Part II — What Would Count as Epistemic Achievement?

Chapter 2 — When Does a Result Become Knowledge?

Epistemic Change, Achievement, and Knowledge

1. From a novel result to an epistemic achievement

The working characterization describes research as an attempt to bring about epistemic achievements new to a society S. This chapter examines one demanding form of success: knowledge. Novelty can belong to an unsupported conjecture, an unread inscription, an accidental truth, or an artifact whose significance is unsettled. None is knowledge merely because it is new or useful. The general orientation does not relax the conditions of this particular achievement.

The knowledge term keeps at least three questions visible. In what respect does the result get its object right? To what subject can the resulting epistemic condition be attributed? What distinguishes an epistemically successful connection to the object from a fortunate accident? These are burdens raised by the knowledge term, not three independently necessary components of a complete analysis. Different epistemologies organize and answer them differently.

The chapter proceeds in this order. Section 2 asks what kind of attitude is at issue, chiefly belief, and whether its bearer must be the producer of the result. Section 3 takes up truth and fallibility, and Section 4 what makes a true success epistemic rather than accidental. Section 5 asks what verification contributes. Section 6 separates what can be social about knowledge: the reference for novelty, production, possession, and retention. Section 7 tests the preceding accounts of belief, verification, and social possession together on a case in which the production of a result and its possession by S come apart. Section 8 turns to epistemic change before and around knowledge, and Section 9 states what each account would require of artificial research and on whom each requirement falls.

This chapter begins with propositional knowledge because its analysis is well developed and provides a workable analytical entry point. The choice of entry point is not a requirement of the working characterization, which does not require knowledge to be propositional; whether knowledge includes abilities or understanding remains the open question registered in Chapter 1, Section 5. Nor does the chapter assume that every research achievement is exhausted by knowing that some proposition is true. A proof, a concept, an interpretation, a model, a practical ability, or an artifact can change what a society can represent, understand, infer, and do. Part of that contribution can be propositionally expressed: a construction is possible; a procedure yields a result under specified conditions; a new concept organizes previously disconnected phenomena in a certain way. Propositional content is not restricted to newly observed natural facts (Thurston 1994). In the history of Euler's formula for polyhedra, for example, the result, that the number of vertices minus the number of edges plus the number of faces is two, was stated by Euler in terms of concepts for the parts of a solid, one of which he first had to name: what are now called edges had, he wrote, no proper name, so he called them acies (Sandifer 2004, 2–3). What the result expressed was not an observed natural fact but a relation among the numbers of a solid's parts. The example thus illustrates, as noted above, that what comes from working out concepts can also be expressed as a proposition. It does not show that the proposition has been established as knowledge, since, as Musgrave and Pigden summarize Lakatos's view in their encyclopedia entry, a conjecture of informal mathematics and its proof are improved together through criticism: counterexamples bring out hidden lemmas and so improve the proof, proofs bring out conditions of validity and so improve the conjecture, and a proof can deserve respect without being flawless (Musgrave and Pigden 2026, §§2.1, 3.1). More generally, that a contribution can be stated as a proposition does not mean that the proposition exhausts the achievement. That a proposition can be written about an output, that the output supports it, that a subject knows it, and that the output itself is knowledge are different claims. Nor does it follow that anything epistemically valuable is knowledge.

In this way the chapter does not reduce every epistemic achievement to knowledge: knowledge is not the name of every valuable process change or achievement, and understanding, ability, conceptual reorganization, empirical adequacy, and recognized ignorance remain assessable in their own terms. Knowledge nonetheless remains a demanding target success, because it forces questions about truth, attribution, and non-accidentality that something new or epistemic improvement can leave open. The chapter determines what knowledge would require under different epistemological accounts, how it relates to other achievements, and what each account makes visible in artificial research.

2. Process states: what sort of epistemic attitude is at issue?

In the central philosophical use relevant to propositional knowledge, to believe that p is roughly to take p to be the case. This introductory gloss is much thinner than a theory of belief. It does not choose among the main accounts of what belief fundamentally is: representational accounts, on which it is a stored representation; dispositional accounts, on which it is a disposition expressed across thought and action; functional accounts, on which it is a state fixed by its role among perception, other states, and action; interpretive accounts, on which it is a state attributable to a subject interpreted as rational; and normative accounts, on which it is a state constitutively subject to norms of truth and rationality (Schwitzgebel 2023). Contemporary discussions also dispute whether belief is necessary for every case of propositional knowledge, even though the belief condition belongs to the traditional analysis with whose critique, as Ichikawa and Steup note, much of the twentieth-century literature on the analysis of knowledge began (Ichikawa and Steup 2026; Myers-Schulz and Schwitzgebel 2013).

Several neighboring attitudes matter because research does not consist only in settled belief. A proposition can be supposed temporarily in order to explore its consequences. A model can be accepted for a task even when its idealizations are not believed literally. Confidence can be graded rather than categorical. Judgment can name an act of affirmation, while belief names a state that persists. On some accounts, suspension is an attitude toward a considered question, not merely the absence of any representation. And a candidate can be entertained without being believed. A historian, for example, describes keeping, beside notes summarizing each archival file, a separate "Idea File" of free association and "uncensored writing" that might never appear in a finished text, in which the historian also records surprising details and significant silences in the documents (Walkowitz 2009).

These distinctions prevent an apparently unified research pipeline from changing the meaning of its central state at each stage. Generating a hypothesis does not require believing it. Comparing mutually incompatible explanations does not require believing them all. Assigning a probability does not automatically create a categorical belief above an agreed threshold. Institutional acceptance can govern collective action even when individual members assign different credences. A new concept can alter the space of candidates rather than merely shift confidence among a fixed set.

For this reason, where later chapters describe the course of research as a whole, they speak more broadly of epistemic attitudes, or of commitments, as Chapter 6 does when it asks which commitment a result gives reason to revise. These words do not replace belief in the analysis of propositional knowledge. They keep belief from being used as the name of every representation, confidence, assumption, policy, or public status that changes during research.

2.1 The bearer of belief need not be the producer of the result

The actor or organized system that conducts research, A, need not be the subject to which the resulting knowledge is attributed. The society S supplies the reference for novelty and identifies for whom the knowledge is sought; it is not thereby the knowledge subject. An artificial system can derive a theorem, execute an experiment, or identify a pattern that a human community subsequently evaluates and comes to know. If a belief condition is adopted, the condition applies to the subject claimed to know. It does not automatically require the artificial producer to believe the proposition.

The distinction becomes unavoidable when knowledge is attributed to an artificial agent or artificial society itself. Consider increasingly integrated relations to a proposition p: a system stores a sentence expressing p; it outputs that sentence when prompted; it assigns p high confidence; it uses p in inference and planning; it changes action because of p; and it revises p when contrary evidence appears. On the functional and dispositional accounts listed at the start of Section 2, which identify belief by what it does in inference and action, storing or uttering a sentence is only weak evidence of belief, and stable integration with inference, prediction, action, surprise, and revision is stronger evidence. Interpretive and normative accounts may ask for something of a different kind, such as a subject that can be interpreted as rational or that answers to norms, and so may require a different kind of representation, agency, or mentality. In either case, evidence of storage and integration bears more directly on whether a state plays the role of a belief than on what the state is about; whether an artificial system's states are about the targets that p concerns is taken up in Chapter 8, Section 2.2.

This is not permission to call the same system a believer or non-believer at will. The system's architecture and behavior provide the evidence; a theory of what belief consists in determines how that evidence bears on literal attribution. This book can therefore describe the functional facts first and state the additional account under which they constitute belief. Whether a system that holds p in this integrated way is responsible for so holding it, or holds it for reasons, is a further question, which Chapter 7, Section 3.3 takes up in asking what thicker attributions to an artificial actor would require.

Whether knowledge can be attributed to an artificial society depends on theoretical choices at two levels. The first is whether knowledge requires belief at all. Ichikawa and Steup describe the belief condition as only slightly more controversial than the truth condition, and report that a few philosophers have argued that knowledge without belief is possible. In Radford's example, which they present, someone asked about English history gives date after date correctly while taking the answers to be mere guesses. On Radford's account, this person does not believe the answers, having no confidence in them; yet the run of correct answers shows that they learned the facts and have not forgotten them, so they know them. Both claims are disputed: some reply that the person has an implicit belief, others that answers held with so little confidence do not express knowledge (Ichikawa and Steup 2026, §1.2). Knowledge-first epistemology is not among these denials. It does not construct knowledge from belief plus additional ingredients, but it rejects that order of analysis rather than the belief condition: it is consistent with the traditional view that knowing entails believing (Ichikawa and Steup 2026). The second level arises among views that do require belief: they differ over what belief consists in, as the accounts set out at the beginning of Section 2 show. Thus, if an artificial society is claimed to know p, either that society must satisfy the selected belief condition or the knowledge attribution must be licensed by an account that does not require belief in that form. The issue is about the subject of knowledge, not a universal mental prerequisite imposed on every tool or participant in research.

3. Truth and fallibility

For its central use of propositional knowledge, this chapter retains factivity: if a subject X knows that p, then p is true. This is a substantive working commitment, but a comparatively modest one. Ichikawa and Steup report that most epistemologists agree that what is false cannot be known, and that most treat ordinary speech that describes a once-confident but false view as what people “knew” at the time as a kind of exaggeration rather than as literally true (Ichikawa and Steup 2026).

Factivity does not turn truth into an operational oracle. Whether p is true is different from whether a researcher can prove, conclusively recognize, or become certain that it is true. A coin can land heads even when nobody observes it. A claim can be true even when the available evidence is misleading. Conversely, a community can apply its best methods responsibly to a false claim and remain unable to detect the error. Truth concerns how things are; evidential and institutional procedures concern the fallible routes through which a subject becomes related to how things are.

This separates two evaluations. At time t, investigators may be warranted in accepting p under the best available evidence and procedures. Retrospectively, if p is false, they did not possess factive propositional knowledge that p, even if they did nothing blameworthy. The first evaluation concerns what the participants were entitled or rationally permitted to believe. The second concerns whether the epistemic attempt reached its factive target.

Research institutions can design methods, gather evidence, organize criticism, estimate reliability, and propagate correction. They cannot observe a separate truth label attached to every conclusion. An artificial system is in the same position. It can be answerable to proofs, observations, interventions, failed predictions, replications, material breakdowns, and criticism without possessing an infallible test for final truth. Each of these can push back against the system's expectations and claims from outside them; call them forms of resistance. They are mediated by instruments, models, concepts, and background assumptions, but a system whose commitments can be revised by them is related to its object differently from one optimized only for textual plausibility or social approval (Chapter 8, Section 2.3 takes up this relation as world-answerability). Criticism, the last of those forms, and social approval are both responses from others, but criticism can carry reasons and the resistance of the world, while social approval can be obtained without carrying either.

4. What makes a true success epistemic rather than accidental?

Gettier cases, as Ichikawa and Steup set them out in their encyclopedia entry on the analysis of knowledge, provide a canonical illustration of the general problem: a subject can reach a supported true belief whose truth nevertheless depends on an epistemically inappropriate coincidence (Ichikawa and Steup 2026, §3). The cases do not determine one replacement analysis of knowledge. They make visible why a true result and some support for it may still fail to constitute the demanding epistemic success at issue here.

A random generator can emit a true proposition. A contaminated benchmark can make an incapable system appear accurate. A researcher can infer a true conclusion from a false intermediate claim. A valid-looking explanation can be unrelated to the mechanism responsible for a correct prediction. These cases motivate a thin commitment:

For the demanding sense of knowledge used as the central target here, accidental truth is insufficient. The truth of the relevant attitude must stand in some appropriate epistemic relation to its grounds, method, competence, environment, or other truth-connected conditions.

The problem of accidental truth is not confined to individual believers or to invented cases. For example, a proof of the four-color theorem that Alfred Kempe published in 1879 appears to have been accepted for more than ten years, until Percy Heawood showed in 1890 that a step in its central argument fails for some maps. The theorem itself was proved only in 1976, by a different route (O'Connor and Robertson 1996). On accounts that require a true belief to be non-accidentally connected to its grounds, those who accepted the theorem on the strength of Kempe's proof held a true belief that was not knowledge on the grounds they relied on.

The thin commitment above says only that the truth must stand in some appropriate epistemic relation to its grounds or method; it does not say which relation. The phrase marks a problem rather than solving it. In the traditional analysis of knowledge as justified true belief (Chapter 1, Section 5), this place is held by the justification condition, which some epistemologists also describe as a belief's being rational, warranted, or evidentially well supported (Ichikawa and Steup 2026, §1.3.1). Among responses to counterexamples to that analysis, some accounts keep the justification condition and add a further condition, such as sensitivity or safety (conditions that test whether a belief was right by accident by considering situations slightly different from the actual one; Section 4.1), while others replace it, for example with a reliability condition on knowledge (Ichikawa and Steup 2026, §§3–6). One way to fill the place, then, would be to call whatever relation is appropriate justification. But calling every candidate relation justification would conceal rather than manage the disagreement over which relation is required. Alston argues that something like this has happened in the theory of justification. As Feldman reports, Alston explains the widely different accounts that philosophers give of justification, in terms of what one ought to believe, good evidence, reliability, proper function, and other notions, by holding that no objective property of beliefs is picked out by justified; there is instead a variety of epistemic desiderata, such as having adequate evidence, being based on it, and being formed by a reliable process, whose nature, viability, importance, and interrelations Alston examines in turn (Feldman 2005; Alston 2005, xi–xii). This chapter, for its part, keeps knowledge as a demanding target, while it treats the relations that might make a true attitude non-accidental as distinct desiderata whose weight depends on the account adopted (Section 4.1).

Some philosophers recognize a lightweight use of knowledge requiring little more than true belief, while reserving a stronger sense for the central epistemological questions (Ichikawa and Steup 2026). The thin non-accidentality commitment selects the stronger sense for the knowledge target examined here. It does not establish that lightweight knowledge attributions are linguistically illegitimate. The stronger target is useful because a system that merely increases the frequency of lucky true outputs leaves precisely the capability and validation problems that motivate the inquiry.

The subsections proceed as follows. Section 4.1 surveys rival approaches to what the appropriate epistemic relation requires, and Section 4.2 sets out, in Table 2.1, what follows for artificial research on each. Section 4.3 takes up basing, a relation asked about within most of those approaches, and Section 4.4 the question it presupposes, what evidence a system has. Section 4.5 uses the case of an accidental generator to separate generating, knowing, contributing, and conducting research.

4.1 Several ways of locating epistemic success

Once knowledge is selected as a demanding success term, no theory-neutral decomposition follows. Epistemology, which has long debated what knowledge requires, offers rival approaches to that question: approaches to what, beyond a true belief, makes a success knowledge rather than luck. They organize the relation among a subject, a content, its grounds, the world, and a wider epistemic environment in materially different ways. They therefore make different features of research visible and imply different judgments about artificial researchers.

Internalist and access-oriented approaches ask what reasons, evidence, or reflective access the proposed knower possesses (Ichikawa and Steup 2026, §1.3.2). Evidentialist approaches ask whether an attitude fits the evidence the subject has. On Feldman and Conee's evidentialism, for example, as Mittag sets it out in the Internet Encyclopedia of Philosophy, whether a belief, a disbelief, or a suspension of judgment is justified turns entirely on its fit with the evidence the subject has at the time, however that evidence was acquired (Mittag n.d., §§2, 3c). Whether norms on acquiring evidence belong to epistemic assessment at all is disputed (Section 5). Externalist and reliabilist approaches can instead locate the relevant standing partly in the causal history and truth-conduciveness of the process, even when the subject cannot fully articulate why the process succeeds (Goldman and Beddor 2021, §§1.1, 2.1). The contrast matters for an opaque artificial system. For such a system, the reliability of the process that produced a result, the reasons a human recipient has for it, and the recipient's entitlement to rely on it can come apart: a reliable process may leave the recipient with no articulable reasons, and a recipient may have reasons for a result that an unreliable process produced. Because they can come apart, which approach one takes matters: a reliabilist looks to whether the process is reliable, an internalist to whether the recipient has reasons.

Virtue-reliabilist and achievement approaches ask whether a true success manifests a stable epistemic ability attributable to the relevant subject. Virtue-responsibilist approaches instead take intellectual virtues to include cultivated character traits such as open-mindedness and conscientiousness, and some ask whether virtue requires the corresponding motivations (Turri, Alfano, and Greco 2021). Where the knowledge of an artificial system itself is at issue, this raises the further question whether anything in the system plays the role of epistemic motivation. Modal approaches ask whether the attitude could too easily have been false (safety) or whether it would still have been held had it been false (sensitivity) (Ichikawa and Steup 2026, §5). Defeat-sensitive accounts ask how contrary information, whether information supporting the denial of the claim (rebutting information) or information casting doubt on the link between the claim and its grounds (undercutting information), or higher-order information, which casts doubt on the reliability of one's own judgment, changes its standing (Section 5). Within most of these approaches a further question arises, which Section 4.3 takes up as the question of basing: whether the subject holds a commitment because of the support that purportedly favors it, rather than merely possessing that support alongside a conclusion reached by another route.

All the approaches surveyed so far, from internalism to defeat-sensitive accounts, separate a correct output from a non-accidental achievement. They need not, however, draw the boundary of the successful subject in the same place: the success may be credited to the artificial system, to the human or other artificial system that uses it, or to a whole that includes both.

Knowledge-first epistemology changes the explanatory order rather than adding another condition. Williamson takes knowledge as the unexplained starting point and explains other epistemic notions partly through it: to believe p is roughly to treat p as if one knew it, justification is primarily a status that knowledge can confer, one's evidence is one's knowledge, and one must assert only what one knows (Williamson 2000, 9–10, 46–47, 185, 243). Taking knowledge as primitive does not make the conditions distinguished in this section idle for evaluation. Williamson gives two reasons for treating knowing as unanalyzable. On Williamson's account, knowing is itself a mental state, whereas every standard analysis includes truth, a condition that is not mental; and the way analyses since Gettier's counterexamples have each been met, as they grew more complex, by more complex counterexamples is what this view would lead one to expect. Williamson still holds that knowledge requires safety and does not require sensitivity. The reason for dropping sensitivity can be seen in an example. Suppose that someone who slightly underestimates the distances they see judges, of a mark one meter above the water, that it is less than twenty meters above the water. The belief is insensitive: had the mark not been less than twenty meters up, it might have been just over twenty meters up, and the subject would still have judged it to be less. Yet the subject knows it, since one meter and twenty-one meters look quite different, so the belief rests on evidence quite unlike the evidence the subject would have relied on had the proposition been false. However slight a systematic bias, so long as it is not zero, the same mechanism can make almost all of a subject's beliefs in contingent propositions insensitive. On Williamson's view, moreover, a necessary condition of knowing need not be a conjunct of a non-circular analysis (Williamson 2000, 30–31, 33, 147, 159–161, 307–310). On this account, then, safety is not a component from which knowledge is defined, but it is still a necessary condition. Evidence about safety is a further matter, used in evaluating an attribution: evidence that a result could easily have been wrong tells against it, while evidence of safety does not by itself establish knowledge. On a knowledge-first account, Table 2.1 in Section 4.2 can be read in this way, as stating what the adopted account takes knowledge to require and what must be established to support an attribution, not the parts of a definition.

Simion's naturalistic version organizes the epistemic domain around one function, generating knowledge. The function is etiological: inquiry has it because its past production of knowledge brought epistemic benefit, and that benefit helps explain why the practice persists; such a history can be supplied by ongoing selection or a learning process as well as by selection over generations (Simion 2025, 7–8). Applied to trained artificial systems, the account implies that whether such a system has that function turns partly on what its training and selection rewarded, not only on how often its present outputs are true. Carried further, the application implies that two systems with the same present rate of correct answers can differ in whether knowledge is attributed to them on this account, if the training and selection of one rewarded correct answers and those of the other rewarded plausible answers or the approval of raters. Taking the stated objective of training to be the etiological function is, however, itself a premise, since what selection actually rewarded can diverge from the stated objective. Bird's extension of a naturalized knowledge-first functionalism to groups is discussed in Section 6.

Social epistemologies introduce further alternatives (Section 6): epistemic standing may depend on testimony, divided expertise, organized memory, critical interaction, or functions realized by a group. Inquiry- and understanding-centered approaches (Chapter 1, Section 3; Chapter 3) also question whether the epistemic change relevant to research is exhausted by the acquisition of settled propositional knowledge.

These approaches stand in different relations to justification. Internalist and evidentialist approaches, and process reliabilism, are theories of what makes a belief justified, and the dispute between internalism and externalism is a dispute about justification (Ichikawa and Steup 2026, §1.3.2). Sensitivity and safety are proposed as further modal conditions on knowledge (Ichikawa and Steup 2026, §5), and reliabilist theories of knowledge put a reliability condition in place of the justification condition (Ichikawa and Steup 2026, §6.1). On virtue-theoretic approaches such as Sosa's, knowledge is apt belief, belief that is true because it was produced adroitly. Knowledge therefore entails justification, but that justification is understood as adroitness, the belief's being produced by skill; and because what is required is that the skill explain the success, justification is not a component separable from a success's being creditable to competence (Ichikawa and Steup 2026, §10). Basing concerns the relation that separates a belief that merely has support available from one held on that support (Section 4.3). Knowledge-first accounts, as noted above, explain justification through knowledge. The thin commitment at the start of Section 4 does not choose among these: it leaves open whether the appropriate epistemic relation is justification on one of these accounts, justification together with a further condition, or a condition that replaces it.

These are not modules to be accumulated into one maximal definition. Nor does this book need to select one account before any further analysis is possible. They can be used conditionally. If accessible reasons are required, one set of opacity and transfer problems follows: the reasons for which a system reached its conclusion may be invisible to the recipient, and it is unclear whether they carry over as the recipient's own reasons. If reliable external connection is sufficient, another follows. If competence, modal robustness, understanding, or social organization is central, the location and meaning of success change again. The philosophical positions are therefore resources for tracing the consequences of locating epistemic success in different places, not different names for one hidden pipeline. What each would require, and what apparent success each is liable to mistake for knowledge, is set out approach by approach in Section 4.2 (Table 2.1); Section 9 draws these together as requirements on artificial research.

4.2 Conditional consequences

The approaches distinguished in Section 4.1 are rival accounts of what knowledge requires. Table 2.1 sets out, for each of them, what follows if that approach is adopted. The left column names each approach by what it emphasizes. The middle column states what knowledge would then require and what an evaluator would need to establish to support attributing it. The right column gives a characteristic false positive: an apparent success that the approach would not count as knowledge but that an evaluation could easily mistake for it. Read row by row, the table shows how the choice of approach changes what must be built and checked in an artificial research system. It is a simplified guide rather than an exact statement of each account: each row compresses a family of accounts into one representative requirement, and each false positive is one example among others. The paragraphs after the table show, for reliabilism and virtue approaches, how accounts within a family differ. The false positives take as their example the case in which the system itself is claimed to know. When a human or a community comes to know through the system, each requirement in the middle column applies to the belief of that human or community, and the system's output bears on it as evidence or as part of the process that formed it; a post-hoc rationale, for instance, can then be a reason that the human really has, if the argument it gives is sound (Section 4.3).

Table 2.1. What follows from each epistemological approach: a simplified overview

If the adopted account emphasizes...	Then what knowledge requires, and what must be established to support attributing it, is...	A characteristic false positive
Accessible reasons	reasons available to and usable by the relevant subject	a persuasive post-hoc rationale unrelated to the operative route
Evidential fit	an attitude proportioned to the evidence possessed by the subject	high confidence produced by selective retrieval despite contrary evidence
Process reliability	truth-conducive formation or maintenance of the subject's belief in the relevant environment	one benchmark score generalized beyond its task distribution
Modal safety or sensitivity	a belief, held by the subject claimed to know, that would not have been false in the nearby possibilities that matter, or that would not have been held had the claim been false	repeated success under near-identical or contaminated conditions
Competence (virtue-reliabilist)	success because of a stable epistemic ability attributable to the subject or system	a lucky run or a pipeline whose decisive selection was performed by hidden humans
Intellectual character (virtue-responsibilist)	inquiry that expresses traits such as open-mindedness and conscientiousness, and whatever plays the role of their motivation	hedging and self-critique reproduced in reports without governing how the inquiry is conducted
Defeat sensitivity	appropriate response to rebutting and undercutting information	a corrected user-facing report while the operative commitment remains unchanged
Knowledge-first functionalism	proper operation of a capacity or organized system whose function is generating knowledge	a conventional checklist that decomposes knowledge into a set of conditions, renamed "knowledge-first" without independent account of the function

The table leaves out three things. Basing is not a rival emphasis but a relation between a belief and its grounds that most of the approaches require in some form, so Section 4.3 takes it up before any choice among them. For that reason the rows for accessible reasons and evidential fit state only that the reasons or evidence are available and that the attitude fits them; that the belief be held because of them is added in Section 4.3. Social accounts have their own table (Section 6; Table 2.2), and inquiry- and understanding-centered approaches are examined in Chapter 3. The requirements in the middle column hold only if the corresponding approach is adopted; they are not to be added up into a single list of criteria that every artificial researcher must meet.

The reliability row shows what such a consequence looks like. Reliabilism can permit knowledge without reflective access to the reliability-producing factors (Goldman and Beddor 2021, §§1.1, 2.1). Reliability belongs to process types, however, and one token process instantiates many types with different reliabilities. Someone who reads a distant clock in a dim room, for example, forms the belief by vision, by vision in dim light, and by reading small, distant figures in dim light, and these types are reliable to different degrees; which type fixes the verdict is the generality problem, to which reliabilists have proposed several responses (Goldman and Beddor 2021, §2.3). For a belief formed through an artificial system, whether it is the system's own commitment or a human's belief that relies on the system's output, the generality problem means that a reliabilist verdict depends on how the producing process is typed: the model alone, the model with its retrieval and human filtering, or the pipeline on a given task distribution. If hidden human filtering removes the model's errors, the whole may be reliable while the model alone is not. Choosing a benchmark is in part choosing such a type, and a benchmark score shows the reliability of one type on one task distribution, not of the system in general; Otsuka notes that in practice AI models are justified case by case through benchmark testing, so that their reliability has become attached to particular models (Otsuka 2026, "Understanding the Machine Virtues"). A claim that an artificial system is reliable therefore needs to state its unit: whether human filtering lies inside it, and on which distribution it was assessed.

Statistical testing supplies a worked instance of the generality problem. Otsuka argues that the stopping-rule problem, in which a single dataset leads to rejecting a null hypothesis under one experimental design and to retaining it under another, is a version of that problem. In Otsuka's example, a coin tossed twenty times comes up heads six times. Under a plan to toss it twenty times, the null hypothesis that the coin is fair cannot be rejected at the 5 percent significance level; under a plan to toss until the sixth head, it can, because the probability of a result at least as extreme as the one observed is 0.0577 under the first plan and 0.0318 under the second. The same sequence of tosses is thus an instance of two testing processes with different verdicts. One frequentist response Otsuka discusses accepts that the same experiment can be described as different tests and requires the experimenter to specify the testing process in advance and to follow it; a verdict is then justified only relative to that process, and switching to another design after seeing the data counts as hacking the test (Otsuka 2023, 99–103). Extended by analogy to an artificial system, the requirement would fix the process type under evaluation before results are inspected and would treat later re-typing as a reportable change rather than a neutral redescription.

Other approaches make other human interventions and evaluation failures salient. Virtue approaches place weight on a success that manifests competence; in Sosa's version, as John Turri, Mark Alfano, and John Greco present it, a success is apt when its correctness, and not merely its occurrence, is attributable to a competence exercised in conditions appropriate for it (Turri, Alfano, and Greco 2021, §5). Modal approaches ask whether the belief could too easily have been false or would have been held even if false. Sosa separates aptness from safety: neither, he holds, entails the other. On this account a performance can be apt though unsafe, when the performer's competence or the conditions for its exercise might easily have failed, and safe though not apt, when its success is secured without being owed to the performer's competence (Sosa 2005, §§8–9). Sosa also holds, as Turri, Alfano, and Greco report, that the competence to which a success is owed may be socially seated, that is, located in a social group rather than in an individual. The correctness of a belief received through testimony, on this account, is owed in small part to the recipient's competence to trust testimony and more fully to a competence seated collectively in those along the chain, as a touchdown pass manifests the quarterback's competence but more fully the team's; the recipient earns partial credit, which Sosa argues can suffice for knowledge (Turri, Alfano, and Greco 2021, §7). Seen through Sosa's distinction between aptness and safety, hidden human selection that corrects a model's errors can make the outputs safe without making their success creditable to the model. Whether a success is creditable to the model, to the model with its tools, or to a wider human–machine organization then depends on where the competence is taken to be seated, a choice parallel to the typing of processes discussed above. Otsuka draws a related conclusion for deep-learning models: because their reliability is established after the fact on benchmarks, model by model and relative to a test set, it comes to be credited to particular named models, and, while granting that reliabilism can accommodate such models, Otsuka suggests that virtue epistemology may offer a clue to their present justification (Otsuka 2023, 132–135).

4.3 Basing and the operative commitment

A subject can possess good reasons to believe p without believing p for those reasons. A researcher may, for example, have data that support a conclusion while holding the conclusion because a senior colleague asserted it. Epistemologists standardly mark this difference with two notions of justification. Propositional justification concerns whether a subject has sufficient reason to believe a proposition; doxastic justification concerns whether a given belief is held appropriately. One common way of relating them takes propositional justification to be the more fundamental, and doxastic justification to be a matter of the subject's belief being appropriately responsive to, or based on, that propositional justification. Ichikawa and Steup report that the precise relation is disputed but that it is uncontroversial that the two can come apart, as they do in the researcher's case, and that the traditional analysis of knowledge is typically thought to invoke doxastic justification (Ichikawa and Steup 2026, §1.3.3). The basing relation is the relation that holds between a reason and a belief when the reason is one for which the belief is held. It is generally thought to be a necessary, but not sufficient, condition for a belief's being justified, that is, for doxastic justification, and it separates the good reasons a subject has that contribute to the justification of a belief from those that do not (Korcz 2026). To ask about basing is thus to ask not whether good support exists, but whether the belief is held because of it.

This distinction lies on a different level from the approaches of Section 4.1. Several of those approaches, among them internalism, evidentialism, and reliabilism, are accounts of what justifies a belief (Section 4.1): reasons accessible to the subject, the evidence the subject has, a reliable process, and so on. Whichever account is adopted, the subject's having what it counts as justifying can be distinguished from the belief's being held because of it. On the common way of relating the two notions above, the former is propositional justification, and adding the latter, basing, yields doxastic justification. On reliabilism, what justifies is the process that actually caused the belief (Korcz 2026, §5), so the element of the belief's being held because of what justifies it is built into the condition of justification from the start. Basing is therefore not one more answer, alongside the approaches, to the question of what must be added to a true belief for it to be knowledge, but a relation asked about within each of them, and Table 2.1 left it out for that reason. The distinction between having good reasons for a conclusion and holding the conclusion because of them can therefore be taken up without first choosing among the approaches.

The evidentialists Feldman and Conee, for example, mark the distinction between having good reasons and believing because of them with two notions of their own. The first is justification: on their account, an attitude is justified when it fits the evidence the subject has. The second is well-foundedness: an attitude is well-founded when, in addition, the subject holds it on the basis of a part of the evidence that it fits, and does not ignore other evidence it has that would defeat that support (Feldman and Conee 1985, 24–25, 33 n. 21). The two notions are close to propositional and doxastic justification as distinguished above.

This subsection asks what that distinction requires of artificial research. Basing is a relation between reasons and the belief of the subject claimed to know, and, as Section 2.1 noted, the system that conducts research need not be that subject. Where knowledge is attributed to the system itself, what is at issue is the basing of the system's own commitment, and the question becomes which of the reasons found beside its conclusion actually produced it. As Section 2 noted, this book uses commitment as a word broader than belief; here it is a conclusion that a system holds, outputs, and may later revise. The title's operative commitment is the commitment that actually governs what the system outputs and how it later revises it, as opposed to a conclusion or reasons merely displayed or stored beside it. Where a human or a community is the subject, what is at issue is instead the basing of that subject's belief, and the system's output is among its evidence or sources; the social subject is taken up from Section 6 on. The subject relying on the output may also be another artificial system, such as an artificial member of a mixed or artificial S; what that recipient takes up from the output is then its own operative commitment, and the questions of this subsection arise on its side.

The following three cases differ in this respect. An AI report may contain excellent evidence for p while the system produced p because of an irrelevant shortcut; if the system is claimed to know p, what governs its operative commitment is the shortcut, not that evidence. A user may believe an AI conclusion because of institutional prestige while ignoring the evidence displayed beside it; here the subject is the user, whose belief is based on prestige rather than on the evidence. A repository may contain both a conclusion and its grounds without any operative process connecting them; here there is not yet a subject whose belief is based on anything, a situation close to the unread discovery of Section 7. In each case, the co-location of support and conclusion does not establish that the attitude or output is appropriately based.

Where the system itself is claimed to know its conclusion, basing bears directly on how it is evaluated. What an evaluator usually has is the conclusion a system outputs together with the arguments, citations, and data presented with it. Examining these can show at most that the system has good reasons for the conclusion, the matter of propositional justification. Basing concerns the further matter of whether the system produced the conclusion because of those reasons, on which doxastic justification depends, and that cannot be read off the output. Since the traditional analysis of knowledge is typically thought to invoke doxastic justification, on accounts that impose a condition of that kind a persuasive paper does not show that the system knows its conclusion. The difference is close to the one between the first two of the adequacy questions that Chapter 3, Section 5.3 distinguishes: whether an account faithfully describes the process that produced an output, and whether it articulates why the available grounds support accepting the claim. Where the subject is instead a human who relies on the output, the same gap appears on the human's side: reading a persuasive paper supplies the human with reasons, and whether the human's belief is based on them or, as in the second case above, on the system's prestige is a further matter.

Where the system is the subject, the difference also shows in later revision. If a conclusion is not based on the reasons presented with it, learning that those reasons are mistaken need not lead the system to revise the conclusion. This is the false positive in the defeat-sensitivity row of Table 2.1: the user-facing report is corrected while the operative commitment remains unchanged. The change can also run the other way. Even if a conclusion first arose through a shortcut or by chance, if the system later obtains independent evidence and whether it keeps or revises the conclusion comes to depend on that evidence, then on some theories of basing the conclusion's status can change at that point. What is in question is not only how the conclusion first arose but what now keeps it in place and what it is revised in response to. In the case of the chance generator in Section 4.5, the same holds for the community's belief.

What, then, would support the claim that a system's conclusion is based on its evidence, or that the system revises the conclusion in response to evidence? Attaching documents to a conclusion after it has been generated only places support beside the conclusion. One means of showing it is a record of where the conclusion came from, its provenance: which inputs, models, calculations, checks, and interventions actually produced the system's operative commitment and governed its later revision. This requirement is stated in causal terms, as a matter of what actually governed the conclusion, and so it fits most directly the causal theories of the basing relation, on which a reason must cause the belief in an appropriate way. On doxastic theories, an appropriate meta-belief, a belief about the reason to the effect that it is a good reason for the belief, suffices for the belief to be based on it, and a belief may be based on a good reason even if the reason did not cause it in the appropriate way (Korcz 2026, §§1, 3, 5). On an account of that kind, what matters is not only what caused the conclusion but also whether something in the system plays the role of such a meta-belief. On either kind of theory, however, the mere co-location of support and conclusion is not basing, so the failure in the examples above is one that both can recognize. Three things can be distinguished here: the conditions under which basing obtains, the evidence for attributing it, and the means of leaving that evidence. A provenance record belongs to the third. It gives strong evidence for an attribution, but the absence of a record is not the absence of the basing relation. Tests of whether the conclusion changes when the inputs are varied, or is revised when a reason is defeated, can also be evidence for an attribution. Which evidence can support an attribution depends on the theory of basing adopted and on the tests that can be run on the system.

4.4 The evidence a system has and the retrieval store

Justification as evidential fit (Section 4.1) and basing on evidence (Section 4.3) both presuppose an answer to a further question: what evidence the subject has. Evidential fit asks whether an attitude fits the evidence the subject has; basing asks whether the attitude is held because of that evidence. This subsection asks neither of these, but the question both presuppose, for the case in which the system itself is claimed to know (Section 4.3). For such a system, part of what it can draw on may sit in a retrieval store such as a database, and whether what is in the store counts as evidence the system has can change the verdicts on both justification and basing. Feldman and Conee's two notions (Section 4.3) are likewise both relative to the evidence the subject has, but they grant that what it is to have something as evidence is unclear, for example whether propositions one could recall only with prompting count (Feldman and Conee 1985, 32 n. 2). For a system that draws on a retrieval store such as a database, that unclarity becomes a question of how the system is organized and evaluated. Suppose the store holds evidence against a conclusion, strong enough that the store as a whole does not support the conclusion to the degree expressed, but the system retrieves only documents that favor it and states the conclusion with confidence. If everything in the store counts as evidence the system has, the conclusion does not fit the system's evidence; this is the false positive in the evidential-fit row of Table 2.1. If only what was retrieved counts, the conclusion may fit the evidence the system has, and the fault lies not in how the conclusion is held but in how the evidence was gathered. How the gathering of evidence bears on epistemic evaluation is taken up in Section 5. The comparison yields one more point. On accounts that require evidential fit, widening the boundary so that the store counts as evidence the system has need not improve the system's epistemic standing: counterevidence that was outside becomes evidence the system has, so that unless the conclusion is revised the same output can lose its justification. Keeping more information at hand is not by itself an epistemic strength.

Williamson's account draws the boundary of the evidence a subject has in another place. On it, one's evidence is what one knows, not what one is merely in a position to know; counting the latter, Williamson argues, would let a hypothesis be probable on a body of evidence to which one has no access as a whole (Williamson 2000, 202–203). Knowing, on this view, is a state rather than an activity, and one knows many propositions without considering them; in one of Williamson's simple models of a creature that stores information in memory, to which Section 8.2 returns, the creature is stipulated to know everything recorded in its memory although retrieval takes time (Williamson 2000, 199, 226–227). Williamson's account, if adopted, would turn the question of how the retrieval store is organized into whether the system knows what its store records or is only in a position to come to know it.

Whether a store that a system relies on can hold part of what the system believes is a question disputed over extended cognition, that is, over whether cognition can extend to tools and records outside the brain or the system, which Chapter 5, Section 6.1 sets out without deciding it. The epistemology of memory meets the same question for memory knowledge. As Matthew Frise reports, J. Adam Carter and Jesper Kallestrup argue that information stored in a notebook can be part of a subject's dispositional memory knowledge, so that the subject does not merely relearn it on access, whereas Kourken Michaelian argues that this case rests on a functional similarity that fails: it requires biological memory to be purely preservative, while on the view that now prevails, remembering is not simply reproducing a record but generating new content (Frise 2023, §§1.1, 3.2). Carried over to a system that contains both a component that generates content rather than returning stored records and a store that returns what was deposited, as a trained model with a retrieval store does, the contrast Michaelian draws between biological and external memory can fall inside the system, so that whether the store's records count among what the system knows would not be settled by drawing the system's boundary around the store. What would bear on it instead is how the records enter what the system does. On the functional accounts of belief in Section 2.1, a stored sentence is only weak evidence of belief, and stable integration with inference, action, and revision is stronger evidence. A record that the system can retrieve but that shapes neither its inferences nor its revisions would then be closer to a notebook kept for reexamination (Chapter 5, Section 6.1) than to part of what the system believes or knows. An evaluation that credits a system with knowing what its store contains would accordingly need evidence of that integration, not only of retrievability. Section 7 meets the corresponding question for a society, whose infrastructure can hold a result that no member has read.

4.5 Generating, knowing, contributing, and conducting research: the case of an accidental generator

A limiting case separates predicates that need to be kept apart when a result comes from an artificial source. Suppose a random generator happens to emit a novel true theorem. That event creates a true candidate content. If nobody can distinguish it from the generator's overwhelmingly many false outputs, it does not yet create knowledge in the demanding sense. If a competent mathematical community later proves the theorem, or checks a proof obtained by another route, and so comes to know it, the accidental origin of the candidate need not make the community's subsequent knowledge accidental. One role of the justification condition in the traditional analysis was to rule out lucky guesses as knowledge, and Gettier cases suggest that even justified true beliefs can be lucky in a way inconsistent with knowledge (Ichikawa and Steup 2026, §§1.3, 3). The luck at issue in either case lies in the connection between a belief and its grounds or its truth, not in how the candidate came to be considered. This is the familiar point that the route that generates a candidate need not be the route that warrants accepting it (Chapter 3, Section 5.1; Chapter 5, Section 4); here what must be non-accidental is the relation between the community's belief and its proof.

The generator can then be causally important to knowledge production without possessing knowledge, conducting a well-organized inquiry, or manifesting a stable research capacity. The case separates six predicates: generating a true content, knowing it, contributing to another subject's knowledge, conducting research, possessing research ability, and deserving credit for an epistemic achievement. What matters is keeping those predicates apart. Some authors separate generating a true content from contributing even where the content is correct and checked. Weatherall and Wolfson, in a draft on the epistemic value of proof, argue that formal correctness is neither necessary nor sufficient for a proof to have epistemic value, which they locate in the activities by which mathematicians produce, check, and come to understand proofs rather than in the artifacts; if so, "the mere generation of correct proofs is not, in itself, a contribution to mathematics." They distinguish proof assistance, in which computers help prove theorems that mathematicians have stated and want established, from computational proof generation, and they hold that a machine-generated proof, however sound, contributes to mathematics, which they treat as a group activity, only insofar as it prompts human mathematicians' engagement with it (Weatherall and Wolfson 2026, 4, 13–15, 35–38). DeDeo adds a point about judgments of value rather than of truth: the judgment that one theorem matters to another rests in the end on someone's experience of the impasses the first resolves, and automated methods hide such experiences behind mechanical processes, leaving proxies for value without that experience behind them (DeDeo 2024, §2). The random generator separates the six predicates for a content whose truth is accidental; on these accounts, written for mathematics as a social activity, generating a correct, checked result and contributing to mathematics can also come apart. Both take that activity to be one of human mathematicians, and Weatherall and Wolfson picture it as human discovery. Where S is mixed or artificial, whether the same separation holds turns on whether engagement with a proof, and on DeDeo's account the experience of impasses on which judgments of value rest, can be realized by its artificial members; neither takes up a community whose members are artificial. The stronger question of researcher status belongs to the analysis of attribution in Chapter 7, Section 3.2.

The case also shows how the working characterization applies. It counts as research an attempt to bring about an epistemic achievement new to S (Chapter 1, Sections 1 and 3). The generator's output contributed causally to the community's later knowledge, but a causal contribution to an achievement is not by itself an organized attempt at it, so the generator's run does not become research because a theorem it emitted was later proved. Whether the generator contributed, whether the community's inquiry was research, and who bears the resulting knowledge are therefore separate attributions.

5. What verification contributes

Research practice can be narrated as proposing an answer and then verifying it. Many epistemological accounts describe knowledge through truth and some epistemically appropriate relation between a subject and the content. One might then map the two vocabularies onto each other: verification produces justification, and justification converts a belief into knowledge. This section examines that mapping in three moves. It first distinguishes verification, an activity, from justification and the other statuses a result may have: a verification can succeed without its result having the status in question. It then takes up a divergence in the opposite direction: a conclusion can be well supported by the evidence at hand while the inquiry that gathered that evidence was badly conducted by standards other than those that govern the result's status. Finally, it returns to the section's question, what verification contributes to a knowledge attribution, by dividing the route from a test to a possible knowledge attribution into steps. Verification directly supplies the first two of them, a result and its evidential bearing; for that to change an attitude takes a basing relation (Section 4.3) or an updating relation, and for the attitude to be knowledge takes whatever the adopted account requires (Sections 4.1 and 4.2). Any of the steps can fail. Where a result is unfavorable in particular, Duhem's point that a hypothesis is tested only together with auxiliary assumptions, and the epistemology of defeat, show what is left open about which element it bears on. Section 5.1 turns to the heterogeneity of verification, Section 5.2 to confirmation, what a result supports and how far, and Section 5.3 to statistical methods as a bounded illustration.

This book uses verification in a broad sense: verification refers to a family of activities through which investigators try to determine whether a candidate answer should be treated as true. The word is not used to promise conclusive establishment. Different disciplines and literatures instead speak of testing, confirmation, corroboration, proof, validation, audit, replication, or criticism, with different objects and success conditions. The important common function is assessment of a candidate's epistemic standing under uncertainty, not one universal procedure.

Verification is something investigators do; justification, warrant, reliability, safety, and knowledge name possible statuses or relations. Warrant is used here broadly for the epistemic support for accepting a claim: the justification of a belief is one form of it, but warrant is also said of results and of the acceptance of claims. An explicit test is not universally necessary for justified belief: perception, memory, testimony, and dependence on an institution can provide epistemic support. Nor is a passed test sufficient. A test can target the wrong quantity, depend on a defective instrument, be analyzed through an unsuitable model, leak its answer, or be interpreted beyond the conditions it supports. The high-gain antenna of the Galileo spacecraft, for example, had passed deployment tests in air and in vacuum, but in air the exposed titanium of its pins kept reacting with oxygen, so that friction never rose high enough to stop deployment, and in the vacuum test the pins and sockets had not moved against each other. In space, vibration from the upper stage after launch did move them against each other in vacuum; the bare titanium pins galled against their sockets and held some of the ribs, so that when the antenna was commanded to open in 1991 it opened only partly. The report of the later investigation concluded that the test conditions had not been adequate to reveal the problem (Johnson 1994, 375–377).

The distinction between verification and justification showed that an activity can succeed without its result having the appropriate status. The gathering of evidence also allows a divergence in the opposite direction. A conclusion can fit the available evidence while the inquiry that produced that evidence was selective, prematurely stopped, or organized so that inconvenient observations were unlikely to appear. Conversely, responsible evidence seeking can end in suspension because the available evidence remains inadequate. Contemporary work on the norms of evidence gathering argues that these diachronic questions should not be reduced to the synchronic question of whether an attitude fits the evidence presently possessed (Flores and Woodard 2023). The distinction need not be built into every theory of justification to matter for research. It shows why a well-supported result and a well-conducted inquiry are not the same achievement. For an artificial system, retrieval policy, source exclusion, measurement choice, active information seeking, and stopping rules can therefore remain epistemically consequential even when the final inference is impeccable conditional on its input.

Of the items just listed, stopping rules are the one whose epistemic consequence is itself disputed, since on one view the rule by which an experiment was ended makes no difference to what its data show. Romeijn sets out both sides. Under the likelihood principle, which standard Bayesian inference satisfies, all the information that matters to inference about hypotheses is contained in the likelihood function, the probability of the observed data under each hypothesis, so the rule that ended an experiment does not change what the data imply. Frequentists resist the principle because the error rates that justify a test's verdict depend on the design, so that the same data can warrant rejection under one design and not under another, as in the stopping-rule problem described in Section 4.2 (Romeijn 2025, §§3.2.2–3.2.3). The point about evidence gathering, however, does not require settling that dispute: even if a stopping rule leaves the import of the collected data unchanged, it can affect what evidence was collected and whether the inquiry was responsibly conducted.

Whether responsible conduct of inquiry is itself an epistemic standard is a further dispute (see also Chapter 4, Section 5). Read through each of the views that follow, the difference between a well-supported result and a well-conducted inquiry nonetheless survives. Feldman and Conee, whose evidentialism Section 4.1 introduced, grant that gathering evidence responsibly has some epistemic merit but deny that the justification of the resulting attitude turns on it. In an example they take from Kornblith, a physicist who fails even to hear a colleague's devastating objection remains justified in believing their theory, provided that they were justified before and their evidence is unchanged; what is open to criticism is their character (Feldman and Conee 1985, 21–22, 31–32). Thorstad denies that there are distinctively epistemic norms of inquiry at all. Thorstad argues that the considerations supporting distinctively epistemic norms for belief do not carry over to inquiry, which is an activity, so that whether to gather more evidence, double-check, or stop is settled by all-things-considered reasons, among which the prospect of truth or knowledge is one (Thorstad 2022, §§3–5, 6.3, 7).

Kauppinen goes further. Where Thorstad counts epistemic considerations among reasons that combine into an all-things-considered ought, Kauppinen argues that all reasons to pursue a question are practical, transmitted from reasons for what answering it would help one do, so that there are no epistemic norms of inquiry, and that where an inquirer does deserve epistemic criticism for how evidence was gathered, the fault lies in a standard epistemic norm on belief or action. In a case Kauppinen adapts from Flores and Woodard, someone consults a single source that is in fact unreliable and happens to reach a true answer. On Kauppinen's account this person is epistemically criticizable without appeal to any norm of inquiry: they plausibly violate an epistemic norm on action, if there are such norms, by treating the source's reliability as a reason for action without knowing it or being justified in believing it, and Kauppinen allows that the fault may lie instead in their holding that belief without evidential support. In a variant in which the person has good but misleading evidence of its reliability, they are not criticizable (Kauppinen 2026, §§3–4).

What differs among these views is whether the second assessment, that of the inquiry, is epistemic, and so whether the requirement of responsible evidence acquisition in Section 9.1 is an epistemic or a practical one. If it is epistemic, decisions about search and stopping are assessed by standards that concern truth or knowledge, independently of what the result is used for. If it is practical, how far to search and when to stop is settled in light of the purposes of the inquiry, its costs, and what is at stake, so that the same stopping point can be appropriate for one purpose and not for another. Where the inquirer is an artificial system, those purposes and costs need not be its own: they may have been fixed by whoever set its objective and budget, or by the system where it sets them itself (Chapter 7, Section 4.3), and whose purposes govern how far inquiry goes is part of the question of allocation (Chapter 10). On Kauppinen's view it would be practical, while any unsupported assumption about the reliability of sources on which a system acts in gathering evidence would remain open to epistemic criticism: criticism of the system, if it is itself a subject of such norms, and otherwise of those who built the assumption into it.

Leaving aside the assessment of the inquiry as a whole, we can return to the section's question of what verification contributes to a knowledge attribution. The bridge from a test to a possible knowledge attribution can be represented without pretending that its arrows are automatic:

target claim H + method M + auxiliary commitments A
    -> outcome R

R + background knowledge and evaluative standards N
    -> evidential bearing E on H or another commitment

E + a basing or updating relation B
    -> changed epistemic attitude D

status of D under the adopted account of knowledge
    -> possible knowledge attribution

Read line by line, the schema separates four steps. In the first, a target claim H is tested by a method M under auxiliary commitments A, such as the instruments, conditions, and preprocessing on which the test relies, and the test yields an outcome R. Suppose, for example, that H is the claim that a model has some capability, M is a benchmark, and A includes the assumptions that the benchmark measures that capability and that its items did not leak into the model's training data. In the second step, the outcome becomes evidence only when it is read against background knowledge and evaluative standards N, such as what score would be significant and how other models perform; it then bears on H, or, if contamination is suspected, on A instead. In the third step, evidence changes an attitude only if someone's commitment actually changes because of it, through a relation B, such as the basing relation of Section 4.3 or the updating by which degrees of belief are revised in response to evidence. In the fourth, whether the changed attitude D is knowledge depends on what the adopted account requires of it (Sections 4.1 and 4.2). The third and fourth steps concern the attitude of the subject claimed to know: the system's own commitment where knowledge is attributed to the system, and the belief of the humans or the community where they come to know through it (Section 4.3).

The last line does not assemble knowledge from conditions shared by all the accounts. On knowledge-first views, the preceding attitudes and their grounds are assessed in relation to knowledge, rather than combined to define it. Further, the last line speaks of a possible attribution partly because attributing knowledge is something an evaluator does, and insufficient evidence for an evaluator to attribute knowledge does not by itself show that the subject lacks knowledge. The schema locates questions for an assessment; it is not an algorithm for settling them.

Every arrow is a substantive site of failure. The Galileo spacecraft case above concerns the first two arrows, and Sections 5.2 and 5.3 return to the second, for confirmation and for statistical methods; what follows concerns the second and third arrows when a result is unfavorable. The method and auxiliaries help produce an outcome, but the outcome does not carry its own epistemic address. A failed prediction may bear on the focal hypothesis, a boundary condition, an instrument model, a preprocessing choice, a background theory, or some combination. Duhem's analysis of physical testing makes the limited but crucial point that an isolated hypothesis is not what confronts an experiment; a negative result implicates a larger theoretical arrangement without itself identifying the faulty element (Ariew 2026). The later process analysis will examine how investigators decide where to revise (Chapter 6, Sections 2 and 4). Here the point is that a verification outcome cannot be mapped mechanically to a change in the main hypothesis.

The epistemology of defeat, which studies how information can take away the justification a belief had, does not settle which element is at fault, but it makes the point finer by distinguishing the ways in which contrary evidence can bear on a claim. In the distinction Pollock drew, rebutting evidence favors the denial of the claim: a careful replication that finds no effect rebuts the claim that a drug works. Undercutting evidence instead gives reason to doubt that the usual connection between a claim and the grounds offered for it holds in the case at hand, without favoring the claim's denial: learning that a sample was biased undercuts the inference drawn from it, although it does not show the conclusion false (Koons 2025, §1.1). Higher-order evidence bears not directly on the claim but on the investigator and the investigator's evidence: it can make the investigator's own competence, reliability, or evidential situation doubtful, as when an analyst learns that the analysis was carried out under conditions known to produce errors (Roush 2017, §3). How much confidence in the first-order claim such evidence rationally requires giving up is disputed. As Roush reports, Feldman and Horowitz hold that one cannot rationally be confident that p while believing that one's evidence does not support p. On Feldman's view, the belief "p, but my evidence does not support p" cannot be both true and reasonable, since its second half undermines the reasonableness of the first. Horowitz argues that a subject who stays confident in this way can explain having been right only as luck. Williamson, Weatherson, and Wedgwood argue instead that such level-splitting can be rational where the doubt concerns the subject's own competence. On their respective views, a proposition can be highly probable on one's evidence while it is also highly probable that its probability is low; a norm of rationality does not cease to apply merely because the subject has evidence of having failed to follow it; and a support relation that actually holds can justify a belief whether or not the subject believes correctly about it (Roush 2017, §3). What counts as an appropriate response in the defeat-sensitivity row of Table 2.1 differs accordingly.

Carried over to an artificial researcher, Duhem's point means that a scalar loss attached to the headline hypothesis cannot by itself exhaust the diagnosis of what an adverse result gives reason to revise, since an instrument model, a sampling assumption, a preprocessing step, or a background regularity is also a candidate for revision. A configuration that represents the loss as a scalar while keeping the candidates for revision separately is compatible with this. That much follows from Duhem's point alone. The distinctions of defeat add two further targets. Undercutting evidence, without favoring the denial of the conclusion, gives reason to doubt the support relation between a claim and its grounds; higher-order evidence can call for revising the estimate of an evaluator's competence, whether the investigator's or a validator's. Defeat is also not confined to adverse results: it can arise after a test has been passed, as when a sample is later found to have been biased. These distinctions do not, however, supply a procedure for locating the defeated relation. Locating it still requires domain knowledge, experimental skill, and possibly further intervention. What was earlier counted as evidence may itself need revising. Williamson argues that standard Bayesian updating, on which a proposition once conditionalized on keeps probability 1 thereafter, cannot represent later evidence undermining something that had been evidence, and proposes keeping a record of evidence separate from the probabilities, so that a proposition can lose its standing as evidence (Williamson 2000, 214, 219–220). On this argument, a system whose updates can only add evidence could not represent such a loss.

What these distinctions require of a system depends on who constructs the verification and who decides, after an adverse result, where to revise. If the system runs checks that others designed, whether human investigators or other artificial agents, and they read the results and decide what to revise, the work of determining which element a result bears on stays with them; for them to do it, the system's report has to show not only a pass or fail on the headline hypothesis but what was tested under which auxiliary assumptions. Where those others are artificial agents, what the report has to show extends to how far they are independent of the system, for example whether they and the system are instances of one base system, since a critic adds to error detection only so far as it does not go wrong where the producer does (Chapter 11, Section 3). If the system itself decides what to test, what counts as passing, and where to revise, these distinctions become a task the system takes on. Chapter 6, Section 4.2 discusses that task for research done by a machine, and Chapter 7, Sections 4.2 and 4.3 distinguish, as comparisons of autonomy, who supplies selection, evaluation, and correction, and whether a system only applies supplied standards or also selects, constructs, and revises them. This is a different axis from the question of who the subject of knowledge is. Where others come to know through the system and the system constructed the verification, they need to be able to read from its reports how it carried out that task if they take on the diagnosis of results themselves. On accounts that require accessible reasons, a recipient needs reasons it can use, but these may concern how the result was checked or, as on reductionist accounts of testimony, the system's reliability for the use at hand (Chapter 11, Section 5.2); the system's reports are then one source of such reasons. On accounts on which the reliability of the process suffices, and on Sosa's account (Section 4.2) applied to this case, on which the correctness of a testimonial belief is owed more fully to competence seated in the chain of transmission, a recipient can know without being able to read the whole structure of the verification. What then supports the attribution is not that the reports can be read but evidence that the system's way of verifying is reliable. Where knowledge is attributed to the system itself, it is further required that its operative commitment actually change in response to such revisions (Section 4.3).

Putting this together, what verification contributes to a knowledge attribution can be stated as follows. An outcome's evidential bearing on a specified research claim depends on relations involving measurement and whether it captures what is to be measured, sampling, interpretation, and the inferential target. Available evidence changes an attitude only through some actual basing or updating relation; the attitude is the system's operative commitment where knowledge is attributed to the system, and, where others come to know through it, their belief, or their operative commitment if they are artificial systems. A well-supported attitude becomes knowledge only if it is true and satisfies whatever additional conditions the adopted account imposes.

5.1 Verification is heterogeneous but not therefore functionless

The start of this section treated verification as a family of varied activities with a common function. This subsection shows that the variety does not empty the common function, and that the common function does not erase the differences among the means.

A verification event can generate evidence, test whether evidence supports a claim, expose a defeater, estimate the reliability of a process, question whether what is being tested is really the intended target, or make a local result available for social criticism. It can leave confidence in the focal claim unchanged while sharply reducing confidence in the instrument or method. It can also reshape the claim being assessed.

The functional commonality does not erase differences in epistemic force. An analogy, pilot result, randomized trial, independent replication, deductive proof, historical source comparison, and interpretive criticism can all alter what is taken seriously. They differ in what they bear on, the alternatives they discriminate, their defeasibility, and the kind of standing they can provide. Calling all of them belief updates would describe one effect while losing the reason their effects differ. For an artificial researcher, this means that a stage among its steps named verification, or a pass or fail reported for that stage, does not by itself say what was checked. Chapter 6, Section 3.1 compares checks by two of these differences: what part of the question an operation bears on (its target), and how firmly it settles that (its force).

Verification also need not follow generation as a separate phase. Candidate formation already uses prior evidence, coherence, analogy, preliminary checks, and expectations about what could survive scrutiny. A mathematical proof can construct a result and display its warrant in the same object. Criticism can change the content rather than certify it. Seen broadly, proposing and assessing are different roles. Seen more finely, they are mixed together in a single course of gradually reorganizing what is taken as settled, and how far. Chapter 5, Section 4 develops that temporal and practical structure.

5.2 Confirmation: what a result supports, and how far

The second arrow of the schema, from a result and background standards to evidential bearing, is the relation that confirmation theory studies: how, and how far, evidence supports a hypothesis or tells against it. On a classification that Crupi calls relatively standard, the main approaches fall into three families. One is confirmation by instances, as in Hempel's theory. Another is hypothetico-deductivism, on which evidence confirms a hypothesis that, together with auxiliary assumptions, entails it. Proposals to strengthen it include requiring that the evidence have been predicted in advance (predictivism), that it not have been used to construct the hypothesis (use-novelty), that it would have been unlikely had the hypothesis been false (eliminativism, of which appeals to severe testing are a mark), that the hypothesis account for it in a simple, unified way (simplicity), or that the hypothesis explain it (explanationism). The third consists of Bayesian theories, which represent support with probabilities (Crupi 2025, §§1–3). Confirmation theory concerns a relation, the support that evidence lends to hypotheses, rather than verification as an activity. Activities in the verification family that try to determine whether a result supports a hypothesis, such as testing and replication, ask about this relation, but the family also includes proof and interpretive criticism, whose work is not exhausted by it (Section 5.1). Several disputes among the approaches of confirmation theory appear elsewhere in the book: prediction against accommodation and severe testing in the appraisal of candidates (Chapter 5, Section 4.2), the Bayesian and error-statistical treatments of where a failed prediction places blame (Chapter 6, Section 2.3), the independence of checks (Chapter 6, Section 3.2; Chapter 11, Sections 2 and 3), and the contrast between Bayesian and frequentist statistics (Section 5.3). This subsection takes up one distinction that those discussions leave implicit, and what it and some related results imply for artificial research.

Within Bayesian confirmation theory, Crupi distinguishes two notions of confirmation. On confirmation as firmness, a hypothesis is confirmed to the extent that it is probable given the evidence. On confirmation as relevance, evidence confirms a hypothesis to the extent that it raises the hypothesis's probability. A positive result on a test for a rare disease can raise the probability of the disease while leaving it low: on the relevance notion the disease is confirmed, and on the firmness notion it is not. Crupi reports Salmon's description of the distinction as "extremely fundamental" and yet "sometimes unnoticed", and notes that it has had to be stressed again and again, in philosophy and also in artificial intelligence and the psychology of reasoning (Crupi 2025, §3.3). In Crupi's presentation the probabilities represent states of belief. Here they can be read as those of whoever assesses a claim, so using them does not presuppose that an artificial system itself has degrees of belief (Sections 2 and 2.1; Chapter 6, Section 1).

For artificial research the distinction bears first on what a passed check reports. On the Bayesian notions, a passed check can show that its result raised the support for a claim, while how well supported the claim then is also depends on how probable it was before the check. Crupi does not discuss this case, but applied to systems that generate candidates cheaply and in large numbers, the distinction has a consequence. If most of the candidates generated are false and nothing distinguishes one from another before the check, each is improbable before it is checked, and a check that a false candidate passes with a fixed probability leaves a passed candidate less firmly supported than the same pass would leave a candidate in an inquiry that tries a few already plausible ones. The consequence holds for whoever is claimed to know. Where the system itself is, it concerns how far the pass supports the system's own commitment. Where others come to know through the system, a report that a candidate passed is part of their evidence, and someone who reads the report without knowing how many candidates were generated, and how many of them were likely to be false, can mistake how far the pass supports the claim. Eliminativist and error-statistical views assess a pass differently, by how unlikely it would have been had the hypothesis been false (Crupi 2025, §2.4). On either view, a report that a hypothesis passed a check leaves open, by itself, what the pass supports and how far; on the argument from inductive risk, on which the more harmful an error would be the more support acceptance requires, whether the support is enough for a given use is a further judgment (Chapter 11, Section 4). The force of a check in Chapter 6, Section 3.1 can accordingly be reported in more than one way.

Four further points in Crupi's survey bear on how such claims are checked. First, on naive hypothetico-deductivism, for any hypothesis and any evidence consistent with it there are auxiliary assumptions under which the evidence confirms the hypothesis. Crupi connects this with Duhem's point, and notes that many particle theorists of light, who reshaped their auxiliaries to recover the facts that wave theorists had deduced, did not claim those facts as confirmation (Crupi 2025, §2.3). Applied to a system that can choose preprocessing steps, exclusions, or model components after seeing its data, this means that such a system can, to the extent that its choices are unconstrained, often produce agreement that naive hypothetico-deductivism would count as confirmation. On predictivist and use-novelty views, whether those choices were fixed before the data were examined becomes part of what a claim of confirmation has to show (compare the testing procedure specified in advance in Section 4.2 and the provenance record of Section 4.3); on eliminativist views, what has to be shown is rather whether, with those choices made, the agreement would have been unlikely had the hypothesis been false (Chapter 5, Section 4.2); on Bayesian treatments, the auxiliaries are instead weighed by how probable they are (Crupi 2025, §3.5). Second, on the use-novelty view, data that served as a constraint in constructing a hypothesis cannot also confirm it; Crupi cites cross-validation, following Schurz, as a routine statistical application of the idea, and reports that the view is itself disputed (Crupi 2025, §2.4). On this view, agreement with the benchmark on which a system generated or tuned its candidate confirms it little or not at all, which is one reason evaluation on routes withheld from selection matters (Chapter 11, Section 8). Third, on the relevance notion, if a result confirms a hypothesis, the result's failure to occur disconfirms it, so an assessment that would count either outcome as support, a judgment of the "heads I win, tails you lose" kind, is inconsistent on that notion: at most one of the two outcomes can confirm the hypothesis (Crupi 2025, §3.3). An assessment of an artificial researcher's claim can accordingly ask what result would have counted against it. Fourth, on the relevance notion, evidence that was less probable beforehand, as more varied evidence often is, confirms a hypothesis that entails it more strongly (Crupi 2025, §3.6). Agreement among repeated runs of one base system is likely whether or not the hypothesis is true, to the extent that the runs can go wrong in the same places, so it is hard to count as varied evidence in this sense (Chapter 11, Section 5.4). Weisberg and Reisman distinguish three levels at which a result of modeling can be robust: surviving changes to a model's parameters, to its mathematical structure, and to the whole representational framework in which it is framed (Weisberg and Reisman 2008, 9–11, 15–17, 25). Where runs or instances share one modeling framework, their agreement could at most show robustness within that framework, not the robustness that varying the framework tests.

5.3 Statistical methods as a bounded illustration

The heterogeneity of verification seen in Section 5.1 does not appear only between fields such as history and the natural sciences. Within a single family of methods, statistics, what is taken to support a conclusion differs from method to method. Otsuka argues that statistical derivations can proceed as mathematics without settling what their probabilities mean, but that the interpretation adopted limits what their results can meaningfully be said to show (Otsuka 2023, 42–43, 176–178). Otsuka treats statistical inference as resting on ontological assumptions—a uniformity of nature behind the data and, in parametric inference, a chosen family of distributions—and reads its schools as differing in their concepts of justification (Otsuka 2021, 3–7, 13).

On Otsuka's reading, Bayesian updating justifies a degree of belief by valid inference from an agent's other degrees of belief, and so resembles internalism, while attempts to secure its priors as a starting point resemble foundationalism, the view that some beliefs are basic, needing no further justification, and serve as the ultimate resource for justifying all others (Otsuka 2023, 53–57, 61–66). On Otsuka's account, a Neyman–Pearson test can instead justify a verdict through the reliability of the testing procedure. Nozick's two tracking conditions require that, had p not been true, the subject would not have believed p, and that, had p been true, the subject would have believed p; the first is the sensitivity of Section 4.1. Otsuka reads a test's confidence coefficient, the probability of not rejecting a true null hypothesis, as a measure of how far the test meets the first condition with respect to the alternative hypothesis, and its power, the probability of rejecting a false null hypothesis, as a measure of how far it meets the second, and so aligns testing with reliabilism (Otsuka 2023, 88–94). Otsuka reads model selection by the Akaike information criterion as pragmatist, because the criterion assesses a model by expected predictive performance rather than by closeness to the data-generating process, and suggests that virtue epistemology may offer a clue to the present justification of deep-learning models (Otsuka 2023, 118–124, 129–135).

Otsuka presents these correspondences as a simple, idealized model rather than a perfect analogy, and denies that the comparison makes Bayesian statisticians internalists or internalists Bayesians (Otsuka 2023, 55–56, 179–180, 182). Otsuka also argues that neither Bayesian nor frequentist statistics can rest content within the single epistemology it resembles. On Otsuka's account, posterior model checking tests the whole network of prior, likelihood, design, and measurement assumptions against data, which takes Bayesian practice beyond relations among the agent's degrees of belief and away from the foundationalist picture (Otsuka 2023, 67–72). The error rates that justify a test's verdict, Otsuka argues, depend on distributional and design assumptions that the verdict itself does not check, so a frequentist who takes responsibility for the reliability of the process must scrutinize those assumptions by other means (Otsuka 2023, 95–98, 106).

The more durable lesson is that a formally correct operation does not determine its epistemic role. Otsuka notes that a computed posterior distribution or a low p-value does not by itself say anything about the phenomenon under study (Otsuka 2023, 105). On the frequentist view, moreover, the same data can warrant rejecting a hypothesis under one stopping rule and not under another, which defenders of the likelihood principle deny (Sections 4.2 and 5). Priors, likelihoods, sampling plans, stopping rules, loss functions, error criteria, measurement models, and target definitions can each alter what conclusion is warranted. Where an artificial system that chooses or applies a method is itself said to conclude, what it is warranted in concluding then depends on how it relates the procedure to the question, data-generating conditions, assumptions, and costs of error. Where others draw conclusions from its output and assess for themselves how far that output supports them, they need to see how the system related the procedure to these; on accounts on which the reliability of the process suffices, a recipient can know without seeing this, but an evaluator who supports the attribution needs evidence that the system's way of doing so is reliable (Section 9.1). Whether the system can criticize and revise the governing standards is a further question about epistemic self-government (Chapter 7, Section 4.3), not a prerequisite for every use of statistics.

6. What is social about knowledge: the reference for novelty, production, possession, and retention

Sections 2 to 5 asked what the attitude of whoever is claimed to know must be like. In research, knowledge is also social, and in more than one sense: the reference relative to which it is new is social, its production can be collective, and its possession may or may not be. This section separates these senses. It first takes up the reference S, then separates production from possession through a distinction Fagan draws, extends that distinction into six relations, including retention, and compares accounts of social knowledge across them (Table 2.2). It then asks what these accounts imply for artificial research, by the combination of A and S. It ends by separating three claims about the social that differ in strength, so that the social character of knowledge is not read as making truth a matter of agreement.

The variable S arose from the observation that Chapter 1, Section 6 made with its examples of learning and rediscovery: in the case of knowledge, when research is said to have produced new knowledge, the novelty claimed is not the novelty of one person's private learning. For knowledge claims, S accordingly has the two connected roles introduced in Chapter 1, Section 1. It is the reference relative to which the knowledge sought, once realized, is new, and it identifies the society for which that knowledge is sought. Neither role attributes the knowledge to S; who knows it, and whether and how it has become S's, are left to the accounts compared below. Both roles presuppose that S's epistemic situation could come to include knowledge of the kind sought (Chapter 1, Section 2), but neither tells us how a society can know. Substituting a group for the individual variable in X knows p does not turn it into one large person or settle whether group knowledge reduces to members' knowledge.

In ordinary discourse, the unstated reference can be humanity and, in practice, a relevant specialist community standing in a wider network of testimony, education, records, instruments, standards, and institutions. Since what is already known differs from one society to another, however a society's prior epistemic situation is construed (Table 2.2), a claim of novelty can be assessed only when the reference is explicit. A proposition can be new to one laboratory and familiar to another, new to a field and established in a neighboring discipline, new to humans and familiar to an artificial community, or new to one artificial society and not another. The difference need not concern novelty alone. In an ethnographic replication of Laboratory Life conducted at the same institute, Havstad, for example, suggests that one sentence from a review by the laboratory under study—that a network of target genes defines the physiological response—would be treated at the time of the study as close to established fact within that laboratory, as contested by rival laboratories, and as speculative or false by scientists outside the field (Havstad 2020, 25–26). What differs across these groups is the standing accorded to the sentence, a standing that this chapter distinguishes from knowledge.

The other two senses, production and possession, are separated by Fagan, who distinguishes two theses that the simple formulation can otherwise conflate. The first is that groups can be necessary to produce scientific knowledge. Fagan identifies three respects in which they can be: in many cases of inquiry, groups are required to satisfy methodological norms, to develop theoretical concepts, or to validate the results of inquiry as scientific knowledge (Fagan 2012, abstract). The claim concerns many cases of inquiry, not all. The second, further thesis is that groups possess scientific knowledge in their own right. Fagan judges the support for it "more equivocal": the cases offered for it can also be explained in terms of "relational complexes of members' beliefs" (Fagan 2012, abstract). As Longino reports, Fagan disputes Bird's account in favor of a relational view, on which shared knowledge is not knowledge held by a social entity existing over and above the individuals who constitute it but a function of epistemic relations among those individuals (Longino 2025, §5). The distinction between the two theses gives claims that knowledge is social analytical force without requiring an immediate commitment to a group mind.

Acceptance is one point at which the two theses come apart. A community's acceptance can be what validates a result, and so belong to production in the third of Fagan's respects; whether it also shows that the community possesses the result as an irreducibly collective belief is a further question. Duygu Uygun Tunç argues that a research collaboration's public affirmation of a result, as in a publication or an announcement, is distinct from the distributed process that produced the evidence for it: the same research process can be combined with dictatorial, majoritarian, or consensus-based procedures for affirming its results, and at CERN, on her report, decisions to publish are taken by dedicated publication committees and are not a function of the members' beliefs (Uygun Tunç 2023, 10 and n. 18). She also notes, without relying on it, a position on which groups cannot have beliefs but can accept propositions as the view of the group, on the ground that beliefs are formed automatically and involuntarily while group views are not (Uygun Tunç 2023, 10 n. 16). That ground matches the distinction Chapter 3, Section 6 takes up from Elgin, on which belief in Cohen's sense is largely involuntary and spectatorial while acceptance is voluntary and agential (Elgin 2025, 11, 35–36); Elgin describes epistemic communities as devising and collectively endorsing the methods and standards they work by, and holds that mere consensus does little unless most members function autonomously (Elgin 2025, 121–122). Uygun Tunç's example separates a decision to publish from members' beliefs. Applied to Fagan's two theses, that distinction leaves open whether public affirmation establishes collective possession.

A knowledge-producing ecology contains several relations that can coincide or separate. They can constitute, enable, communicate, maintain, or promote the acquisition of knowledge, but they are not thereby alternative meanings of one event called production:

• a candidate content and its grounds are generated locally;
• the result is articulated in a form that others can inspect, such as a written claim, a formal proof, a dataset, a mathematical or physical model, or a demonstration;
• one or more subjects acquire the relevant attitude and standing;
• qualified participants criticize or validate the result (the relations Chapter 11 examines as criticism and certification);
• the result enters a group's live structures of inference, decision, and action;
• it is retained in a form that can be retrieved, defeated, corrected, or withdrawn.

In separating these six relations, this book extends the distinction between producing and possessing knowledge that Fagan draws. These are analytical dimensions, not a mandatory pipeline. Criticism can occur before public release. A subject can acquire the relevant attitude toward a result before the result has been articulated in any form that others could inspect. Whether an artificial system's internal states can realize such an attitude depends on the account of belief adopted (Section 2.1); understanding that may be tacit is taken up in Chapter 3, Section 6. A result can also be available for inspection without discriminating between rival theories. Measurements of how the electron's mass varies with velocity, for example, were regarded at the time as the arbiter between rival models of the electron and the broader theories perceived to be attached to them. Lorentz's theory and Einstein's, however, agreed in all their empirical predictions, including those for this variation, and data presented in 1908 seemed to support a prediction that was by then Einstein's as well as Lorentz's (Janssen and Mecklenburg 2004, 2, 31). Since the two theories shared that prediction while differing in their foundations, a measurement favoring it could not decide between them. On one historian's account, the essential difference between Einstein's theory and that of Lorentz and Poincaré lay in whether it went as far as changing the concepts of time and space; until around 1910 the Lorentz–Poincaré theory was still widely used, and whether the ether should be given up became a question for physicists, the historian suggests, only after 1910 (Hirosige 1964, 107). An instrument or protocol can integrate a result into practice before broad belief forms. An archive can retain a result after the people who once understood it have died. Publication can leave a result socially inert. For example, unusually low ozone values measured at Syowa Station in Antarctica in 1982 were printed in a 1984 report that listed the lowest values since 1966 second among four summary points; the observer later recalled that presenting the results in Japan and at an international symposium drew no particular response, although the printed proceedings of the symposium became the basis on which the observer was credited with first reporting the depletion (Chubachi 1984, 13, 18; Chubachi 1993, 817).

Different accounts assign knowledge at different points or deny that the points form one scale. Table 2.2 sorts these accounts into eight families. For each family, the middle column states what would bear social knowledge and through which relation it would possess it, and the right column states what follows for a result produced by an artificial system, in a form that holds whether S is human or artificial; the paragraphs after the accounts separate the two cases. The table is a simplified guide rather than an exact statement of each account. Each row states its family through one representative account, named in the paragraph after the table, and two rows join positions that differ: the member row covers both counting enough members and counting only the operative members, those with decision-making authority in the relevant domain, and the group-attitude row covers both group belief and group acceptance.

Table 2.2. Candidate bearers and possession relations for social knowledge: a simplified overview

Account family	Candidate bearer and possession relation	Consequence for an artificial result
Individual-possession	suitably situated individuals know, even when production is distributed	for the result to become S's knowledge, it must reach individual knowers in S through evidence or justified dependence
Member or operative-member	enough members, or the members with decision-making authority in the relevant domain, hold the relevant attitude, subject to group-level conditions such as coherence among their bases	routing to the right members of S is part of successful social acquisition, and counterevidence held elsewhere in the group can defeat it
Relational	a group's belief is a system of members' beliefs and the epistemic relations among them; groups can be necessary for production without thereby possessing knowledge in their own right	the result must be connected to members through epistemic relations among them, not only deposited
Group-attitude	a structured group forms belief or acceptance under group-level norms	aggregation, authorization, coherence, and responsibility must be specified, and whether an accepted position is a belief is itself disputed
Distributed cognition without distributed knowing	the larger system performs cognitive work, but knowledge remains attributable to individuals	records and computational components can take part in production without being knowers; whether an artificial system that conducts research could itself know is not settled by this account
Knowledge without a knowing subject	scientific knowledge is ascribed to publicly certified results of reliable processes rather than to an individual or collective knower	a record's status turns on certification and public acceptance, not on anyone's present attitude
Distributed or group-functional	a properly organized collective realizes the cognitive function associated with knowing	members, records, retrieval, and action can jointly realize a social knower
Critical-contextual	accepted content that conforms and has survived criticism under specified social conditions is knowledge for that community	public venues, uptake, standards, and authority relations matter, not visibility alone

The accounts discussed above and below represent the rows in order: Uygun Tunç's individual possession, Lackey's operative-member account, the relational view Fagan describes, the joint-acceptance account that Gilbert states, Giere's distributed cognition without distributed knowing, Giere's adaptation of Popper's epistemology without a knowing subject, Bird's functionalism, and Longino's critical contextual empiricism. Compared with one another, the accounts differ partly in which of the six relations listed above they treat as decisive for social knowledge; in the terms of Chapter 1, Section 2, they differ over which relation must hold for a result to have entered S's epistemic situation. For example, individual-possession and member accounts weigh most heavily the acquisition of an attitude by subjects; critical-contextual accounts, criticism together with acceptance by members; Giere's adaptation of knowledge without a knowing subject, certification and public acceptance; and distributed or group-functional accounts, the result's entry into live structures of inference and action and its retention.

Uygun Tunç argues that distributed scientific systems can collectively produce evidence and knowledge while the subjects who possess knowledge remain individuals. On this account, what collaborative research achieves collectively is the establishment of evidence supporting a proposition. That is not a process of belief formation, so it yields no group belief. Moreover, an account that denied possession to individuals would imply that no one learns anything even from the most successful collaborations (Uygun Tunç 2023, 2, 9). Giere likewise argues that distributed cognition need not entail distributed knowing: attributions of belief, knowledge, and agency should be restricted to the human components of a distributed cognitive system. Giere treats extending such attributions to the system as a whole as a higher-level interpretation, justified only if it brings some theoretical advantage, and judges that it brings none. Questions such as whether the Hubble telescope system as a whole expected to find something, or believed that it had found it, "do not make much sense," Giere writes. Restricting the attributions to humans, by contrast, yields a theoretical framework that everyone who studies science, from historians to anthropologists, can share (Giere 2007, 315–319). These views make collective production compatible with denying that the system as a whole knows. Giere's position also points to a further option. Adapting Popper's idea of epistemology without a knowing subject, Giere holds that the institution of science produces scientific knowledge with no need to posit a collective agent that knows it; individuals can claim to know its conclusions because they know that the conclusions were certified by reliable, accepted procedures (Giere 2007, 318). Popper's own formulation is more austere: knowledge in the subjective sense consists of mental dispositions and thought processes, while knowledge in the objective sense consists not of thought processes but of thought contents, the content of linguistically formulated theories (Popper 1978, §X); the conditions of certification and public acceptance are Giere's. Uygun Tunç accordingly classifies Giere's account as a no-subject account and rejects it in favor of individual possession. The reason given is that the notion of knowledge in the objective sense says nothing about where knowledge is located. If no one who understood physical cosmology remained, then, even if the publications were preserved, that the universe is expanding would not be known until someone read them (Uygun Tunç 2023, 12–13).

Among the accounts Giere answers is Knorr Cetina's. As Uygun Tunç summarizes the exchange, the two describe how collaborative experiments produce knowledge in similar terms, as distributed cognition, but differ sharply over whom to count as the epistemic subject of the knowledge so produced (Uygun Tunç 2021, 3–4). Comparing collider experiments at CERN with molecular-biology laboratories, Knorr Cetina argues that the experiments erase the individual as an epistemic subject, defined as the entity through which knowledge is procured. Papers appear under the names of all members in alphabetical order, and the capacity to produce knowledge shifts to the experiment; the conversation the collaboration holds with itself, Knorr Cetina argues, replaces the individual subject with a collective and dispersed one, although no member knows all that is produced. In the molecular-biology laboratories, by contrast, authorship conventions and the assignment of projects keep individual contributions distinct and sustain the individual as the epistemic subject (Knorr Cetina 1999, 236–237, 249–250, 303–304). Knorr Cetina denies that the experiments act as corporate actors, and reports that individuals and institutes persist within them, most visibly while a new experiment is being formed (Knorr Cetina 1999, 268, 271–274). Giere, on the same summary, reads Knorr Cetina as positing the experiment itself as the knower and rejects that extension of epistemic agency as too much of an ontological commitment (Uygun Tunç 2021, 3). Because Knorr Cetina defines the epistemic subject by the procuring of knowledge, however, the thesis can be read as bearing most directly on the first of Fagan's two theses, that groups can be necessary to produce knowledge, and as leaving possession open.

Bird develops a collective account of possession. On Bird's naturalized, knowledge-first functionalism, knowing is not a kind of believing but the state that a properly functioning cognitive system produces. A cognitive system here is a system whose function is to produce factive representational states that guide action and are carried forward into inference and memory; a function is a working that has been selected for, whether by natural selection or as an artifact's. Because functions of this kind can be realized in different materials and arrangements, a group too can know by realizing an appropriately organized cognitive system, and the account does not reduce the attribution proposition by proposition to what individual members know (Bird 2024, §§2–3).

The joint acceptance account, of which Gilbert gives a prominent statement, holds, in Jeff Dunn's summary of the version Lackey discusses, that a group believes that p just in case its members agree jointly to accept p (Dunn 2021). On Dunn's reading, Lackey rejects both views that simply add up what members believe, are justified in believing, or know, and views on which a group's belief, justification, or knowledge floats free of its members'. Against the joint acceptance account, Lackey argues that it cannot account for group lies. Suppose a tobacco company's directors believe that smoking is harmful but jointly accept and announce that it is safe. The company appears to be lying; yet if its board's joint acceptance is all it takes for the company to believe that smoking is safe, the account rules out calling its assertion a lie. On Lackey's diagnosis, as Dunn reports it, the account makes group belief too voluntary and so makes the charge of lying too easy to escape (Dunn 2021). Lackey also rejects group beliefs that are what she calls base fragile, in which the bases of a significant subset of the members' beliefs conflict with the bases of another significant subset. Such a state, Lackey holds, is not always open to epistemic evaluation, and it cannot figure in the group's deliberation about its future actions, since the different reasons call for different plans (Dunn 2021). On Lackey's account, as Dunn quotes it, a group believes that p if and only if a significant percentage of its operative members believe that p and adding together the bases of their beliefs yields a belief set that is not substantively incoherent (Dunn 2021); operative members, in the sense Kallestrup takes from Tuomela, are members with decision-making authority, and so responsible for the content of the group's state (Kallestrup 2024, 571). As Jessica Brown quotes Lackey's account of justified group belief, a group justifiedly believes p if and only if a significant percentage of its operative members justifiedly believe p on bases that, added together, yield a coherent belief set, and full disclosure of the evidence relevant to p, with rational deliberation about that evidence among the members in accordance with their individual and group epistemic normative requirements, would not yield further evidence that, added to those bases, fails to make p sufficiently probable (Brown 2023, 2). On this account, then, a collectivity with no operative members in the relevant domain, humanity for example, has no group belief; an S so specified serves as the reference relative to which a result is new, while possession falls to members, subgroups, or organizations within it. The presupposition of Chapter 1, Section 2, that S's epistemic situation could come to include the knowledge, is then met through those members, subgroups, or organizations. Whether a scientific field has operative members in this sense is not settled here (compare Section 7). Lackey's account of group assertion, unlike her account of group belief, turns on a spokesperson who speaks with the group's authority (Chapter 11, Section 5.1).

Kallestrup occupies another intermediate position: some group knowledge is irreducible to member knowledge but remains anchored in individual features. On Kallestrup's account, a group's conventions can set epistemic standards that its members do not follow as individuals. A criminal jury instructed to disregard hearsay evidence, for example, can know that the defendant is guilty, while none of the jurors knows it, because each has received reliable but misleading hearsay to the contrary. Even so, the jury's knowledge is fixed by the members' states and processes and by the relations in which they stand to one another within the group's structure (Kallestrup 2024, 578–579, 581–582).

Longino's critical contextual empiricism (Longino 2025, §5) places the social at a different point from the accounts above. They ask mainly who possesses knowledge; Longino argues that social organization enters what supports accepting a claim, its warrant, itself. Because evidential relations are mediated by background assumptions, the argument continues, critical interaction among the members of a scientific community, or among different communities, is the only check against the arbitrary dominance of subjective preference in accepting such assumptions (Longino 2025, §5). Under Longino's substantive account, venues for criticism, uptake, publicly recognized standards, and tempered equality of intellectual authority (any perspective has a prima facie capacity to contribute to the community's critical interactions, a standing that can be lost by failing to engage with or respond to criticism) help constitute scientific objectivity rather than merely communicate an already completed private achievement (Longino 2025, §5). Uptake, in this sense, is a requirement on the community: the distribution of belief in it must change over time in a way that is sensitive to the critical discourse taking place within it (Longino 2025, §5). As Dutilh Novaes and Dede report, Longino presents these conditions as features of an idealized community that real communities meet by degree (Dutilh Novaes and Dede 2025, §2). They are not adopted here as a theory-neutral list of conditions for all knowledge. They show one important way in which social organization can enter warrant itself.

Longino also addresses who possesses knowledge. As Wray reports her earlier account, knowers are interdependent individuals, rather than a group that transcends them (Wray 1998, §IV). Eigi explains Longino's conformation as success that admits degrees and varies with the respects assessed, with truth as a limiting case (Eigi 2016, 49). In Eigi's exposition, knowledge is indexed to communities and their projects without making acceptance arbitrary: communities must satisfy conditions of effective criticism (Eigi 2016, 47–49, 138). Where this chapter draws on Longino's social conditions on warrant, it keeps its factive usage.

What these accounts imply for artificial research depends on what A, the agent conducting the research, and S, the society for which knowledge is sought, are. Two combinations are taken up here: an artificial A with a human S, and an artificial A with an artificial S; mixed societies lie between them (Chapter 8, Section 1.2). Where A is artificial and S is a human community, the accounts differ on what it takes for A's result to become S's knowledge: reaching individual knowers in S through evidence or justified dependence, being routed to its operative members, being jointly accepted, being certified by accepted procedures, being integrated into an organized cognitive system, or surviving criticism in public venues and being taken up. On none of them, as their authors state them, is it enough that A generates and publishes an output, and what has to be added differs from account to account (Section 7 asks whether, on an extension of Bird's account, a validated record available through normal channels could suffice); this concerns the third of the judgments in Chapter 1, Section 2, whether the result has entered S's epistemic situation. They also differ on whether A can be among the bearers of that knowledge, which concerns the second, where the achievement is realized. On Uygun Tunç's and Giere's views, a human–machine organization does not know as a whole, and knowledge is located in individuals. Giere, as Matthew J. Brown and Uygun Tunç report his position, restricts belief, knowledge, and agency to the human components, as a theoretical choice defended for systems such as the Hubble telescope, in which humans conduct the research with instruments and computers (M. J. Brown 2015; Uygun Tunç 2021, 3–4). Neither view considers an A that is itself an artificial system conducting research, so whether such an A could be an individual knower is not settled by them; it turns on the accounts of belief in Section 2. Read as a thesis about production, Knorr Cetina's shows how far production can be dispersed across a collaboration, so that no member knows all that is produced, without settling who possesses the result. On Bird's group-functional account, records and retrieval can belong to an organized collective that knows, so the question becomes whether A is organized into such a collective as one of its components. Whether A contributes to knowledge that S comes to possess and whether A is itself a knower, or part of one, are separate questions (Section 2.1; Chapter 1, Section 2). In this case, nothing in these accounts requires A itself to be a knower (Chapter 8, Section 1.1). Other relations can make it matter. If the system's reports are taken as assertions through which humans come to know, accounts of assertion and testimony place conditions on the system as speaker, and on Williamson's knowledge rule it is the system, or the group for which it speaks, whose knowledge the rule concerns (Chapter 11, Section 5.1); and where the achievement sought is realized in the system itself, as with a skill, the system is its first bearer.

Where S is itself an artificial society, the second question, whether A is itself a knower, can no longer be set aside, and it takes a wider form: whether artificial members, A among them if it is one, can be knowers at all. Both roles of S presuppose that its epistemic situation could come to include knowledge of the kind sought (Chapter 1, Section 2), and that situation now consists of artificial members, their organization, and their records; the question therefore arises before any member of S knows the result, and whether or not A is one of its members. Whether a result has become S's knowledge then turns, on accounts that require a bearer, on whether its bearers can be found among those artificial members or their organization, and, on Giere's adaptation of knowledge without a knowing subject, on certification within S. On individual-possession, member, and relational accounts, the question returns to whether artificial members believe the result (Section 2), and on relational accounts also to whether epistemic relations hold among them. On joint-acceptance accounts, it becomes whether they can take part in joint acceptance; on Bird's account, whether the artificial members that make up the society together realize a properly organized cognitive system; on Giere's adaptation of knowledge without a knowing subject, whether certification by artificial members counts as certification, a question that sharpens where they are instances of one base system, since instances of the same base can err in the same places, so that their certifications of one another may add no independent check (Chapter 11, Section 3); and on Longino's account, whether their interaction involves the diversity and the public standards of criticism that a knowing community requires, which multiple instances or a majority vote do not supply by themselves. Giere's restriction of agency to humans, taken as a premise, would leave such a society with no member to whom knowledge could be attributed, but since Giere defends it as a theoretical choice for systems in which humans conduct the research, it does not settle this case. Chapter 8, Section 1.3 develops these differences.

Finally, the social dependence of knowledge production that these accounts describe does not imply that truth is created by agreement. To keep the two apart, this book distinguishes at least three claims that differ in strength:

1. inquiry causally depends on socially inherited language, concepts, instruments, testimony, labor, and records;
2. evidential standing, objectivity, or knowledge is partly constituted by organized social practices;
3. whatever a community accepts is thereby true.

Individualist epistemologies can accept the first. Accounts in Table 2.2 on which social organization enters warrant or possession, such as Longino's critical-contextual account, defend versions of the second. Neither entails the third, which no account discussed in this chapter holds. Social recognition can affect warrant and possession without making a false proposition true. Evidence, failed interventions, resistant materials, and other world-involving constraints continue to matter. For artificial research, the distinction guards against reading agreement among artificial members, or a majority vote across them, however they were made, as making a result true; what the second claim leaves to be asked is whether an artificial society's organized practices could partly constitute warrant (Chapter 8, Section 1.3; Chapter 11, Section 3).

Selecting S settles the relativity of novelty; it does not settle possession, which the working characterization leaves to the accounts compared here. A claim that a result has become S's knowledge can therefore be assessed only when it reveals enough of its account. Chapter 8, Section 1 returns to the roles of S when it separates A from S, asks about the bearer where success is at issue, and distinguishes cases by whether S is human, mixed, or artificial. Within this chapter, Section 7 tests the accounts compared here on a single case, and Section 9 states them as requirements; Chapter 11 takes them up for certification and criticism.

7. The unread artificial discovery

This section tests the preceding accounts together on one case, in which the production of a result and its possession by S come apart. The result is produced by an artificial A, which may itself know it on an account of artificial belief (Section 2), and S is human society, no member of which has read it. The case is then strengthened: the result passes recognized validation (Section 5) and enters retrieval, use, and correction, so that of the relations distinguished in Section 6, the one chiefly still missing is a member's acquiring an attitude toward the result. The accounts compared there divide over whether that absence matters. The verdicts are stated first for a human S and then carried over to an artificial or mixed S.

The case, deliberately difficult, is this. An artificial system derives a true proposition previously unknown to human society, preserves strong grounds and a complete record, and deposits the result in a public database. No human has read it, believed it, or acted on it. Does human society know the result?

The physical inscription exists. The artificial system may itself know the result under an appropriate account of artificial belief and warrant. It may also have the knowledge on an account that does not require belief (Section 2). Humphreys, for example, modifying a reliabilist definition for automated methods, proposes that an instrument has the knowledge that F if and only if it contains a representation of the entirety of F, the representation holds of the target, and a reliable process, one that produces a high proportion of accurate representations, formed it. Humphreys adds that the definition makes good sense where the representations are accessible to humans but becomes problematic where they are opaque and their accuracy is unknown, and that it may then be necessary to fall back on something like Burge's entitlement approach to computer-assisted mathematics, on which some sources, such as normal visual perception, are taken as primitive epistemic authorities (Humphreys 2020, 24–25). Symons and Alvarado, writing earlier, argue against carrying that approach over from computer-assisted proof to computer simulation. As Winsberg and Alvarado set out the argument in their encyclopedia entry on computer simulations in science, a simulation is not what Burge calls a transparent conveyer, one through which the warrant for its results is of the same kind as the warrant for the operations it carries out, since features of simulations let warrants of another kind, a posteriori ones, enter the pipeline; a simulation must have its own warrants and its own sanctioning, pursued independently of the sanctioning of the methods that enabled its construction, rather than being trusted by default (Winsberg and Alvarado 2026, §2.3). Symons and Alvarado write about simulation, not about learned representations, but applied to Humphreys's proposal their account would not let an entitlement supply what unknown accuracy leaves missing: ascribing the knowledge to an instrument whose representations are opaque would rest instead on evaluations of the instrument itself, such as comparison with real-world data, the kind of validation the case acquires when it is strengthened below. On such a definition the depositing system could have the knowledge as an instrument has it, without belief, and whether human society knows the result would remain a separate question. The record is a candidate contribution to human knowledge. Neither the inscription, nor the system's possible knowing, nor the record's standing as a candidate establishes by itself that human society already knows. A file disconnected from the society's epistemic organization could remain indistinguishable, from the perspective of human inquiry, from an indefinitely large collection of unnoticed truths and falsehoods.

The limited conclusion is:

Bare inscription or public deposit is not a theory-neutral sufficient condition for human social knowledge.

This statement is weaker than the claim that storage can never partly constitute knowledge. Strengthen the case. The record has passed validation by procedures that S's institutions recognize. Its provenance is authenticated. It is indexed in canonical infrastructure, retrieved automatically when relevant questions arise, used in inference and action, and connected to mechanisms that propagate correction or withdrawal. No human has yet read the proposition itself.

Now the philosophical verdict divides. An individual-possession account can say that no human yet knows, although humans are well positioned to know. A member-based account such as Lackey's reaches the same verdict for groups. Applied to human society as a whole, which has no operative members in the relevant domain, Lackey's account yields no group belief (Section 6), so the verdict concerns groups within it that have such members: since none of their operative members believes the result, none of these groups yet believes or knows it. A joint-acceptance account asks whether the members have jointly committed to letting the result stand as the group's view. A distributed-cognition account can classify the infrastructure as a genuine cognitive resource without attributing knowledge to the whole. An account that allows knowledge without a knowing subject detaches the result's status from anyone's attitude, but Giere's version does not count a claim as scientific knowledge until the relevant community has publicized and accepted it (Giere 2007, 318). Applied to the present case, that condition suggests that the unread result would not yet qualify, although the same passage lets certification by accepted procedures support a claim to know, which the recognized validation in the strengthened case may approach. Popper's own example is close to the present case: a series of logarithm tables computed and printed by a computer, sent to libraries and never read, contains what Popper calls objective knowledge, since what makes a thing a book is its possibility of being understood, not its actually being read (Popper [1972] 1979, 115–116). On that formulation the deposited record already belongs to objective knowledge, true or false; but because objective knowledge is there defined as independent of anyone's claim to know (109), the formulation does not answer the question this section asks, whether human society knows the result. Extended to a record no one has ever read, Bird's group-functional account might say that the result is already part of the society's knowledge because the organized system can retrieve, use, and revise it; Bird's own case, discussed below, concerns retention instead. A critical-contextual account, on Longino's definitions, gives a more determinate verdict.

In the definitions Büter reproduces, accepted content counts as knowledge for a community when it conforms to its intended objects well enough to enable members to carry out their projects and is epistemically acceptable there. At a given time, acceptability requires support by data evident to the community in light of reasoning and background assumptions that have survived criticism from as many perspectives as are available to it, in a community characterized by venues for criticism, uptake, public standards, and tempered equality of intellectual authority (Longino 2002, 135–136, as quoted in Büter 2012, 84). Applied to the present case, Longino's definitions imply that the unread result is not yet the society's knowledge, whatever the quality of its automated validation, since no member has accepted it. It would further follow that whether that validation could contribute to acceptability once the result is taken up depends on whether the validation itself has been exposed to criticism under Longino's conditions.

The basic and the strengthened case differ as a dormant record differs from a live epistemic capacity. The former merely persists; the latter is integrated with retrieval, competent use, dependency tracking, and correction, and among the relations of Section 6 it is these that the strengthening added. The distinction names what it takes to make a record a live capacity. Whether it also marks the boundary of knowledge, however, differs from one account to another. Bird's own case of a journal article that no one any longer remembers is close to the present one in that no one is currently aware of the result. Dr N. publishes a peer-reviewed result; everyone who read it dies or forgets it; years later another scientist finds it through the indexes. Bird holds that the result remained part of the scientific community's knowledge throughout, because it stayed available through normal channels to those who needed it, and Bird defends this as genuine knowing that was causally isolated and inactive (Bird 2024). On that view a validated and indexed record can remain social knowledge while dormant. Lackey challenges the claim that the community knew Dr N.'s result all along. Kallestrup, for whom irreducible group knowledge remains anchored in individual features, sides with Lackey here and sets out an objection inspired by her discussion: the familiar distinction between knowing and being in a position to know pertains to the difference between information that has been accessed and information that is merely accessible, so that a group none of whose members has accessed a result is more naturally described as only in a position to know it. In Kallestrup's own argument, because Dr N.'s results are stored in no one's memory or an extension of it, the community does not even dispositionally believe them and has only a disposition, or is only in a position, to know them, and storage in a library or on a journal website does not suffice for knowledge (Kallestrup 2024, 583–585). The unread artificial discovery thus enters an existing dispute, though at a different point. In Bird's case the result had passed peer review and been read before it became dormant, so the case concerns retention. Whether a record that no member has ever read or accepted can become social knowledge is a further question, and the strengthened case differs from Bird's in exactly this respect. On Bird's view, the dormant/live distinction does not mark where a community's knowledge ends. On a reading of Lackey's account, even a live capacity leaves such a group only in a position to know until operative members believe the result. An account on which retrieval, competent use, and correction constituted possession would treat the distinction as the point at which an epistemic resource becomes knowledge.

The dispute turns on how knowledge is tied to action. A second objection, which Kallestrup attributes to Lackey, combines two principles. One is that if an agent knows that p, it is epistemically rational for it to act as if p, a principle Lackey accepts in this direction; the other is what Lackey calls the Group/Member Action Principle, on which a group performs an act only if at least one of its members performs some act that causally contributes to it. The worry is how a group could act on knowledge that none of its members has. Applied to a community none of whose members is aware of a result, Kallestrup concludes, it becomes a mystery how the members could contribute to an act of the community in a way that bears non-accidentally on the result (Kallestrup 2024, 572 and n. 15, 585–586). Bird, who illustrates the objection with a case adapted from Lackey, in which a leading expert asserts on a mere hunch what Dr N.'s forgotten paper had established and the field acts on the assertion, replies that the connection holds only for action based on the knowledge in question: one can know that p and still act irrationally when that knowledge plays no part in the action, so the community's irrational action is no objection (Bird 2024). Lackey also presses, as Kallestrup reports, a case meant to show that Bird's account cannot distinguish what a group knows from what it should have known. Three officers in one police unit each know one fact: that a boy has been reported missing, what he was wearing that morning, and that a boy so dressed was seen walking with an adult man in a nearby neighborhood. On Lackey's reading, Bird's account implies that the department knows that the boy seen is very likely the missing one, yet no single officer seems accountable for failing to act on it. Avram Hiller and R. Wolfe Randall modify the case: each officer enters what he or she knows into a central system, which collates the entries and prints the conclusion once they fit together. Since every officer then has easy access to the printout, they find it intuitive to hold the group accountable for failing to act, and they argue that the group knows, because its structure is designed to access information of that type and is functioning properly. Kallestrup replies that the department in the modified case may be accountable because it should have known the conclusion, not because it knows it: knowledge entails dispositional belief, and a non-biological repository would sustain such a belief only if it served as a functional analogue of the way individuals store and retrieve information in memory (Kallestrup 2022). The automated variant comes close to the strengthened case. Kallestrup's reply asks how the group itself could dispositionally believe the conclusion: without an account of that, collation would at best give the group a disposition to believe it, and the natural candidate he notes, an extended cognitive process at the group level, is one that Hiller and Randall consider but do not adopt (Kallestrup 2022). On this reply, as on the reading of Lackey's account given above, automatic collation and use would not by themselves make the result the group's knowledge unless the group's integration of the record amounts to such a process; the reply adds that a group whose system collates the result can still be answerable for what it should have known.

The verdicts above concern a human S. Carried over to an artificial or mixed S whose artificial members retrieve the record automatically and integrate it into their inference and action, the accounts as set out above do not all move in the same way. On an individual-possession account, the question is no longer whether anyone has read the result but whether an artificial member's automatic retrieval and use amounts to its believing the result, which returns to the accounts of belief in Section 2. On Lackey's member-based account, it becomes whether the society's artificial operative members believe the result, and whether their retrieval counts as accessing the information rather than its being merely accessible; the automated variant of the missing-child case does not decide this, since the collating system there was not among the members. Since operative members are, on this account, those with decision-making authority in the relevant domain, which artificial members are operative depends on how such authority is distributed within an artificial S, a matter Chapter 12, Section 5 takes up. On a joint-acceptance account, it becomes whether artificial members can take part in joint acceptance, and on Giere's version of knowledge without a knowing subject, whether certification by artificial members counts as certification, a question that sharpens, for the reason given in Section 6, where they are instances of one base system (Chapter 11, Section 3). On Longino's definitions, acceptance by members could be met by artificial members if they can be members of a knowing community in Longino's sense, with the diversity and public standards of criticism that Chapter 8, Section 1.3 discusses; whether the result has been accepted would then turn on what the artificial members do with it, not on whether any human has read it. Bird's group-functional verdict does not change in this way, since it already turned on whether the organized system can retrieve, use, and revise the result rather than on any member's awareness; for an artificial S, the question remains whether the artificial members together realize a properly organized cognitive system (Section 6). Since on Bird's account a function is a working that has been selected for, it also matters whether that organization was selected or designed for producing factive states that guide action rather than, for example, for passing checks or winning approval; as with Simion's account in Section 4.1, what selection actually rewarded can diverge from the stated objective.

The producing system may itself be a member of an artificial or mixed S. If it is, and if it knows the result on the adopted account of belief, then on an individual-possession account a member of S already knows, and the result may have entered S's epistemic situation as soon as it was produced. On Lackey's account, the system's belief counts toward a group's belief only if the system is among the group's operative members, and the group believes the result only if a significant percentage of them do. On Longino's, the system's acceptance is at most one member's, and acceptability requires that the result survive criticism from the other perspectives available in the community. On Bird's, the verdict still turns on the organized system as a whole. It follows that the group-functional account is not the only route to crediting an artificial society with knowledge. Even on an individual-possession account, if what counts as a knowing individual changes, a result that no human has read can no longer be said to be known by no member. That requires, however, that the machine be a knowing member on the adopted account of belief, not a member in name only (Section 2).

At the limit of the basic case, where the record is so disconnected that no member of S, human or artificial, can observe, retrieve, assess, or use the result, the accounts above that answer whether S knows it largely agree that it does not yet; Popper's objective knowledge, as noted above, answers a different question. The basic case itself, a public deposit, need not reach that limit, since a deposit is available through normal channels. On Bird's account, however, knowing is the state that a properly functioning cognitive system produces (Section 6), and the forgotten journal article had passed peer review; extended to a record no one has read, the verdict that such a record can be knowledge while dormant therefore fits the strengthened case, whose record has passed validation that S's institutions recognize, better than a bare deposit. What remains open is only whether such a wholly disconnected true record is to be called knowledge in some sense, and however that is settled, what has to be built for the result to enter S's epistemic situation is the same: routes by which it can enter validation, inference, action, and correction. These need not lie within the system that conducted the research; they may belong to the infrastructure between it and S, or to S itself. The boundary case nonetheless remains philosophically important, because it prevents upload from being equated with society knows.

In the strengthened case, where such routes exist, the accounts divide, and the division bears on what a project claiming that the result has entered S's epistemic situation has to build. On Bird's account, extended as above, the routes may already suffice for the result to have entered S's epistemic situation. On Lackey's, such a project would further have to bring the result to members of S and have it believed by them, and, where entry is claimed through a group's knowing it, to the operative members of a group within S that has them in the relevant domain; on Longino's, to have it accepted by members after criticism under the account's conditions. What such a project has to build therefore depends on the account of possession adopted, which is why, as Section 6 noted, a claim that a result has become S's knowledge can be assessed only when it reveals enough of its account. Whether the activity is research for S does not depend on whether the result meets that account's conditions of entry, as the next paragraph notes.

The case also shows concretely two distinctions drawn in Chapter 1. If the artificial activity is organized toward producing knowledge new to human society, it can be research in the sense of the working characterization before any human acquires the result, since attempt separates research activity from its success (Chapter 1, Section 3). The examples there, such as an inconclusive experiment or a proof strategy that collapses, separated the two through the absence of the achievement. Here the achievement may be present, only not yet in S, so that success itself divides. Of the three judgments of Chapter 1, Section 2, the first, whether the activity is research for human society, can hold. The second, whether knowledge new to human society has been realized somewhere, can hold as well, in the system itself if the system can know on the adopted account of belief. The third, whether the result has entered human society's epistemic situation, depends on the adopted possession account and may not yet hold. For an artificial or mixed S, the judgments divide in the same way, unless the producing system is itself a member whose knowing already counts on the adopted account, and the third turns on the accounts as carried over above.

8. Epistemic change before and around knowledge

Research often ends short of knowledge, and what it achieves need not be knowledge even when it succeeds. This section asks how such outcomes can be described and evaluated without either counting only knowledge or counting every change as an improvement. The literature this section draws on has mostly pursued this question in one narrower form: whether such outcomes count as progress. Accounts of scientific progress in particular disagree largely over whether only the accumulation of knowledge counts as progress or other outcomes count as well. The section follows the debate in this narrower form, because doing so shows what an evaluation of such outcomes needs to keep apart.

It first separates research activity from success at knowledge (Section 8.1). Section 8.2 takes up the case in which research falls short of knowledge: changes before knowledge, chiefly the making explicit of ignorance, together with formulating questions, acquiring evidence, and changing credence, and whether they count as progress; the arguments it draws on concern epistemic progress in general, not only in the sciences. Section 8.3 takes up, together with the accounts of scientific progress, the case in which the outcome is not knowledge itself: achievements around knowledge, such as a false theory's approximate truth and problem solving. Understanding belongs here too on views that do not take it to be a kind of knowledge; whether it is one is taken up in Chapter 3, Section 3. The section ends with what this implies for evaluating artificial research.

8.1 Research can be well conducted without reaching knowledge

Section 7 separated research from success in a case where knowledge may already have been realized but has not yet entered S. This subsection takes up the case from which Chapter 1, Section 3 began, in which the knowledge sought is not reached at all: a central claim turns out false, or a result remains inconclusive. Because attempt separates research activity from its success, such an investigation does not thereby cease to have been research, and whether it was research can be judged without waiting for the result, since the judgment turns not on whether the activity succeeded but on how it was organized (Chapter 1, Section 3). Independence from success does not, however, mean that the judgment is settled when the activity is undertaken: an activity can be reorganized into research along the way, and its organized direction may become clear only from later records.

For the case where the target is knowledge, this book therefore keeps at least the following six evaluations distinct:

1. whether an activity is research for S in the sense of the working characterization;
2. whether it is responsibly or competently conducted relative to the available situation;
3. whether target knowledge new to S has been realized, in whatever bearer, A included, or, on accounts that allow knowledge without a knowing subject, in a certified result;
4. whether that knowledge has entered S's epistemic situation, which the accounts of possession in Section 6 answer differently;
5. whether failure at the target produces a collateral epistemic contribution;
6. whether the contribution is significant or worth its costs.

The first, third, and fourth are the three judgments of Chapter 1, Section 2.

A well-conducted investigation can fail at its target. A poorly conducted one can reach a true answer by luck. A failed experiment can reveal a limitation in a method, an auxiliary assumption, or the original problem formulation. Whether that collateral result is knowledge requires its own assessment; whether it is important requires a further relation to purposes and affected parties. As the term is used here, whether the result is knowledge depends neither on whether it came from research nor on whether it is important.

8.2 Changes before knowledge: explicit ignorance and whether such changes are progress

Failure to reach knowledge need not be a mere absence of progress. A society can be wrong about a matter, unaware that the matter is unresolved, aware of a question but unable to identify plausible answers, or able to specify both a question and the evidence still needed to answer it. These are materially different epistemic positions even when none yet contains the target knowledge.

Wilholt distinguishes, within total ignorance, conscious ignorance—roughly, the questions whose answers a subject recognizes that it does not know—from opaque ignorance, what a subject does not know that it does not know. The transition from error or opaque ignorance to conscious ignorance is what Wilholt calls Socratic progress. Conscious ignorance can itself require substantial conceptual and epistemic competence (Wilholt 2020, 2–6). Within conscious ignorance, Wilholt reserves the label deep ignorance for cases in which a subject is aware that no plausible candidate answer is available, and identifies these cases with what Bromberger called a p-predicament: a subject is in a p-predicament with regard to a question if and only if, on the subject's views, the question admits of a right answer, but the subject can think of no answer to which, on those views, there are no decisive objections (Wilholt 2020, 9–10). Wilholt argues that such questions hold particular attraction for researchers, since overcoming one promises to open new questions and connections, but concludes only that being perceived as deep favors a question's inclusion in the research agenda: such questions can be passed over, for example when no available method offers an approach to them (Wilholt 2020, 10–11). Wilholt's account of Socratic progress gives a precise example of epistemic improvement that need not yet be knowledge of the answer. It also helps connect the achievement analysis in this chapter to the later account of how questions are understood and made tractable (Chapter 4, Section 3).

The scope of this point is limited. It does not make every unanswered question a valuable research contribution, identify ignorance with knowledge, or show that articulating a question is sufficient for research. Significance, tractability, novelty, and the quality of the articulation remain separate matters.

What this section takes from Wilholt's account is a contrast. On one side is missing knowledge that is represented nowhere. On the other is a limitation that is represented in some form, such as a note in a record, an error estimate, or a stated question, and that can therefore guide attention, inquiry, and criticism. Conscious ignorance need not take the form of a categorical belief that one does not know; what matters is whether the limitation is represented in some such form.

Working records can keep such limitations explicit. In a 1972 design study for a laser-interferometric gravitational-wave antenna, for example, Weiss estimated the instrument's noise sources one by one and recorded where the needed knowledge was missing: the seismic spectrum had not been measured above 100 Hz, and no data existed on high-frequency gravity gradients near the earth's surface (Weiss 1972, 64, 67). Dismantling a claimed discovery can also begin by making such a limitation explicit. After a 2009 report associated the retrovirus XMRV with chronic fatigue syndrome, a multi-laboratory working group, before testing clinical specimens, first checked whether the assays could detect the virus in blinded samples to which it had been added. It thus first established what the assays could and could not show. A later blinded multicenter study, whose design all participating laboratories had agreed on in advance and whose authors included the original claimants, found no association (Simmons et al. 2011a; Simmons et al. 2011b; Alter et al. 2012).

Some fields have made such limitations a sustained object of study. Knorr Cetina, for example, reports that the collider experiments under study cultivated what Knorr Cetina calls negative knowledge and distinguishes from non-knowledge as knowledge of the limits of knowing, of the mistakes made in trying to know, of what interferes with knowing, and of what one is not interested in and does not want to know: one physicist described the systematic error, the uncertainty of a measurement that comes from the apparatus and the analysis rather than from chance fluctuation, as a way of measuring the group's ignorance. When some forty-five competing theoretical descriptions of the proton's internal structure gave different inputs to an analysis, the spread among the resulting values was entered into the result as a theoretical systematic uncertainty rather than left as a disagreement among theories (Knorr Cetina 1993, 269–273). Knorr Cetina presents this as a preference of that culture, which fields such as molecular genetics do not share (Knorr Cetina 1993, 270–271). Chapter 6, Section 4.3 takes up the molecular-biology side of this contrast.

The distinction between missing knowledge that is represented nowhere and an articulated limitation has a direct consequence for artificial research. Evaluation by the number of questions closed or answers emitted can favor a system that closes questions before the evidence settles them. Such a system can fill its outputs with plausible claims while making uncertainty, missing evidence, incompatible candidates, and unformulated problems harder for those who rely on them to see. Conversely, a system may improve the research position of S by showing that a seemingly settled answer is unsupported, by identifying which alternatives remain live, or by making a previously inarticulable limitation askable. An artificial research system may therefore need to retain open questions, reasons for suspension, failed operationalizations, and missing-evidence relations as first-class epistemic states rather than as empty slots left after answer generation. Keeping such states in the system's own record is not yet a change in S's position. For the move to conscious ignorance to be made in S, the represented limitation must enter S's epistemic situation: where S is human, it must reach human members in a form they can take up as a question; where S is artificial, its artificial members must be able to treat it as an open question in their inquiry. Chapters 4 to 6 examine how such states are formed and transformed in inquiry; the present claim is only that they can matter to epistemic progress. Whether they constitute progress or only promote it divides accounts of progress, as the end of this subsection shows.

Williamson's arguments bear on how complete a system's record of what it does not know can be made. Williamson argues that almost no nontrivial condition is such that, in every case in which it obtains, one is in a position to know that it obtains, and in particular that one can fail to know something without being in a position to know that one fails to know it (Williamson 2000, 13–14, 107). In Williamson's simple models of creatures that store information, a creature that cannot survey what it has stored can fail to know something without knowing that it fails to know it; in one of them, retrieval from memory can take any length of time, so the creature cannot tell a missing memory from one not yet delivered. Williamson describes such a failure as one of self-knowledge rather than of rationality in any ordinary sense (Williamson 2000, 226–227, 233–234, 316–317). What does the work here is less that the creature cannot survey its store at once than that retrieval can take any length of time, so that a missing memory cannot be told from one not yet delivered. Applied to an artificial system that meets this condition, these arguments, if they hold, imply that its record of its own open questions and missing evidence cannot be guaranteed complete by the system's own checks alone. A configuration that can confirm that it has finished examining a finite store is not covered by this argument as it stands. Even then, however, what such a check confirms concerns only what is represented in the store; an absence represented nowhere, like the opaque ignorance above, does not show up when the examination of the store is finished.

As used here, the process vocabulary does not turn every change into an improvement. Ross argues that an intellectual community can progress before it knows an answer: it may formulate a worthwhile question, acquire relevant evidence, or increase its confidence in the correct answer. The reason given is that events widely recognized as progress occur while a community is still divided. After Galileo reported telescopic evidence that the surface of the moon is not smooth, the community neither believed nor knew it; yet to say that the community made no progress would, on this view, be too large a revision. Ross also rejects the view that such events count as progress only because they facilitate later knowledge, since by that standard even making coffee for a research group would count as progress (Ross 2021). Harris objects that formulating questions and collecting evidence facilitate progress rather than constitute it: a community would not progress merely by formulating many questions or by collecting evidence fruitlessly, although increased confidence in true propositions might be accommodated by a credence-based version of a truth-centered account (Harris 2024, 142–143). Carried over to Wilholt's account above, Harris's point would apply to the move from opaque ignorance to recognizing a question that Wilholt calls Socratic progress, counting it as facilitating rather than constituting progress. Whether it applies to giving up a false belief is less clear, since Harris's concession about credences may cover that case. The claim above about Socratic progress is therefore made on Wilholt's account rather than on every account of progress.

Without settling the theory of progress, the dispute still yields a constraint. Question formation, evidence acquisition, credal change, and acknowledged ignorance are candidate epistemic transformations. Whether a particular change constitutes progress depends on its relation to a worthwhile target, a correct answer, error exposure, or an improved capacity for inquiry. More confidence, less uncertainty, more information, and more questions are not independently sufficient.

8.3 Achievements around knowledge and accounts of progress

This subsection turns from changes before knowledge to achievements around it, beginning with historically successful but false theories. This chapter takes propositional knowledge to be factive: what is known is true (Section 3). A false theory, however successful, is therefore not itself known, and the question becomes how its achievement is to be described. We may identify true restricted claims that were known, take the parts of a false theory that were responsible for its successes to be approximately true (Chapter 6, Section 3.2), distinguish literal belief from practical acceptance, appeal to approximate truth or empirical adequacy, or treat understanding and problem-solving ability as achievements not exhausted by knowledge accumulation; this subsection takes up several of these options, among them Bird's restricted claims, Niiniluoto's truthlikeness, Laudan's problem solving, and Dellsén's understanding. Factivity does not decide among these analyses. It fixes only how this chapter uses the term propositional knowledge; how to describe the wider epistemic achievement of such theories is left to those analyses and to later examination.

Several of these options have been developed as accounts of scientific progress. Dellsén's survey of that literature distinguishes four accounts, each of which places its own type of cognitive achievement at the heart of progress: increasing truthlikeness, increasing problem-solving effectiveness, accumulating knowledge, and increasing understanding, the last being Dellsén's own account, termed noetic (from the Greek for understanding). Dellsén adds that the most plausible account may be pluralistic, counting more than one type of achievement (Dellsén 2018, §1). On this survey, progress is analyzed not as one more achievement alongside knowledge and understanding but as some type of achievement's being increased, accumulated, or more nearly achieved. In this book's terms, what accounts of progress dispute is which achievement's increase counts as progress. As Chapter 1, Section 1 noted, the book's working sentence characterizes the orientation of a research activity and is not meant to serve as a theory of scientific progress. In particular, characterizing research through epistemic achievement does not commit the book to Bird's epistemic account, on which progress consists in the accumulation of knowledge (Dellsén 2018, §§1, 4); Chapter 3, Section 4 takes up understanding-based views, including Dellsén's, to examine whether, and how, understanding and knowledge differ as targets and measures of progress. Dellsén suggests that there is a close conceptual connection between progress and the aim of science: roughly, X is the aim of science just in case science makes progress when X is increased, accumulated, or more nearly achieved (Dellsén 2018, §6). If that connection holds, it follows that adopting one of these accounts would amount to establishing the kind of global aim that Chapter 1, Section 1 declines to establish. The paragraphs that follow draw on these accounts for distinctions that bear on evaluating artificial research, without choosing among them. The accounts were proposed for scientific progress, whereas the working characterization is not restricted to the sciences; the distinctions drawn from them are used here for research more generally, and whether a given account's measure fits a field outside the sciences is a further question.

Historically successful but false theories pose a difficulty in particular for Bird's account, on which progress consists in the accumulation of knowledge, since a false theory cannot itself be accumulated knowledge. Yet a sequence of false theories may approach the truth, solve increasingly important problems, or improve understanding. Truthlikeness accounts, on which progress is a theory's coming closer to the truth, represent one route through this difficulty. Others are represented by problem-solving accounts, on which progress is an increase in problem-solving effectiveness; noetic accounts, on which it is an increase in understanding; and functional accounts, described below, on which it lies in defining as well as solving research problems (Niiniluoto 2024, §§3.2, 3.6; Nickles 2017, §1.5; Dellsén 2021, 11259–11262; Shan 2019). Laudan's problem-solving account shows how such a route works. As Dellsén notes, the account is explicitly designed not to require that progress be factive (Dellsén 2025, 2128), so it can count as progressive a succession of theories each of which may later prove false. Shan's functional account revises the problem-solving approach it attributes to Kuhn and Laudan by counting the definition of research problems, as well as their solution, among the constituents of progress. On that account, Darwin's theory of pangenesis marked progress in the study of heredity by providing a useful, if imperfect, framework of problems and hypotheses, although the gemmules it posited do not exist (Shan 2019, §§2–4; see Dellsén 2021, 11266). These accounts are rival analyses of progress, not interchangeable repairs to a knowledge definition.

Returning to the truthlikeness accounts named first above, Niiniluoto's version analyzes progress as increasing truthlikeness, measured as closeness to the most informative true answer to a cognitive problem, so that a step from one false theory to another can be progressive. Niiniluoto distinguishes real progress, which can go unrecognized while that answer is unknown, from estimated progress, judged by expected truthlikeness on the available evidence and open to revision as the evidence changes (Niiniluoto 2024, §3.5). Bird, who holds that progress is the accumulation of knowledge, had objected that accounts defining progress by truth or truthlikeness alone would count as progress even a belief that happened to be true or truthlike, reached by irrational methods without justification. Niiniluoto replies, following Cevolani and Tambolo, that estimated progress already gives justification a place: justification is not part of what real progress consists in, but estimated progress is relative to evidence and justification, and claims of real progress can be justified by appeal to expected truthlikeness. Niiniluoto also raises as a further problem whether Bird's way of handling false theories, by treating an approximately true theory as the fully true claim that the theory holds approximately, can distinguish a sequence of theories that approaches the truth from one that recedes from it (Niiniluoto 2024, §3.6). Brought back to the choices listed at the start of this subsection for historically successful but false theories, the problem bears on the first of them, identifying true restricted claims that were known, since Bird's claims that a theory holds approximately are claims of that kind.

The separation drawn in Section 3, between being true and being shown to be true, bears on an objection to factive accounts of scientific progress. Laudan argued, as Dellsén reports, that factive accounts make progress epistemically inaccessible to scientists themselves: since even the most empirically successful theories might turn out to be false, scientists could not tell whether later theories are more faithful to the facts than their predecessors, and progress so inaccessible cannot guide rational decision making in science (Dellsén 2025, 2128, 2140). As Niiniluoto puts it, Laudan repeats the classical skeptical argument that knowing the truth is a "utopian" task (Niiniluoto 2024, §2.4). Dellsén replies that a rational choice requires only rational credences about how far each available course of action would promote progress, in the sense defined below, not access to whether it does, so a factive standard of progress can still guide choice (Dellsén 2025, 2128, 2140–2141).

Niiniluoto's distinction between real and estimated progress, above, can be read as answering the same objection from within the factive framing retained here, and Reichenbach as having earlier met a difficulty of the same form from outside that framing. Reichenbach treated truth as an idealization of a high weight and knowledge as a system of posits rather than of true propositions, and argued that induction can be justified without knowing whether it succeeds: if limiting frequencies exist, persisting with induction and correcting its estimates must eventually bring the estimates within any chosen distance of those limits and keep them there, although, as Reichenbach conceded, convergence may come too late to be observed within a human lifetime (Reichenbach 1938, 187–191, 348–351, 360–362, 403–404). Reichenbach granted that this argument does not single out induction, the rule of taking the observed frequency so far as the estimate of the limit, since any rule that adds to the observed frequency a correction that shrinks to zero as observations accumulate yields estimates that approach the same limit, and preferred induction among such rules for the practical reason that it carries the smallest risk (Reichenbach 1938, 354–356). If so, a guarantee of this long-run kind does not by itself settle which method to follow at a given time, the guidance whose absence Laudan's objection pressed and that Dellsén's appeal to rational credences addresses. Where assurances about an artificial system's learning or updating rule take this long-run form, the same reasoning suggests that they do not by themselves settle what the system is warranted in concluding at a given time.

Comparisons of truthlikeness also depend on how a problem is framed. Oddie and Cevolani report an argument, due to Miller, that on likeness accounts of truthlikeness, which measure it by a theory's similarity to the truth, Niiniluoto's among them, the ranking of two theories can reverse when the same theories are stated with different primitive predicates. In their example, weather can be described with hot, rainy, and windy, or with hot, minnesotan (hot if and only if rainy), and arizonan (hot if and only if windy). Each vocabulary can be defined from the other, yet a theory that makes fewer errors than another when counted in the first can make more when counted in the second. Of the two responses they describe, one holds that some predicates, such as hot and rainy but not minnesotan, carve reality at its joints; estimating truthlikeness then depends on estimating which predicates do so. The other, which denies that the restated theories say the same thing, makes theories framed in different concepts hard to compare (Oddie and Cevolani 2022, §1.4.3). This suggests that estimated progress in Niiniluoto's sense is relative to a cognitive problem and the concepts in which it is stated, and revising those concepts (Chapter 4, Section 4; Chapter 5, Section 3; Chapter 6, Section 5) can change what is being compared. Applied to an artificial system that also chooses its representations, Miller's argument suggests a further point. If the system itself selects the vocabulary in which its results are stated, the same results may be rated closer to the truth under the vocabulary selected. Whether revising a representation captured the target better or changed the yardstick of evaluation then needs to be asked separately. On the first response above, estimating the former also requires estimating which vocabulary carves reality at its joints.

A further distinction concerns temporal contribution. A dataset, formalism, instrument, failed experiment, or classification may constitute an epistemic advance, promote a later advance, or merely be rationally expected to promote one when the work is undertaken. Dellsén gives the second and third predicates a precise form. On this account, an episode promotes progress to the extent that the chance-weighted amount of progress across the developments that could follow it exceeds what it would have been had the episode not occurred; an agent is rational in expecting an episode to promote progress to the extent that the same comparison, made with the agent's own coherent credences, favors the episode. In Dellsén's example, a botched experiment that leads scientists to a successful rerun has not promoted progress if running it correctly the first time would have led further (Dellsén 2025, 2131–2135). These predicates can diverge from one another and from the results that in fact follow. Kept apart, they give a useful artifact no automatic credit for every later result in its causal future, and they do not make a responsible attempt irrational only because unforeseen events prevent its later use.

Consider representing the results of an artificial research system by a single progress score. That requires deciding how much each kind of change counts: new knowledge, an eliminated possibility, an improved method, and so on. The accounts of progress above differ on exactly this, so a claim that a single score represents the progress of the results as such already chooses among them. A proxy that measures only a limited aspect, or a value combined for a provisional decision, carries no commitment of that kind. An evaluation of the system's results that leaves that choice unmade would instead keep a record sorted by kind of change, distinguishing, among other things, target knowledge, collateral knowledge, changed evidence or credence, eliminated possibilities, recognized ignorance, improved method or capacity, promotion in Dellsén's sense, rationally expected promotion, and realized downstream contribution. The last records in retrospect what in fact followed, and this book adds it alongside promotion, which compares prospects. The items can overlap. Where the research system is artificial, each item can further be recorded as realized only in the system or as having entered S's epistemic situation (Section 7); on accounts of possession that require entry, only the latter is yet a change in S's epistemic situation, and so in what an account of progress for S would count. Kept apart, they let an evaluation count more than knowledge thresholds without treating every epistemic motion as progress. Someone who adopts one account of progress can still compute a single score from this record, weighting the items as that account does.

9. What the epistemological accounts require of artificial research

Sections 2 to 8 drew consequences for artificial research where each question arose. This section gathers them into requirements and says on which claim, and so on whom, each requirement falls. The unread artificial discovery of Section 7 shows why the second task, saying on whom each requirement falls, matters. The same record can, once incorporated into S's practice, contribute to the result's becoming S's knowledge, be available as a collective resource on a different account, or fail to be anyone's knowledge despite containing a true result. What must be built thus includes the relation between a result and its possible bearer, and the accounts examined here specify different ways that relation can succeed in the case of knowledge. What is required changes with the claim being made: that an activity is research for S, that knowledge new to S has been realized and in whom, or that it has entered S's epistemic situation (Chapter 1, Section 2). Throughout, the subject claimed to know need not be the actor that conducted the research (Section 2.1). Whether a requirement falls on the system therefore depends on whether the system is that subject, on whether it gathered the evidence and constructed the verification, and on the account of possession adopted.

The requirements are conditional. Each states what knowledge would require on the account in question and, where it differs, what an evaluator would need to establish to support the attribution; as Section 5 noted, insufficient evidence for an evaluator does not by itself show that the subject lacks knowledge. The requirements of different accounts are not to be added up into a single list that every artificial researcher must meet (Section 4.2), and this section does not choose among them. The conditionals concern claims of knowledge; the corresponding consequences for claims of understanding are drawn in Chapter 3, Section 9.

Stated for whoever is claimed to know, the accounts compared in Sections 4 and 6 impose the following. If accessible reasons are required, a post-hoc explanation is insufficient unless the relevant subject can use those reasons in inference, challenge, and revision. If evidential fit is required, the live confidence or acceptance of the subject claimed to know must fit the evidence that subject has at the time; since fit is assessed at a time, an attitude left unrevised when that evidence changes can cease to fit. If the attitude must also be held on the basis of the evidence it fits, as Feldman and Conee's well-foundedness requires (Section 4.3), the operative basing relation becomes part of what is required. If responsible evidence acquisition is also required, the decisions on search, measurement, omission, and stopping through which the subject's evidence was gathered must be assessed separately from the final inference. If reliability is required, the process that forms or maintains the subject's belief must be reliable in the relevant environments; for an evaluator to support the attribution on that ground, that performance must be established across those environments. If safety is required, the subject's belief must not be one that could too easily have been false, and an evaluator supporting the attribution must examine contamination and nearby possibilities of failure along the route by which the belief was formed. If sensitivity is required, what must be examined is whether the subject would still have held the belief had the claim been false. If virtue-reliabilist competence is required, success must be attributable to a stable capacity of the relevant agent or organized system. If responsibilist virtue is required, the conduct of the subject's inquiry, and not only its reports, must express something that functions as intellectual character and motivation. If defeat sensitivity is required, the response to rebutting and undercutting information must reach the subject's commitments that rest on the affected claims, instruments, assumptions, and validators, not only the reported conclusions; provenance that preserves the defeasible support relations among these is one way to make that response traceable. If social criticism is constitutive, formal publication without uptake in Longino's community-level sense is insufficient. If group functionality is sufficient, the boundaries and organization of the collective, whether a human–artifact system or a wholly artificial one, become part of the epistemic claim. These last two concern possession as well as warrant, and Section 9.4 returns to them.

Drawn together, Sections 4 to 7 also imply that a knowledge claim in artificial research has to state its units: the process type whose reliability is claimed and whether selection by others lies inside it, where competence is seated (Section 4.2), the boundary of the evidence the system has and whether a retrieval store lies within it (Section 4.4), the boundary of a collective claimed to know (Section 6), the reference S, and the account of possession (Section 7). For an individual human investigator, who holds the belief, how far the evidence extends, and where competence is seated are largely fixed by there being one person. For an artificial system they need not be fixed in advance. Where a system works through instances, external stores, tools, or other systems, they can be drawn in several ways, depending on how these are connected, and whoever draws them, whether designers or a system that reorganizes itself, the verdicts change with where they are drawn (on the boundaries of an actor, Chapter 7, Section 2). Until the units that a given account turns on are stated, a claim cannot be checked against that account. The subsections that follow take the requirements by where they fall: on the system as the conductor of research, whoever knows (Section 9.1); on the system as the subject claimed to know (Section 9.2); on others who are claimed to know through it (Section 9.3); on the entry of a result into S's epistemic situation (Section 9.4); and on changes short of knowledge (Section 9.5). Table 2.3 at the end of the section sets out the result in simplified form.

9.1 What falls on the system as the conductor of research

Some requirements fall on the system because it conducted the research, whoever is claimed to know. None of them is a condition of the first judgment, whether an activity is research for S. That judgment does not depend on whether the result meets the conditions of entry of the account of possession adopted (Section 7): an activity organized toward knowledge new to S can be research before anyone knows its result, or when it fails (Sections 7 and 8.1), and conversely a true output does not make a run research, since a causal contribution to an achievement is not by itself an attempt at it (Section 4.5). What falls on the conductor concerns how well the attempt was conducted, the second of the evaluations in Section 8.1, and what others need in order to come to know through it.

The first concerns the gathering of evidence. Where the system did the gathering, its decisions on search, measurement, omission, and stopping are assessed separately from the final inference. Whether that assessment is epistemic or practical is the dispute set out in Section 5: on Kauppinen's view it is practical, while an unsupported assumption about the reliability of sources on which the system acts remains open to epistemic criticism, of the system if it is itself a subject of such norms, and otherwise of those who built the assumption into it. The assessment thus applies to the system as gatherer even where it is not claimed to know.

The second concerns verification. Who constructs the verification is a different axis from who knows (Section 5). If the system runs checks that others designed, its reports have to show what was tested under which auxiliary assumptions, so that those who read them can locate what a result bears on. If the system constructed the checks itself, those who come to know through it and who take on the diagnosis of results themselves need to be able to read from its reports how it did so (Section 5): how many candidates were generated and how many of them were likely to be false, which choices were made after the data had been seen, whether the data that support the claim were used in constructing it, what would have counted against it (Section 5.2), and how a formally correct procedure was related to the question (Section 5.3). On accounts that require accessible reasons, recipients need reasons they can use, which may concern how the result was checked or the system's reliability for the use at hand (Section 5; Chapter 11, Section 5.2); the system's reports are then one source of such reasons. On accounts on which the reliability of the process suffices, a recipient can know without reading these, but an evaluator who supports the attribution still needs evidence that the system's way of verifying is reliable. Whether the system also constructs and revises the governing standards is a question of autonomy (Chapter 7, Section 4.3), a different axis from who knows, and not among the conditions of knowing (Section 5.3; Chapter 8, Section 1.3); for research for a human S, inherited human standards can serve.

The third concerns the independence of checks. A critic adds to error detection only so far as it does not go wrong where the producer does, so where checks are run by instances of the producer's own base system, or by other systems that may go wrong in the same places, a report would need to show how far they are independent (Section 5; Chapter 11, Section 3). This holds whether S is human or artificial, and agreement among such instances does not make a result true (Section 6).

The fourth concerns the unit of a claim credited to the system. A reliability claim has to state its process type and the distribution over which it holds, and selection performed outside the system, such as human filtering, must not be hidden inside a claim about the system, since it can make outputs safe without making their success creditable to the model (Section 4.2). For a recipient, human or artificial, who relies on the whole pipeline, the whole may be reliable while the model alone is not; what is required is that the unit be stated, not that there be no filtering.

Answerability, the responsibility to answer when questioned about an output, is a further matter, and it need not track who knows either: those who answer for a system's outputs need not be among those who know. If a system's reports are treated as readings of an instrument rather than as assertions, applying Goldberg's account as reported in Chapter 11, Section 5.1 directs reliance toward the epistemic work of its designers and makers; whether its designers and makers can take informed responsibility for an essentially opaque system is disputed, as Koskinen argues (Koskinen 2024a, 464–465; Chapter 11, Section 5.3).

9.2 When the system itself is claimed to know

Where the system itself is claimed to know, the conditionals above fall on its own states. Several of the false positives of Table 2.1, which take this case as their example, then share one form with a point from Section 2.1: what the system displays differs from what governs its output and revision. A stored sentence is weak evidence of belief beside stable integration with inference, action, and revision (Section 2.1); a post-hoc rationale can be unrelated to the operative route; hedging in reports can leave the conduct of inquiry unchanged; and a corrected report can leave the operative commitment as it was (Table 2.1; Section 4.3). Taken together, these cases suggest that much of what this claim needs is evidence about what governs the system, not about what it shows.

For several relations, the chapter has said what that evidence is. For the attitude, the system's states must satisfy the belief condition of the account adopted, or an account that does not require belief in that form must be defended (Section 2.1); and since a state can be a belief, an acceptance, a credence, or a supposition, which of these it is has to be settled before a condition for belief is applied to it (Section 2). For basing, the system's operative commitment must be based on the evidence it fits, on accounts that impose such a condition; co-location of a conclusion and its grounds does not establish it, and provenance that records which items governed a commitment is the device that fits causal theories of basing most directly, though, as a means of leaving evidence for an attribution, its absence does not by itself show the absence of basing (Section 4.3). For evidence, on the functional accounts of belief, crediting the system with knowing what its store contains requires evidence that the store's contents are integrated with its inference and revision, not only that they are retrievable (Section 4.4). For defeat and verification, the system's operative commitment, not only its reports, must respond appropriately to rebutting and undercutting information and change with the revisions that verification prompts (Sections 4.3 and 5); on Williamson's argument an updating that can only add evidence cannot represent the loss of evidence (Section 5). On Simion's knowledge-first account, if the system is a trained one, whether it has the function of generating knowledge turns partly on what its training and selection rewarded, not only on its present accuracy (Section 4.1). And a system whose commitments can be revised by proofs, observations, failed predictions, and criticism is related to its object differently from one optimized only for textual plausibility or social approval (Section 3).

Where the system is not a member of S, whether S is human or artificial, a system that knows is a bearer outside S; the case in which it is a member is taken up in Section 9.4. Its knowing realizes knowledge new to S somewhere, the second judgment, without by itself securing the third, that the result has entered S's epistemic situation, and a claim about it needs an account of how its achievement is new relative to, and sought for, S (Chapter 8, Section 1). The claim that the system knows is thus a further claim, beyond both the claim that it conducted research for S and the claim that the result has entered S's epistemic situation, and unlike the latter it carries the burden of this subsection (Chapter 8, Section 1.1).

Two further points bear on applying these requirements. First, the conditions can return conflicting verdicts. Williamson argues that safety and sensitivity come apart. Safety does not entail sensitivity, and knowledge does not require sensitivity: as the example of the mark one meter above the water showed in Section 4.1, even a very slight systematic bias can make almost all of a subject's beliefs in contingent propositions insensitive while they are still known (Williamson 2000, 147, 159–161, 307–310). For evaluation, this means that a sensitivity test applied to a system with a small systematic bias could report failures that a safety-based account would not count against its knowledge.

Second, the alternatives that the conditionals set out do not require artificial research to reproduce the internal route used by a human investigator. A system unlike a human could realize a reliable, safe, reasons-responsive, or socially integrated relation through very different mechanisms. What cannot be inferred merely from difference of method is that no epistemic standard applies. The burden shifts from resemblance to humans to evidence that the alternative realization satisfies the relevant account, which is one reason the account has to be named.

9.3 When others are claimed to know through the system

Where humans, or artificial systems other than the producer, are claimed to know through the system, the system need not believe its output. An AI can, for example, contribute to human knowledge without literally believing what it outputs. The conditionals then apply to the belief of those claimed to know, or to their operative commitments if they are artificial, and the system's output bears on these as evidence or as part of the process that formed them (Section 4; Chapter 8, Section 1.1). The same output can then ground knowledge in one recipient and not in another. A human investigator who accepts a hypothesis that an AI proposes only after an independent test that would have failed had the hypothesis been false can hold a sensitive belief, although the AI would have proposed the hypothesis either way; conversely, if the output would have asserted the claim even had it been false, worded differently or not, a belief based only on that output is insensitive. An attribution of knowledge to those who know through the system is therefore an attribution about their belief and what it rests on, of which the output is one part, not about the system alone. The gap in basing that Section 4.3 found on the system's side reappears on theirs: reading a persuasive report supplies reasons, but whether a belief is based on them or on the system's prestige is a further matter.

What the system needs to supply in this case differs with the account adopted for the recipient. On accounts that require accessible reasons, the recipient needs reasons it can use in inference and objection. Where these concern how the result was checked, what the reports make visible is central; where, as Section 9.1 allows, they concern the system's reliability for the use at hand, the reports are one source of them. On accounts on which the reliability of the process suffices, or on Sosa's account from Section 4.2 applied here, the recipient can know without being able to read the reports as a whole; what then falls on the system is less that its reports be readable than that the process actually be reliable, and an evaluator who supports the attribution needs evidence of that. The requirements of Section 9.1 on what reports need to show come from the recipient's side chiefly where the recipient takes on the diagnosis of results itself or where its reasons must concern how the result was checked, and otherwise from what an evaluator needs in order to support an attribution. In every case they fall on the system because it conducted the research and constructed or ran the checks, not because it knows.

Where the recipient is artificial, the requirements of Section 9.2 move to it: whether it believes, and whether its operative commitment is based on the output and revised under defeat (Section 4.3). Carried over to such a recipient, the sensitivity example connects with the question of independence that Chapter 11, Section 3 examines: a recipient whose check of the output is made by an instance of the producer's own base system, or by another system that errs where the producer does, and which would have accepted the output whether or not it was true, holds an insensitive commitment, however many instances agree.

Other relations can make the system's own knowing matter even where others are the knowers. Where its reports are taken as assertions, accounts of assertion place conditions on it as speaker, although whether current outputs are assertions, and whether the norm of assertion requires knowledge, rational belief, or support that varies with context, are disputed (Chapter 11, Section 5.1); and where the achievement sought is a skill realized in the system, the system is its first bearer (Section 6).

9.4 When a result is claimed to have entered S's epistemic situation

The verification schema of Section 5 ends with whoever is claimed to know. Unless S itself is the subject claimed to know, entry into S's epistemic situation is a further step that the schema does not contain, and what it requires depends on S and on the account of possession adopted (Sections 6 and 7). What has to be built are routes by which the result can enter validation, inference, action, and correction, and they need not lie within the system that conducted the research (Section 7). Whether such routes suffice differs by account. On Bird's account, extended to a record that no one has read, they may. On Lackey's, since a society as a whole without operative members has no group belief, the result must reach and be believed by members of S, and, where entry is claimed through a group's knowing it, by the operative members of a group within S; on Longino's, it must be accepted by members after criticism under the account's conditions; and on Giere's adaptation of knowledge without a knowing subject, it must be certified by reliable, accepted procedures and publicly accepted (Sections 6 and 7). A claim that a result has become S's knowledge can therefore be assessed only when it reveals enough of its account (Section 6). The distinction between a dormant record and a live epistemic capacity names what makes a record a live capacity, but it marks the boundary of knowledge only on some accounts; and in the dispute over Lackey's missing-child case, whether a group knows a result comes apart from whether it should have known it, so that a group whose system collates a record can be answerable for a result it does not yet know (Section 7).

Where S is artificial, a question arises before any of these. Both roles of S presuppose that its epistemic situation could come to include the knowledge sought (Chapter 1, Section 2), and that situation now consists of artificial members, their organization, and their records; so whether such members can be knowers arises before any member knows the result, and whether or not A is one of them (Section 6). The accounts then ask of the members what Section 9.2 asks of the system: whether they believe the result or can take part in joint acceptance of it. Of their organization they ask whether it realizes a properly organized cognitive system and, on Bird's account, for what that organization was selected or designed (Section 7), whether its interaction involves the diversity and public standards of criticism that Longino's account requires, and whether certification by artificial members counts as certification, including where they are instances of one base system (Section 6; Chapter 8, Section 1.3; Chapter 11, Section 3). No human-membership condition is built into the working characterization, but nonhuman subjecthood is not established by coordination or plurality alone. As Section 9.2 noted for the system, the relations need not be realized as they are in a human community; what has to be shown is that the realization satisfies the account.

The producing system may itself be a member of an artificial or mixed S. Of the cases in which the producing system is the one that knows, this is the only one in which its knowing can do the work of the third judgment: on individual-possession accounts, if the system knows, a member of S already knows, and the result may have entered S's epistemic situation as soon as it was produced. On Lackey's account this holds for a group only if the system is among its operative members and a significant percentage of them believe the result; on Longino's, the system's acceptance is at most one member's; on Bird's, the verdict still turns on the organized system as a whole (Section 7). For a mixed S, a claim of entry has to say which members, human or artificial, the result reaches and how possession, memory, and correction are distributed among them (Chapter 8, Section 1.2).

9.5 Changes short of knowledge

One requirement comes not from an account of knowledge but from the claim about progress that Section 8.2 makes on Wilholt's account. If making ignorance explicit counts as epistemic progress, evaluation must also distinguish the absence of an answer from a represented and actionable unknown, and say for whom it is actionable: for the system's own further inquiry, or for members of S, human or artificial, who take up its work. Answer volume and closure rate can then be anti-correlated with research quality when they erase warranted suspension, live alternatives, or the evidence still required to resolve a question. Williamson's arguments, if they hold and the system's retrieval has no upper bound on its time, so that it cannot tell what is missing from what has not yet been reached, bound what can be asked here: the system's own checks cannot guarantee that its record of what it does not know is complete (Section 8.2), so the requirement concerns what is represented and for whom, not completeness.

Where an evaluation does not select an account of progress, it keeps the record of Section 8.3: changes sorted by kind, and, where the research system is artificial, by whether each is realized only in the system or has entered S's epistemic situation. A claim that a single score represents progress as such already chooses among the accounts of progress (Section 8.3).

Table 2.3 sets out where the requirements of this section fall. Its rows are the claims distinguished in Chapter 1, Section 2 and Section 8.1, with the claim that the system knows separated from the claim that others know through it, since the requirements differ most between those two. Its last column states what each claim does not require, which Tables 2.1 and 2.2 leave implicit. The table is a simplified overview: each cell compresses requirements that differ by account, and the rows are not exclusive, since the system can, for example, be a member of S. As with Table 2.1, the entries in the second column hold only on the accounts concerned and are not to be combined into one list that every system must meet (Section 4.2).

Table 2.3. Where the requirements fall: a simplified overview

Claim	What must hold (on the adopted account, where the claim depends on one)	What an evaluator needs to be able to see	What the claim does not require
The activity is research for S	organized direction toward an achievement new to S that S could come to hold	the organization of the attempt and the reference S	that knowledge was reached; that the system knows or believes the result
The system conducted it well, whoever knows	responsible evidence acquisition; reports adequate to the checks it ran or constructed	how evidence was gathered; what was tested under which assumptions; how many candidates were generated; how independent the checks were; the unit of any claim credited to the system	that the system construct or revise its governing standards, an autonomy question (Chapter 7, Section 4.3)
The system itself knows the result	its states meet the account's belief condition, if it has one, and its operative commitment the account's further conditions	what governs its output and revision, not only what it displays	that it follow a human route; that the result has entered S's epistemic situation
Others know through the system	their belief, or their operative commitment if they are artificial, meets the adopted account, with the output as evidence or as part of the process	what their belief rests on, including whether any check was independent of the system	that the system believe the output (though where its reports are taken as assertions, accounts of assertion place conditions on it as speaker, and where the achievement is its own skill, it is the first bearer)
The result has entered S's epistemic situation	routes into validation, inference, action, and correction, and whatever the adopted possession account adds	which account is relied on, and which members or organizations the result has reached	that the routes lie within the system; that the system be a knower, unless it is a member of S whose knowing is claimed to count
For an artificial or mixed S, in addition	that S's members or organization can hold the achievement, and the possession account applied to them	independence and diversity among members, including where they are instances of one base system; the organization of the collective; for a mixed S, which members the result reaches	that the relations be realized as in a human community
A change short of knowledge	a represented limitation or other change, recorded by kind	for whom it is represented and actionable, and whether it has entered S's epistemic situation	that the change be knowledge; a single progress score

Comparing the requirements of Table 2.3 account by account brings out consequences that the accounts share and consequences on which they part. One shared consequence is that, whichever account it is checked against, a knowledge claim about an artificial configuration can be checked only once it is settled where the units that account turns on are drawn. Why the units are needed differs by account. On evidentialism, for example, how they are drawn settles what evidence the system has; on reliabilism, which process type has its reliability assessed; and on virtue accounts that require competence, where competence is seated (Sections 4.2 and 4.4). One consequence on which the accounts part is whether a recipient who knows through the system must have reasons it can use and, if it must, whether those reasons must concern the verification itself. Accounts that require accessible reasons demand reasons the recipient can use in inference and objection, but these need not be about the verification; they may concern the system's reliability, and the reports are one source of them (Section 9.1). On reliabilism, on which the reliability of the process suffices, and on Sosa's account applied here, the recipient can know without reading the reports. A recipient needs to read from the reports how the verification was constructed where its reasons must concern the verification itself or where it takes on the diagnosis of results itself. The same requirement on what reports show is therefore imposed, in those cases for the recipient's sake, and otherwise so that an evaluator can establish that the process is reliable (Section 9.3).

Knowledge does not exhaust research achievement, but it is a demanding target. On the accounts compared, novelty, truth, reasonable acceptance, reliable success, understanding, social possession, and research capacity belong to different claims, and their requirements fall on different parties, so none of them is to be inferred from another without an argument from the account adopted. Carried through artificial configurations, these distinctions have concrete consequences. On evidentialism, counting only the retrieved documents as evidence the system has can leave the conclusion justified despite poor evidence gathering, while widening the boundary to count the whole store can cost the same conclusion its justification, since the counterevidence in the store then counts as evidence the system has (Section 4.4). On accounts that require basing, good reasons attached to a report are one thing and a conclusion held in place by them another, and a corrected report can leave the operative commitment unchanged (Section 4.3). A chance generator can contribute to another subject's knowledge without knowing and without conducting research (Section 4.5). For a system that generates candidates in bulk, passing a check can raise the support for a claim without making the claim probable (Section 5.2). And whether a result that no human has read is knowledge for S differs not only between the accounts of Bird, Lackey, and Longino but also with whether machines count as knowing members (Section 7). Memory, updating, and the arrangement of members are not only external means of carrying knowledge; on some accounts they bear on whether there is knowledge at all. Counting apparently novel conclusions makes none of these differences visible.

One question now becomes unavoidable. Suppose an artificial system produces reliably validated true predictions that a community can use, but no human member of the community can explain why the predicted relation holds, and the system supplies no usable representation of the dependencies behind it. The present chapter has asked whether knowledge has been brought about. Even where it has, whether that reaches the intended target depends on what the inquiry sought. The next chapter asks when understanding is achieved, and by whom: what would count as understanding, how it relates to knowledge, who, including the system itself, can have it, and what follows when the two come apart.