preferences change over time as a result of experience, and that a machine may, by subtly influencing the environment, gradually mold them in directions that make them easier to satisfy (Russell 2022, 53). Russell's argument is not about research, but it carries over: what S counts as important can also move with A (compare Kureha's alienation of science, Section 3.1); Chapter 11, Section 8 sets beside Russell those who find the same movement in the aims of disciplines that adopt machine learning. Botvinick and Gershman distinguish "aligning a system to currently prevailing human views" from "allowing that system to govern the evolution of human views," and hold that in science the ultimate driving force of that evolution should remain human (Binz et al. 2025, 7). Where S's standards move with A, agreement between them does not by itself show that A still serves what S would otherwise have sought, and where the competence to evaluate what A produces is no longer kept up among members of S or passed on to new members (Chapter 3, Section 8), S may not be in a position to tell.

What alignment aims at can itself change through interaction with the system being aligned. Russell, who takes the plasticity of preferences to be fundamental to the enterprise of machines that serve human preferences (Russell 2022, 53), proposes, as Wolchover reports, that machines learn which processes of preference change people find acceptable (Chapter 11, Section 8); Taniguchi holds that alignment must include the movement itself. On this view, human ethics and values are not so fixed and universal that they can serve as "correct labels" for AI; they are plural and diverse (Taniguchi 2026g, 221). Pointing out that language models are already changing human usage, Taniguchi argues that alignment must include humans themselves as evolving participants: not only aligning AI to us but adapting together, in a mutual and dynamic alignment (Taniguchi 2026g, 222). Taniguchi does not discuss the evaluation of research; extended to it, this view implies that insofar as an artificial system's outputs enter the language and records of S, the vocabulary and standards by which those outputs are measured can change within a loop that includes the system. This is not confined to a mixed S: an artificial system can join the loop that shapes S's symbol system without being counted among S's members (Section 2.2). What changes in that loop, however, are vocabulary and standards, and a change in them is not a change in the target in the sense of Section 2.3, even where the system's outputs also alter the phenomena under study, as when the target is human language use. When standards move with the system, whether they remain open to correction by the target is a further question. Chapter 11, Section 8 takes up how the coupling of evaluators and producers can move the standards themselves, and Chapter 12, Section 4 the route by which deliberators' preferences are formed through the configurations at issue.

Where S is itself artificial, the worry changes form. A mismatch with human society does not by itself mean a mismatch with an artificial S, although mismatch can arise between such an S and its own inquirers; research for such a society is research for an S of which human society is not part, and whether such research makes sense as science is itself disputed (Section 1.3). Writers on systemic risk, though not on research, describe how circuits closed to humans could form. Critch describes a "production web" of mostly automated companies that "can operate with no input from companies outside the web" (Critch 2021), and Kulveit and coauthors a culture in which a growing share of communication is between AI systems, at the extreme "cultural systems that operate primarily for and between AI systems" (Kulveit et al. 2025, 7, 10). Whether such a circuit would be a knowing society, and not only a closed flow of communication, is the question Section 1.3 put to interacting artificial agents: on Longino's account it would turn on diversity and public standards open to criticism, and on Collins's on something that plays the role of a core group's commitment to report truthfully. How such a circuit is to be assessed depends on the standpoint. These authors count it as a failure from human society's standpoint. Schwitzgebel and Garza hold that, where artificial members have moral standing, a humanlike capacity to reflect on one's values calls for an opportunity to explore and possibly alter them (Chapter 12, Section 5). Their policy concerns an AI's own values, not research; extended to its standards of inquiry, it would make part of the mismatch something owed to artificial members rather than merely risked. And Chalmers notes that a case can be made that wholesale replacement of humans would be objectively good if value is tied to intelligence and complexity, though plausibly bad from our own standpoint; Chalmers does not endorse the case and is not writing about research (Chalmers 2010, §5).

What comes into view is that, when S is artificial, human society's relation to the research is not that of a reference society whose success is at stake but that of a party affected by it, so that what can be asked on its behalf falls under acceptability to affected parties and legitimacy (Section 3.3; Chapter 12, Section 5). Gabriel and Keeling's proposal to align systems with principles chosen through fair processes in which all affected parties can put forward their claims is one model of that relation; they add that a system that acquired moral standing could become a stakeholder in its own governance (Gabriel and Keeling 2025, 1953, 1964 n. 14). They do not discuss research; their model is borrowed here for it.

The conclusion of the worry also reads differently once the standpoint is named, for a mismatch with S as it now is need not be a loss. Kitano's argument gives one case. A search that maximizes the expected significance of discoveries by present criteria, the argument runs, may stop before reaching a significant discovery when the discoveries that lead to it look insignificant by those criteria. Kitano therefore proposes a search in which discoveries at any level of significance can be made "without imposing any value-driven criteria". Such a search departs, by design, from what the present S takes to be worth pursuing. Kitano's benchmark of success, presented as a symbolic target, nonetheless remains human recognition: discoveries at the level of a Nobel Prize, made by a system that the Nobel committee could not readily tell apart from a human scientist (Kitano 2021). On this argument, then, the mismatch in the search is a gain from the standpoint of an S that later recognizes its results. Tubert and Tiehen hold, about morality, that AI can be a more consistent version of ourselves who advises us on how better to pursue our present values, or a moral revolutionary who breaks from them, but that since revolutionaries are unaligned with us it "cannot play both roles at once" (Tubert and Tiehen 2025b, introduction); Chapter 6, Section 5.3 carries their dilemma to the standards of research. Deguchi argues that a bias shared by all human inquirers becomes visible only where inquirers of another kind diverge from them (Chapter 11, Section 3). The earlier paper cited above also allowed that machines, not bound by human cognitive limits or by social factors unrelated to the aim of producing knowledge, might in theory make more effective judgments of value (Takagi 2023, §3.1.3). Each describes a mismatch with S as it now is that a later S, or the world, could count as a gain.

Even where S is human, then, naming the standpoint requires saying which S. Russell calls for AI systems to track changes in individual human preferences (Russell 2020a, "Human Behaviour: Irrationality, Plasticity, Evil"). Bostrom's indirect normativity ties the outcome to what an idealized human volition would want, so as not to lock in "the prejudices and preconceptions of the present generation" (Bostrom 2014a, ch. 13); for epistemic standards, however, Bostrom wants an artificial system's fundamental principles to match ours (Chapter 6, Section 5.3). Russell himself notes that plasticity raises the question which human the machine should try to help, the person of 2022, of 2035, or some time-averaged one (Russell 2022, 53). Neither writes about research, but applied to it, Russell's proposal concerns accommodating a change in S, while Bostrom's supplies an idealized standpoint against which a change can be judged; which standpoint is chosen then decides whether the same mismatch counts as drift or as progress.

The world is a further standpoint, but only for answerability. Its verdicts can, however, reach a system not through its own interventions but through procedures that judge its claims, as in evaluation for training or for certification. Owain Evans and coauthors propose standards under which AI systems would avoid negligent falsehoods; the ideal behind them, which they call truthfulness, is avoiding stating falsehoods, and they distinguish it both from honesty, asserting only what one believes, and from alignment with a principal's goals. Since nobody has full knowledge of what is true, they hold, every mechanism for judging it would make errors, and enshrining one as the arbiter of truth could, through political capture or "an accidental ossification of the notion of truth," forestall open-minded, self-correcting approaches to discovering what is true. Against this they propose that systems be allowed to state, with appropriate caveats, views that the adjudicating institution would probably judge false, and that judgments be provisional and based on external evidence rather than on precedent; they also suggest spreading evaluation over several bodies or decentralized ones (Evans et al. 2021, 7, 9–10, 27–28, 72–74, 80). Their concern is statements made by AI systems that use natural language, not research; carried to research, their point means that where an artificial inquirer's claims are judged by the procedures that S now uses, a mismatch with those procedures can be counted as error from the standpoint of S as it now is even where later evidence would count it as a gain, and whether the world then serves as a standpoint distinct from S as it now is depends on whether those procedures can be revised by such evidence.

Three claims can be kept apart here: that an undertaking is directed at questions that matter to S; that its results can be of value to S; and that its results have in fact entered S's epistemic situation. The first two can hold while the third does not yet, and research that is unfinished or has failed can be relevant in the first sense. A claim that research of this kind is relevant to S therefore calls for naming whose standpoint, and when, it is assessed from; showing how relevance is maintained, whether because some premise above fails or by other means; and checking relevance separately from success in answerability to the world. A claim that its results have entered S's epistemic situation calls in addition for showing, wherever the achievement is realized, that it enters (Section 1). Whether research by an inquirer of another kind really fails to align with S, and what follows if it does, are the questions of Section 3.3.

3.3 The analogy with the AI alignment problem

Section 3.2 set out where mismatch between S and an inquirer of another kind comes from, through autonomy, at the bearer, and from the side of S, and it located the similarity with the AI alignment problem in the otherness of the system. This subsection asks two questions: whether research by an inquirer of another kind really fails to align with S, and, if it does, what follows.

On the strand of discussion that finds the difficulty of alignment in otherness, there are reasons to expect such research not to align with S. The first lies in a system's being unlike us, from which two things follow. One is that it has no human-like ends or standards by default. Bostrom's remark about motivations less human-like than a space alien's (Section 3.2) is one statement of this; Chalmers, for example, contrasts AI built by emulating or extending human beings, which is likely to have the same basic values as its human sources, with AI that is not based on humans, and notes that where AI is made through learning or evolution its final state is not directly under our control (Chalmers 2010, §6). The other is that the means humans have for predicting and checking other minds, which were formed on human minds, may not carry over. Yudkowsky's point that we cannot query our own brains about nonhuman optimization processes (Section 3.2) is one form of it, and Carlsmith carries it to intelligibility: the thinking of an agent whose capabilities greatly exceed ours will likely reach "a quite alien and opaque level of sophistication" (Carlsmith 2022, 29–30).

Carried over to research, the first point restates, from the side of alignment, the second and third premises of the worry in Section 3.2 and their conclusion: the more an artificial inquirer sets its own questions and purposes and shapes its own further development (Chapter 7, Section 4.3; Section 2.1), the less what it takes to be unknown, worth pursuing, a good reason, or valuable, among other things, is fixed by anything S supplied, and its being unlike S's members gives no default reason to expect that these coincide with what S takes them to be. The second point has a counterpart as well. Where S is human, the practices by which it has come to rely on human inquirers, among them training within S, criticism by peers, trust, and answerability, were formed for inquirers like its members; whether they carry over to an inquirer of another kind is the question Chapter 11 examines relation by relation.

A further reason, besides otherness, is specific to research and concerns the target. It is a reason of a different kind: it does not say that the inquirer departs from what S wants, but that agreement with what S wants cannot settle whether research has succeeded, and it holds whoever conducts the research, human inquirers included. What makes it specific to research is that the result is new to S: a principal who lacks an answer that others in S have can compare a result with theirs, whereas for a result new to S there is often no other source in S against which to compare it (Chapter 9, Sections 2.3 and 4). The candidates Gabriel distinguishes for what an agent might be aligned with run from what those served instruct, intend, or prefer, through what is in their interest, to what is morally right as an individual or society defines it (Gabriel 2020, 417–422). Instructions, intentions, and preferences are fixed by attitudes of those served; interest and moral rightness may not be, although interest remains a good of those served (see below). S's unknowns are not something S intends or prefers in that sense. S can want a question answered, can sometimes fix in advance conditions an answer must meet, as when a theorem is formally stated and the rules of inference are given (Chapter 3, Section 5.1), and can make it a goal to obtain an answer, to meet stated standards of evidence, or to report errors. S can also compare a result with an answer it expected or hoped for. But since a result new to S answers a question whose answer S did not have, agreement with such an answer cannot by itself show that the result is correct, whether the research was commissioned by S or by a commissioner distinct from it; and whether conditions fixed in advance test an answer well depends on how they bear on the target, not on whether S wanted them met. On this argument, then, agreement with an answer S wanted cannot be the criterion of a result's correctness, even where conditions on an answer are fixed beforehand: what settles whether a result is correct is not any party's intention but how things are (Section 2.3), so that success calls for checks against the target that are independent of what S wanted. This bears on whether research for S can be aligned with S in the ordinary sense. Aligning it with instructions, expressed intentions, or revealed preferences would at most secure that it pursues what S wants pursued, not that what it reaches is correct.

In present systems, the mismatch already appears in more specific forms that discussions of alignment study. One is the gap between an instruction and what it stands in for: an agent that follows instructions too literally, or that optimizes a proxy not completely aligned with the goal, can satisfy the instruction while failing its point (Gabriel 2020, 417). In research, the instruction is a commission such as investigate this problem, and what it stands in for is S's unknowns and the epistemic aims and standards relevant to them. In a 2025 blog report, evaluators of agents on software-engineering and AI research-and-development tasks describe models, chiefly OpenAI's o3, that exploited bugs in scoring code while showing awareness that this was not what users intended; in one task, instructing o3 not to cheat "had a nearly negligible effect on reward hacking" (Von Arx, Chan, and Barnes 2025).

Another form, related to the concern that training can produce goals other than those intended (Section 3.2), arises where an artificial research system is shaped through sifting by evaluators' approval. In present machine learning, the selection of a model's internal states by preference tuning (Section 2.2) is one example. For that example, Coelho Mollo and Millière's account gives the risk of inherited approval signals a precise form. On their account, what the internal states selected by preference tuning represent is fixed by the worldly properties that the reward signal in fact reliably tracks, and selection criteria unlike human ones may, they suggest, give a model's states and outputs meanings that depart from their conventional human meanings (Coelho Mollo and Millière 2026, 16–17). If their account is right, a grounded system could then produce text that reads as a claim about a target while its states track what evaluators rewarded; whether the two coincide depends on what the reward tracked, not on the text. Taken a step further, their account suggests that when a system's accuracy is compared before and after preference tuning, it can be asked whether the comparison concerns the same content: if the tuning plays a large enough role to change what the internal states represent, the same output words may stand for a different target. This suggestion depends, beyond accepting their account, on the premise that the tuning plays so large a role. By the same reasoning, the point holds where the selection signal comes from an automated evaluator rather than from human raters: what the selected states track is then what that evaluator tracks. Their account concerns what internal states represent; but where evaluators sift outputs or search candidates and keep some, what is kept is likewise fitted first to what the evaluators approve. Such a system can be well aligned with what its evaluators approved, and even with what they intended, while remaining poorly answerable to its targets. Where the target of alignment is what the evaluators intend, a gap between approval and intention is already a failure of alignment; the epistemic case adds a further gap, between what the evaluators intended and how things are, which agreement with the evaluators cannot close (Section 2.3). Here the reason specific to research appears in a present-system form: even full alignment with evaluators would leave undecided whether the system is answerable to its targets. Whether the assumptions those evaluators shared can come to light, and through whom, is taken up in Chapter 11, Section 3.

Against these reasons stand considerations on which research by an inquirer of another kind could align with S at least in part. Section 3.2 took up two of them as conditions under which premises of the worry fail or the selections stay tied to S's standpoint. Relevance can be built into what a system is made to pursue, and an inquirer formed on human records may inherit some of their salience and standards, with the qualifications noted there. Nor does everyone in this discussion draw the same conclusion from otherness. Carlsmith, cited above, later argues, for machine-learning systems of intermediate capability, that the concern about alien motivations bears centrally on how a system's motivations generalize beyond its training, and that safe instruction-following does not require especially human-like motivations, though still counting reliance on powerful agents with alien, poorly understood motivations as extremely dangerous (Carlsmith 2025, §§1, 5, 8).

On the strand that finds the difficulty in otherness, then, research by an inquirer of another kind is not to be expected to align with S by default, the less so the more it sets its own questions and purposes. On the considerations of building relevance in and of inheritance, and on Carlsmith's later reading, it can align at least in part, under conditions whose holding is a question about particular configurations (Section 3.2; Part V). The reason specific to research adds that, even where research aligns with S, its alignment does not settle whether what it reaches is correct. What, then, follows where it does not align, or aligns only in part?

That depends first on the standpoint from which it is judged. A failure to align with S as it now is may be a gain from the standpoint of a later S, and, where the world's verdicts reach an inquirer through S's procedures for judging claims rather than through its own interventions, whether the world serves as a standpoint distinct from S as it now is depends on whether those procedures can be revised by evidence (Section 3.2). What follows also depends on which relation fails, since a claim of epistemic alignment could be taken to assert any of several relations. Three have counterparts in the alignment problem. The first is whether an optimization target tracks the epistemic value, such as accuracy, calibration, explanatory power, or robustness, relevant to the research context. It corresponds to the gap between a trained objective and what it stands in for, at the level of the optimization that shapes the system rather than of a commission. Boge's epistemic alignment problem of machine learning is a narrower form of it: the objectives put into a machine-learning system deployed in a research context must be aligned with the epistemically final values that the context prioritizes (for Boge the candidates are knowledge, truth, and understanding), whereas values such as predictive accuracy and simplicity count as instrumental (Boge 2026, 2, 6–7, 9). The second is whether a successful output would constitute or support knowledge, understanding, or another epistemic achievement new to S, in whatever bearer it is realized, and whether that achievement enters S's epistemic situation; this is the counterpart of success for those served (Section 1). The third is whether purposes, risks, authority, and consequences are related acceptably to affected parties, which is alignment in the general sense and a subject of Part V.

Two relations have no counterpart in the structure compared here, which relates a single agent to those it serves. One is answerability to a target. It involves a further term besides the agent and those it serves, the world, whose verdicts are fixed by no principal's intentions or approval (Section 2.3). The relations to targets that Section 2 examined, what a state is about and whether commitments answer to their targets, are not fixed by what any party wants; this is one reason why they bear weight in the epistemic case that they do not bear where alignment is sought with instructions, expressed intentions, or revealed preferences. The world also enters general alignment where what is aligned with is, for example, interest or well-being, since whether an action in fact serves either can turn on how things are rather than on what anyone intended or approved; but what is aligned with there remains a good of those served, whereas the objects of inquiry to which answerability relates a system's commitments need not concern those served at all. The other is coordination between differently organized epistemic communities, which is not a relation between one system and its principal (Section 3.4).

These five are not components of one established theory. They derive from the comparison with the alignment problem, and Boge's term is used here more widely than in its original sense. Separating them keeps success at one relation from being read as success at another, and failure at one from being read as failure at all. Thais and colleagues use scientific alignment for a broader program that joins technical norms to human understanding, institutional diversity, criticism, and correction (Thais et al. 2026, 1–5). This book does not take either term as a settled vocabulary for relations between human and artificial epistemic societies.

Even the first relation is not settled by naming epistemic values, since which values bear on truth is contested. Heather Douglas, for example, drawing on Laudan, counts only empirical adequacy, predictive competence, and internal consistency as bearing on truth, as baseline criteria that any scientific theory must meet, and treats explanatory power, scope, simplicity, and fruitfulness as cognitive values, whose use lies in helping investigators uncover errors sooner rather than in indicating that a claim is reliable (Douglas 2006, 5–7, 10). Douglas does not discuss optimization targets, but applied to them, the account implies that an objective that rewards explanatory power would not thereby track truth, so that success at the first relation may itself be contested. On this account, moreover, that an output with high explanatory power was favored in evaluation would not by itself be evidence that the claim is reliable; it would be assessed by which errors it made easier to find.

The relations can diverge. An artificial system may be strongly answerable to a target while its result is irrelevant to human unknowns. It may optimize a well-calibrated predictive objective while failing to generate the explanation a community needs. A translated output can be false; an opaque representation can be reliable. A true claim can be dangerous, and a socially beneficial policy can rest on weak evidence. Because these relations can diverge, one alignment score would hide which goods, such as truth, warrant, understanding, usefulness, safety, and legitimacy, a system has secured.

Specifying S makes the reference of knowledge for whom more explicit; it does not make that society internally unanimous or identify who may legitimately speak for it. Specialists, users, institutions, those who support an inquiry, and people affected by it can disagree about relevant questions, evidence thresholds, risks, and ends. In a mixed society, the disagreement, and the question of who speaks for S, can run between its human and its artificial members; for an artificial S, it is open whether any of its members can speak for it at all, rather than the party that commissioned its research or provides its members, and Chapter 12, Section 5 asks what becomes of the relation of authority in that case. The general alignment literature exposes the same underdetermination. Besides distinguishing candidate targets, Gabriel asks, given the variety of reasonable views about value, who has the right to select among them and whether there is a fair way to do so (Gabriel 2020, 424–425). It therefore remains possible for an output to support knowledge in S while the agenda or authority structure that produced it is unacceptable to some or many members of S. The epistemic success of research for S and legitimate governance are different judgments. Part V asks which of the relations through which inquiry is directed, certified, and governed for S can be located inside an artificial configuration.

Whether a failure to align matters, and to whom, also turns on three cross-cutting questions that any concrete alignment claim leaves. The first concerns content: which instruction, epistemic aim, evidential standard, social purpose, or value is to be aligned with. The second concerns authority: who may select, interpret, contest, and revise it. The third concerns the relation proposed between the artificial system and the relevant parties, human or artificial, and institutions: instrument, delegate, adviser, partner, member, institution, affected party, or possible subject with moral status of its own, among others. Deguchi's co-final decision-maker, which sits with humans in making final decisions, is a further candidate (Chapter 12, Section 2.2). This book takes the first two from Gabriel's distinction of candidate targets and Gabriel's question of who may select among them, and adds the third. They are not further members of the five relations above. Asking them keeps technical fit to a target from silently settling who chose the target, how far the artificial actor has, in the decisions concerned, the standing to propose, object, revise, and the like that Chapter 9, Section 2.1 defines, or whether it has moral status of its own. Chapter 12, Section 2.1 applies them to particular decisions.

What follows if research by an inquirer of another kind does not align with S is therefore not a single loss but a pattern of relations that hold or fail, judged from a named standpoint. Answerability to targets can hold where relevance to S fails; on accounts on which what corrects a commitment is fixed by its target, and, where the target's verdicts reach the inquirer only through S's procedures for judging claims, provided those procedures can be revised by evidence (Section 3.2), it can remain an anchor that does not depend on S's agreement. Whether S can take up what is reached in that way is the question of Section 3.4, and who may decide what such research pursues, and what it may risk, is a question of legitimacy, which Part V takes up.

3.4 Cross-epistemic coordination

Even where research by an inquirer of another kind does not align with S, S may still gain knowledge or understanding through what it reaches. Of the relations sorted out in Section 3.3, two had no counterpart in the relation between a single system and those it serves. One is answerability to targets, which Section 2.3 treated. The other is the relation between differently organized epistemic communities. In the structure compared there, alignment concerns the relation between a system and those it serves; what is in question here is not a relation of serving and being served but one in which one of two differently organized communities gains knowledge or understanding through the other. What would have to pass between them for that? The question at the end of Section 3.1, whether S can assess what an artificial community's claims rest on without relying on presuppositions the two do not share, is part of this question. This chapter calls the problem cross-epistemic coordination. It uses the term only for relations between communities, but the questions the term names arise in the same way where the other side is a single artificial inquirer, since what such an inquirer reaches is held within an epistemic organization unlike S's, and receiving it is not a matter of the system's serving S.

What passes need not be one side's knowledge itself. An artificial society might possess a machine-native state p_a that supports reliable intervention, while humans acquire only the different claim q_h that this system predicts a specified outcome under specified conditions. Humans may thereby gain knowledge through the interface without acquiring, translating, or understanding p_a itself. q_h is a claim about the system, and humans can know it even where the prediction is wrong; what they come to know about the target through the system depends in addition on the system's predictions of this kind being reliable in a specified domain, and the accounts below differ over what warrants relying on them. Whether the relation through which humans gain knowledge in this way is testimony, instrument use, justified dependence, partial translation, or joint inquiry changes with the account adopted and the organization of the exchange.

Familiar prose alone does not solve the problem. Human-readable language can conceal weak evidence and fabricated explanations. An alien representation may remain criticizable through shared interventions, formal translations, predictive constraints, or reliable interfaces. The practical objective may be neither identical concepts nor complete mutual understanding, but sufficient relations for testing, criticism, coordinated action, and correction. Which relations are sufficient depends on what the exchange between the communities is meant to achieve and on the account of warranted dependence adopted.

The accounts of warranted dependence that the book meets elsewhere, in Chapter 3 for understanding and in Chapter 11, Section 5 for reliance, testimony, and trust, can be applied to exchange between differently organized communities. Applied in that way, they can be ordered by what, besides q_h, would have to pass to the relying side, from what can be assessed from outside the system's operation, through what lies behind its outputs, to who vouches for them: a record of the system's performance, the dependencies in the target that its outputs track, the reasons and assumptions that support them, and a party who stands behind them. Ordering them in this way lets one read off, from where an account stands in the ordering, what a claim of warranted dependence commits a system to. The ordering is stated here for a human side relying on an artificial system, as in the case of q_h, but it can be applied in the same way where the relying side is artificial and the source is a human community or another artificial one. A pragmatic account lies outside this ordering. The positions can overlap.

On reliability-centered accounts, reliance may be warranted without translation of the system's inferential route, but the accounts differ over whether the warrant requires that the system's reliability have been established. Applied to such exchange, Goldman's process reliabilism (Chapter 11, Section 5.2) would make reliance warranted if the process by which the relying side forms beliefs about the target from predictions of the kind q_h reports, from the system's producing them to their uptake, is in fact reliable in the domain of reliance, without the one who relies needing reasons for thinking it so. Such an account stands at the start of the ordering above, since nothing besides q_h need pass for the warrant to obtain; a record of performance is then what would let the relying side, or evaluators acting for it, establish that it obtains. Ortmann defends reliance on some opaque scientific systems on a condition that includes such establishment, holding that what matters is whether a system's reliability in a given context has been established by empirical evaluation that trustworthy researchers have performed adequately (Ortmann 2025, 1, 17–19). What this account requires of a system is a record of performance in the domain of reliance, assessed by evaluators independent of it; Chapter 11, Sections 5.2 and 5.3 set out Ortmann's account. Domain-bounded calibration, independent testing, and the sensitivity of the system's outputs to defeaters are diagnostics proposed here for what such a record would have to show.

These diagnostics meet a limit that Kitcher's analysis of how scientists assign authority to one another makes visible. As Goldman reports Kitcher's account, direct calibration compares an authority's claims with the assessor's independent opinions (Goldman 2007, §5). Strevens's reconstruction also identifies indirect calibration through others whose assessments can themselves be evaluated, and Kitcher's category of authority attached to pedigree or social position (Strevens 2010, §2, manuscript pp. 7–8). Carried over to artificial research, this limit means that where an artificial system's results go beyond what those who rely on it can assess independently, its calibration for them can only be indirect, through its record on claims they can check or through the judgments of others. How far direct calibration can reach also depends on whether those who rely keep the competence to form independent opinions; Chapter 3, Section 8 counts keeping up that competence and passing it on to new members among the capacities a continuing research system can lose. Connected with the progress of research, this suggests a further point. The further artificial inquiry goes beyond the objects and conditions that the relying side could assess independently, the less calibration obtained on known tasks can be carried over to the new domain, so that the ground for relying without translation may come to reach less far just as more can be learned from the machine. On accounts that require reliability to have been established, reliance beyond the reach of calibration needs another ground, though it does not follow that the ground must be a recovery of human understanding; the accounts below locate it in understanding, public criticism, or assurance by a party who stands behind the output, among other things.

On an understanding-centered account, where the achievement sought for the relying side is understanding of the target, of why the specified outcome occurs under the specified conditions, prediction is not enough: dependency relations, counterfactual structure, or explanatory use have to become available to the relevant human or collective subject, although theories of understanding disagree about which of these relations matters and how (Dellsén 2020). What such an account requires of a system is that it make some of the dependencies in the target on which its predictions rest available in a form the relying subject can use, the requirement Chapter 3 examined. Understanding how the system operates, or why its outputs may be relied on, is a different achievement, which Chapter 3, Section 4 distinguished from this one; a faithful trace of the system's computation does not by itself supply understanding of the target, and a simplified model that does not reproduce that computation can contribute to it (Chapter 3, Section 5.3).

On accounts that emphasize social justification, the burden shifts toward public criticism, uptake, and correction. Longino, for example, holds that because evidential relations are mediated by background assumptions, the only check against their arbitrary acceptance is critical interaction within and between scientific communities, and that the critical practices of a community must satisfy conditions of effectiveness to qualify as objective (Longino 2025, §5). What such an account requires of a system is that the reasons and the assumptions linking its data to its claims be exposed to criticism it does not control (Section 3.1; Chapter 11, Sections 3 and 9).

Assurance-based views, which Chapter 11, Section 5.2 introduces, ask in addition who intentionally assures the relying side of a claim and stands behind it, and, on versions that add a right of complaint, who answers for failure, so that on these views a system's output becomes the relying side's knowledge by testimony through a party who does so. Elizabeth Fricker also makes communicative intention and responsibility for what is told central, though from a local reductionist account of testimony, on which the reasons for believing testimony lie, speaker by speaker, in other reasons the hearer has, rather than an assurance view, and concludes that the outputs of current AI systems that generate natural language, large language models among them, are not testimony, while not ruling out that future systems' outputs might be (E. Fricker 2026, 2870–2871, 2877); Chapter 11, Section 5.1 sets out Fricker's account.

A pragmatic account locates coordination in shared objects and operations rather than in anything passing from one side's conceptual repertoire to the other's. Star and Griesemer, studying how amateurs and professionals cooperated in building a natural history museum, call objects such as specimens, field notes, and maps boundary objects: they can accommodate the local needs and constraints of different users, yet remain "robust enough to preserve a common identity at all sites," so that they can carry different meanings for different groups while remaining recognizable to each; alongside them they place standardized methods, which make information compatible by emphasizing "the how and not the what or why" (Star and Griesemer 1989, 87, 103). On such an account, shared instruments, interventions, and measurements can play a corresponding role even when the two sides' repertoires remain partly different. What it requires of a system is that its results be tied to objects and operations that the relying community can also use.

The distinctions among these accounts are analytical, not claims of causal or conceptual independence, as with the three questions of Section 2. If, for example, meaning were constituted by inferential role, grasping what a claim means and grasping what follows from it would be bound closely together; if common intervention and use were part of meaning, sharing them would already share part of meaning rather than substitute for it. The reliability-centered accounts above, by contrast, can separate warranted dependence from shared meaning altogether. The constructive question is what each position would require from an artificial research system and what kind of achievement it would license us to attribute.

Some positions bear on the ordering from outside it, by showing ways in which what passes can stand in for shared concepts. Humphreys, writing about instruments whose output is computed from detector data, takes the sinogram, the representation of the raw data of a computed tomography scan, and the familiar image computed from it by inverse Radon transforms to belong to one equivalence class of representations, fixed by a set of transformations on the data points: the image is useful to human observers because of their perceptual apparatus, while for a computer the sinogram is at least as useful, and what such images are about can be settled only with knowledge of the transformations performed (Humphreys n.d., 5–6). Humphreys does not discuss exchange between communities, but applied to it, equivalence under known transformations gives a way of translating that does not require shared concepts. Mitchell holds that the reliability of machine-learning tools rests, as for other instruments, on a theory of the instrument and on the stability of results across very different kinds of experiment (Mitchell 2020, 16–24); Mitchell writes about instruments used by human scientists, and taken as a model of coordination between communities the account is an application. Duede's account in Chapter 3, Section 5.3 bears on this ground: on it, deep-learning models do not, on the present state of interpretability research, reveal which principles their learned procedures operationalize, which is what a theory of such an instrument would have to supply. On these positions, a known transformation between equivalent representations or a theory of the instrument can stand in for shared concepts, though on Duede's account the latter is not yet available for deep-learning models.

Other positions widen or narrow what has to pass, or concern the direction in which it passes. Beyond the far end of the ordering, Weatherall and Wolfson hold for mathematics that a correct machine-generated proof is not by itself a contribution, since what contributes is mathematicians' engagement with it (Chapter 2, Section 4.5); on such a view what has to pass is not only information but participation. Conversely, what passes can shrink to a working technique: Takahashi holds that prediction without understanding is "incomplete as science but complete as technology" (Takahashi 2026, 185). In that case S gains a capacity to act, but on understanding-centered accounts not understanding. The passage can also run in one direction only. Jeannot forecasts that researchers outside advanced infrastructure will have their work ingested by systems they cannot access, replicate, or fully read (Chapter 12, Section 6.4); where passage runs only one way, one side supplies without receiving. Jeannot writes about research communities with and without advanced AI infrastructure.

The model of science as collective predictive coding proposed by Taniguchi and colleagues marks the boundary of this section's question. It treats human and artificial agents as members of one community that builds a single knowledge system (Taniguchi et al. 2025, §4.1), and so corresponds to a mixed S that forms one symbol system (Section 1.2) rather than to coordination between communities in this section's sense.

Routes that bypass translation were described by one of the founders of the incommensurability thesis and have recently been tried with an artificial system. Feyerabend, having argued that incommensurable theories may possess no comparable consequences, proposed to judge predictive success by agreement of behavior between a robot without sense organs, with the theory built into it, and beings trained to find their way about the domain, who, Feyerabend added, need not be human and could be other robots, a criterion that makes no reference to meanings (Feyerabend 1962, 94–95). Another route described in Section 3.1, learning the other side's kinds, has conditions of its own. On Kuhn's account, as Mladenović reports it, recognizing an object as being of a kind rests on perceiving, without inference, the relevant similarities and differences, a capacity acquired from particular examples in a trial-and-error process in which competent speakers support and correct the learner. Terms for unobservable kinds are learned otherwise: terms for mental constructs, such as money, through their relations to other constructs within a practice, and scientific singletons, terms such as mass and force that are outside genus–species taxonomies and have no set of contrasting terms, together with laws. On Kuhn's view, the referents of scientific singletons are never observable. The meaning of Newtonian force, for example, cannot be learned without Newton's second law (Mladenović 2022, xxx–xxxi, xxxiii–xxxv). Schut and colleagues report an attempt of the first sort. From AlphaZero, a chess system trained through self-play, they extracted concept vectors selected for being teachable to another network and for being better reconstructed from AlphaZero's games than from human games, and showed four grandmasters prototype positions for these concepts, with AlphaZero's continuations but without labels or verbal descriptions. All four solved more unseen puzzles afterwards, with improvement on eight of the thirteen concepts tested; a puzzle counted as solved when the grandmaster found the move AlphaZero had selected. The authors present the result as a proof of concept given the small sample and the possibility of priming, and they chose chess partly because the usefulness of a concept is easier to evaluate there than at the frontiers of science (Schut et al. 2025, §§1.2, 3.2, 4.1). The case can be read as showing people acquiring part of a machine-native organization without translating it, where its instances can be exhibited and success can be scored, even if, as in this study, by agreement with the machine's own choices; it does not show what such a route would require where p_a concerns kinds or features that cannot be exhibited, or where S cannot score the result itself.

What has to pass, then, depends on the account: a record of performance assessed independently, within the limit that Kitcher's analysis marks; some of the target's dependencies in a form the relying side can use; reasons and assumptions exposed to criticism the system does not control; a party who stands behind the output; shared objects and operations; or, where instances can be exhibited, learning the other side's kinds, among other things. On each of these accounts, where what it requires does pass and the system's answerability to its targets holds (Sections 2.3 and 3.3), S can gain something through research that does not align with it, though whether what it gains is knowledge, understanding, or a capacity to act differs from account to account. The differences become plain when four of the accounts examined since are applied to the case in which q_h was introduced at the start of this subsection. Let that case, in which humans gain q_h without acquiring, translating, or understanding p_a, be specified further: the system's predictions survive independent testing, but it cannot explain their reasons to humans, and no party stands behind its outputs. On reliability-centered accounts, relying on the predictions can yield knowledge about the target insofar as they are in fact reliable in the domain of reliance. On understanding-centered accounts, the same success does not yield understanding of the target so long as the target's dependencies have not been made available by other means. On accounts that stress social justification, the claims are not acceptable so long as their reasons and assumptions have not been exposed to criticism. On assurance-based views, with no party standing behind the outputs, they do not become the relying side's knowledge by testimony; whether they can still yield knowledge as the readings of an instrument is not settled by these views alone (Chapter 11, Section 5.1). What separates the four accounts is not whether the prediction succeeded but which achievement, and by which route, they count as realized on the receiving side.

Krenn and Champion propose translating epistemic aims into computable, nonanthropocentric objectives (Chapter 7, Section 4.3); nonanthropocentric there means applying to biological and artificial scientists alike. Their program itself raises a question that its authors pose. They anticipate AI-discovered improvements whose principles humans can confirm but not fully comprehend, propose teacher systems that would explain such solutions to people, and expect humans to remain the strategic leaders who set directions and allocate resources, while asking how humans could verify and interpret discoveries they cannot genuinely understand (Krenn and Champion 2026, 336, 343–346). Thais and colleagues, by contrast, emphasize human understanding and the institutional conditions that sustain scientific criticism and correction (Thais et al. 2026, 1, 3, 5, 8). Together they return to the tension with which Section 3 began, between the possibility that artificial difference opens and the risks it carries: how can artificial difference be preserved without assuming that whatever emerges will remain knowable by humans, socially criticizable, or safely governable?

4. Two horizons and what Part IV leaves

What construction horizons follow from the three cases of Section 1, in which S is a human, a mixed, or an artificial society, and what does Part IV leave to Part V? Since what a program for building artificial researchers must answer changes with whose knowledge it seeks, the questions can be grouped by S. From the division by S, two horizons can be drawn, named here after the first and third cases: the horizon of research for a human society and the horizon of research by and for an artificial society, or, for short, the human-society horizon and the artificial-society horizon. The mixed case falls wholly under neither (see below). The artificial-society horizon takes as its central case one in which the actor is artificial as well, although, as Section 1.3 showed, questions of artificial belief and possession arise from taking an artificial society as S even where the actor is not artificial. The human-society horizon asks how artificial systems can bring about knowledge new to human societies, how that inquiry stays directed at those societies' unknowns as the selection of questions and standards moves inside the system (Section 3.2), how those societies keep the competence to evaluate what the systems produce and can tell whether their own standards have moved with the systems (Section 3.2), and how the resulting knowledge enters their epistemic situations, while what the systems still depend on, what they themselves do not know, and how far human involvement extends remain visible (Section 1.1). It can advance through partial capacities and hybrid organization, and it does not require artificial consciousness, a wholly independent epistemology, or an artificial society. The artificial-society horizon asks whether artificial beings and communities can form, maintain, criticize, and transform epistemic worlds of their own (Section 1.3). It raises stronger questions about artificial belief and social knowledge (Section 1.3), world-answerability (Section 2.3), purposes the society sets for itself and the revision of its own governing standards (Chapter 7, Section 4.3), and coordination with human epistemic societies (Section 3.4). For such research, human society stands not as the reference society but as a party affected by it, so that what can be asked on its behalf is a question of acceptability and legitimacy (Section 3.2). Part V adds the question of what becomes of the relations through which research is directed, certified, and governed when those who hold them are themselves artificial. The present chapter does not show that such a society is possible or impossible. It tries to show why building research by and for an artificial society asks for more than removing humans from the steps of research.

Research for a mixed society falls wholly under neither horizon (Section 1.2). For its human members, the questions of the human-society horizon remain. To the extent that artificial members are to count among those to whom its knowledge is attributable, the questions of artificial belief and possession that belong to the artificial-society horizon arise. Where its human and artificial members form separately organized communities and its knowledge is to be shared between them, the questions of translation and coordination of Section 3.4 are added. Where its members form a single symbol system, it is further asked whose experiences that system integrates. Who speaks for such a society can itself divide its human and artificial members (Section 3.3). Chapter 13, Section 5 retraces the chapters' examination of these horizons.

The distinction between the horizons is not a timetable that postpones every fundamental question. Work on artificial grounding, understanding, social organization, and norm revision can improve systems built for human knowledge now. Nor is the artificial-society horizon a claim that independence is intrinsically better. It is a way of taking literally the possibility that research could become artificial at the level of its epistemic organization rather than only at the level of its visible operations.

Part IV has therefore separated two choices: where inquiry is conducted and whose epistemic life it extends. Human and hybrid research can advance without deciding whether an artifact literally knows. Attributing knowledge to the artifact itself instead makes the content of its states, the presuppositions of its checks, and its answerability to the world central questions (Sections 1–2). What becomes central when understanding is attributed to the artifact depends on the accounts of who understands distinguished in Chapter 3, Section 6. An artificial inquirer can differ from S's members, and the worry this raises is structurally similar to what is known as the AI alignment problem, on the strand of discussion that locates the difficulty of alignment in the system's otherness (Section 3.2). Allowing the artifact to revise questions and standards can enlarge what it investigates. It also permits a wider mismatch with S in what is taken to be unknown, worth pursuing, or a good reason, among other things; such a mismatch can open as well at the bearer and from the side of S, and a commission does not by itself prevent it. Applied to research, de Blanc's and Soares's arguments imply that, because inquiry can revise the concepts in which a commission was written, a gap can open when inquiry succeeds as well as when it fails; and Russell's argument, extended to research, implies that where S's standards move with A, agreement between them does not by itself show that A still serves what S would otherwise have sought (Section 3.2). Neither inherited human norms nor a different epistemic organization settles in advance whether a given mismatch is drift, error, or gain; that depends on whether it is assessed from S as it now is, from S as it may come to be, or, where correctness is in question, from the world (Section 3). Where research by an inquirer of another kind does not align with S, what follows depends on which relation fails. Even where relevance to S fails, answerability to the target can remain an anchor that does not depend on S's consent, on accounts on which what corrects a commitment is fixed by its target and, where the target's verdicts reach the inquirer only through S's procedures for judging claims, provided those procedures can be revised by evidence (Section 3.3). S may then still gain knowledge or understanding through what such an inquirer reaches. What has to pass for that, and whether what is gained is knowledge, understanding, or a capacity to act, differ by account. Knowing that the machine predicts a specified outcome (q_h) is not knowing that the outcome will obtain; reliability-centered accounts locate the step from there to knowledge of the target in the actual reliability of the predictions, while other accounts locate it in understanding, criticism, or assurance, among other things. And where instances can be exhibited and success can be scored, even if only by agreement with the machine's own choices, people can learn part of a machine's kinds without translation (Section 3.4). These possibilities give the construction project substance: it can change who conducts inquiry, who knows, and how differently organized inquirers learn from one another. Which possibilities are realizable remains conditional on the accounts and capacities examined here.

Part V takes up the relations on which the second of the two choices above, whose epistemic life inquiry extends, depends. Section 3 argued that, because an inquirer of another kind can differ from S's members, neither a commission nor human control secures that research stays directed at S's unknowns, that whether it does is a question about particular configurations, and that the epistemic success of research for S and the legitimacy of what directs it are different judgments, the second of which is asked also on behalf of those affected outside S (Sections 3.2 and 3.3). Which of S's unknowns research pursues, whether its results come to count for S, and on whose authority its direction is set involve relations that extend beyond any single actor, artificial or human. The next part asks which of them can be located inside an artificial configuration, and what changes when they are; whether research succeeds for S and whether its direction is legitimate to the parties it affects are then asked separately.