Chapter 11 — Criticism and Certification for S

Defeat, Independence, Trust, and Correction

1. Which evaluation this chapter examines, and why

Chapter 9, Section 2.1 derived three relations through which research is tied to S (the society to which an achievement is to be new) beyond the activity of whatever unit conducts a particular inquiry. The second, criticism and certification, concerns the assessment of epistemic success, and this chapter develops it first through knowledge. Producing a result does not make it something S can rely on. Of the three, it is the relation that bears especially directly on what it is for a result to be knowledge and to have become S's, though only on accounts on which social organization does part of the work by which knowledge is warranted or held (Chapter 2, Section 6; Chapter 9, Section 2.1). This chapter asks what happens to that relation when an artificial configuration can take it over. A configuration here is, in the sense of Chapter 7, Section 1, the unit to which the performance of research is attributed: a single model, a collection of agents, an organization of humans and AI, among others. In the example of Chapter 9, Section 1, a single system generates a hundred instances, assigns some of them to check and reproduce what others produce, and its builders claim, among other things, that the results which pass its internal checks are certified for S. The chapter asks what such a claim would require, which part of the evidence for it can come from inside the artificial configuration, and why the part that cannot come from inside remains a relation to S.

This chapter examines, among the many evaluations in research, the evaluation of results addressed to S and the criticism on which it draws; it calls the former certification. There is a further reason for this focus besides the one just given: the changes an artificial configuration brings to evaluation put pressure on the independence of evaluators, on which this evaluation draws, and on the visibility of evaluation to S. Evaluation in research is plural. In human research, for instance, funders have assessed proposals before any result exists, reviewers and editors have assessed manuscripts, institutions have assessed researchers for credit and positions, and those who use a result have assessed it in use. Chapter 9, Section 4 set out four things that any of these involves, among them that a standard tied to a purpose is applied for some party's choice and that the evaluator and the addressee who uses the verdict can differ. Two of them are used later in this chapter. One is that proxies are unavoidable, because what an evaluation of research is about is usually not observable when the judgment is needed. The other is that, for something new to S, the evaluator may not yet have a standard (Section 8).

An artificial configuration changes at least three things about evaluation directly. First, it can generate evaluators rather than recruit them, as copies of one base system or as evaluators that differ widely from one another. Second, it can place those who produce results and those who evaluate them within one artificial configuration, and bring evaluation continually into results while they are still being formed. Criticism then changes not only in cost but also in range, judgment, and arrangement: some kinds of check become nearly costless, checks that humans did not ordinarily perform come to be attempted, and evaluators can err in places other than where humans err and pick up different things. And evaluation can dissolve into production, so that the very moment of "having passed review" disappears and evaluation is no longer visible to S. Third, it can couple evaluators and producers over time, adjusting each to the other's outputs (for example, by training or selecting each on them). Sections 3, 7, and 8 take up these three in turn.

Other evaluations are taken up elsewhere: the ex ante evaluation of lines of inquiry, as part of allocation, in Chapter 10; credit and the attribution of contributions in Chapter 7, Section 3.3; the evaluation of claims about artificial researchers in Chapter 7, Section 5 and Chapter 13; and who may set and revise the standards and error weights that evaluation applies in Chapter 12. Allocation, though, also uses judgments produced by criticism, so the judgments of allocation that Chapter 10 examined rest in part on the criticism examined here.

Chapter 3 drew a limited permission from the distinction between discovery and justification and from Duede's application of it to deep learning: the route that generates a candidate answer need not be the route that warrants accepting it. Even where the generating process is opaque, the answer may be used if the grounds for accepting it are supplied by another route. Reliabilist and virtue-epistemological accounts narrow the range of cases that need the permission, since on them a reliable generating source can itself supply warrant without a separate route. Chapter 3 also added a condition. The permission holds only when the candidate is assessed through an adequate independent route, and evidence of shared evaluators, selection, contamination (evaluation items or their answers having entered the training data of what is evaluated), and hidden human filtering is needed to judge whether the independence is real. That discussion concerned a single result. When an artificial configuration can instantiate not only the investigator but also the reviewer, the replicator, and the critic, criticism itself becomes something the configuration can organize internally. Here and below, when names such as review and replication are used for parts of an artificial configuration, they name functions that human institutions have bundled, not the persistence of those institutional forms; Section 7 asks what those functions are.

This chapter distinguishes two relations: the defeat relation and certification. It begins with the first. Consider three checks. A predicted property of a candidate material is measured with an instrument whose readings are not inferred from the prediction. A generated proof is checked against formal rules. A forecast registered before an event is compared with what happens. In each case, the candidate or its grounds can encounter evidence against it that is not merely a repetition of the producing process's reasons for accepting it. Call such an evidential route a defeat relation. The term names not the event of a result's actually being overturned but the existence of a route that could show the candidate wrong if it were wrong; a candidate that passed a check also stood in such a relation. Chapter 2 distinguished evidence that rebuts a claim from evidence that undercuts its grounds; Chapter 8, Section 2.3 called an epistemic organization world-answerable when it is constrained by its targets in such a way that its commitments would differ if the targets were different, and remains corrigible through further contact. A defeat relation specifies a route through which such correction can arise. Whether a particular measurement, formal check, or comparison supplies one depends on what it tests and how its result bears on the claim, not on its name. Neither an infallible check nor a guarantee that every false candidate will fail is built into the term.

Keeping such a route beyond the producer's power to alter is one way of protecting it, not the definition of the evidence it adds. A program may have permission to revise a checking routine and nevertheless submit a candidate to an unchanged version that detects errors its generation process misses. Permission alone does not remove that information. Conversely, a routine that the program cannot change may reproduce the generator's mistaken assumptions. Its protection alone adds none. As these two possibilities show, the evidential contribution of a check is distinct from the arrangements that maintain it. Section 2 develops the first of these, the evidential contribution of a check; Section 8 asks how the contribution can survive adaptation, selection, and changes to the check.

Failure also needs interpretation. Chapter 6, Section 2 called the work of deciding which commitment a contrary result gives reason to revise localization. A contrary measurement can bear on the candidate, an auxiliary assumption, a calibration, or the match between a formal statement and the question. The producer may correctly locate the failure; assigning localization to another party does not by itself make that party right. What matters here is whether the diagnosis has support beyond the wish to preserve the candidate and can itself be challenged by further evidence. If every adverse result can be reassigned without such support, the measurement may remain informative while the judgment whether to accept the candidate ceases to answer to it. Separating custody of evidence, recording revisions, or sharing authority over disputed diagnoses can protect this answerability. Their value depends on the challenge they preserve; they are not universal conditions that every informative check must meet.

The other relation is certification: the relation through which others are given reason to rely on a result for specified uses. Certification refers to S. It concerns who within S may rely on a result and for which action, and on Longino's critical contextual empiricism (examined in Section 3) it also concerns the perspectives from which the result has been criticized, since acceptability is relative to the perspectives available to the community, as Büter's quotation of the definition makes explicit (Longino 2002, 135–136, as quoted in Büter 2012, 84). An artificial configuration may secure strong defeat relations internally and still leave open whether its certification gives S reason to rely on what it certifies. The three changes named above bear on both relations, the defeat relation and certification. The chapter asks how much of the former can be secured inside an artificial configuration and why the latter remains a relation to S, and further whether time and the reduced visibility of evaluation reopen to the outside even the part of the former secured inside.

Whether a defeat relation itself refers to S depends on the account adopted. On accounts that locate the value of criticism in the errors it detects, a defeat relation can hold inside a single actor: a mathematician, or an algorithm, can propose a proof and have it checked against rules the proposer does not control, and nothing in that relation refers to S. On Longino's account, evidence bears on hypotheses only through background assumptions (Longino 2025, §5), so what a measurement counts against is relative to assumptions that a community holds. On that account, the claim that a defeat relation does not refer to S holds only so far as the background assumptions through which its outcome is read are not themselves among those that S would contest.

This divergence follows the branch drawn in Chapter 2, Section 6. There, the claim that inquiry causally depends on social practices was separated from the stronger claim that evidential status, objectivity, or knowledge is partly constituted by organized social practices, and the book declined to adopt either account of social knowledge. The question of this chapter can therefore be read in two ways. On the first branch, criticism matters because it detects error, and the question is what makes one check independent of another; that is the question of Section 2. On the second, organized criticism partly constitutes objectivity or warrant, and the question is what kinds of difference among critics the constitution requires. Section 3 compares several accounts on this branch, alongside the view of the first branch (Table 11.1). What becomes a topic on the second branch is objectivity in the sense of being partly constituted by organized criticism; this does not keep the view of the first branch from speaking of objectivity in other senses.

At several points the chapter uses the sorting of Chapter 9, Section 6. Chapter 9, Section 2.3 proposed that the division of the relations surrounding research among separate parties may have supplied certain goods without anyone designing them. The sorting of its Section 6 asks what each such good depended on, and distinguishes four kinds of answer. One is premises about human realizers, the people who actually held the relations (their limited attention, their motives, their lifespans, and the like). The second is premises about the present organization of science, how the relations have historically been organized (that only investigators have access to the primary data, for instance, or the gate formed by review); these can be reorganized even while the realizers remain human. The third is conditions of inquiry, which hold whatever realizes it. The fourth is relations to parties outside the artificial configuration, which the chapter, following Chapter 9, sometimes calls relations to S for short.

Section 2 defines independence, on the error-detection view, as the information a check adds. Section 3 asks where difference comes from when critics can be made, by how each of five families (Table 11.1), this view among them, assesses it. Before the second and third changes are taken up, Sections 4–6 turn to certification, since what Sections 7 and 8 ask about those changes is stated in terms of the thresholds and grounds of reliance set out there; they add two questions to the two branches of this section: whose valuations the threshold at which a result counts as sufficient for a use answers to (Section 4), and what grounds reliance on an unchecked result and whether that reliance has to be trust (Section 5). Section 6 asks what remains with S at the limit at which S as a whole stands as a layperson relative to the artificial configuration. Section 7 asks how the cost, range, judgment, and arrangement of criticism change when production and evaluation are joined in one artificial configuration, and what can be lost when evaluation dissolves into production and ceases to be visible. Section 8 asks what coupling over time does to critics' independence, to the standards by which the new is evaluated, and to correction. Section 9 returns to the builders' claim and considers it by combining these accounts. In outline: an artificial configuration makes the properties that the accounts each track (the independence of checks from one another, there being someone who answers when a result is wrong, those who produce and those who rely sharing valuations of the weight of errors, and the visibility to S of the fact that evaluation has taken place) separately settable. In human research they were tied together by realizers and organization, so verdicts plausibly tended to coincide; for artificial configurations the accounts' verdicts on the same case diverge, and a builder's claim cannot be assessed unless it shows which account it answers to. And on some accounts, improving internal criticism can weaken certification for S.

These verdicts, however, are not where the chapter ends. The account-by-account verdicts are also a means of describing what can happen to research when criticism and certification move inside an artificial configuration, and the picture the chapter's material yields can be given in advance. Where critics are made rather than formed, what adds evidence is their difference rather than their number, and a body of critics that shares its assumptions can, on Longino's account, entrench them the better its internal criticism works (Sections 2 and 3). If evaluation dissolves into production, the moment of "having passed review" disappears and evaluation ceases to be visible to S; cheap criticism, multiplied, can then conceal a scarcity of independent criticism so far as it errs where the producer errs (Section 7). Evaluators can come to tell less about producers merely by being used for selection, without anyone deceiving. The very standards by which S judges results can move with the coupling of production and evaluation, and agreement with S's present standards is no evidence that they have not moved (Section 8). The weighting of errors is shaped in advance inside the configuration (Section 4), and certification can come to be treated less as a researcher's assertion than as the reading of an instrument (Section 5). Meanwhile the members of S, as they lose the places in which they gained experience of evaluating, can lose the very capacities needed to carry the relations that remain (Section 6). Where these coincide, passing internal checks may come to serve as a proxy for certification, the lack of relations to S may be selected along with it, and research may approach a state in which relations to S remain in form while no one is able to carry them (Section 9.3). The account-by-account examination that follows also serves to separate which accounts count which of these as a problem, and what each would take to prevent.

2. Independence on the error-detection view

On the first branch of Section 1, the value of criticism lies in its detecting errors. To ask about independence on this branch is to ask what makes a check able to find errors that other checks have not yet found. The question matters in this chapter because the three changes listed in Section 1 all put pressure on the independence of evaluators. When two checks return the same verdict, how much evidence does the second add? The literature on expert testimony has given a precise answer to this question for the case in which several sources agree. This section starts from it.

Nicolas J. Tanchuk reports Alvin Goldman's minimal condition. Agreement from a second source adds evidence only if that source is at least partly conditionally independent of the first: more likely to agree when the claim is true than when it is false, given what the first source has said (Tanchuk 2026, 4). Goldman and Thomas Blanchard likewise explain that merely multiplying uncritical followers of a leader adds no independent support (Goldman and Blanchard 2015, §3.2). The condition says nothing about persons or organizations. It concerns the relation between the ways two judgments can go wrong. Guerrero reports Goldman's warning that sources which merely reflect one another or a shared third party do not supply the needed independence (Guerrero 2017, 177). Nor does independence require that the second source bypass the first. Tanchuk reproduces Goldman's example of a source that hears the first source's reasons, considers defeaters the first never considered, rebuts them, and only then agrees; such a source can still be partly independent (Tanchuk 2026, 5).

This book extends the condition within this view. On the extension, a critic is independent to the extent that it adds error-detecting information, with respect to a given claim, beyond what existing critics already supply. This is an application to checks in general of the distinction Goldman draws for agreement among sources, not Goldman's own account of such checks. On the extension, organizational separation is neither necessary nor sufficient for independence. It is not necessary, because a single configuration can check a proof with a kernel (the small program at the core of a proof assistant that checks each step of inference) whose failure modes differ from those of the process that found the proof, or test a prediction against a measurement it does not control. It is not sufficient, because evaluators in different organizations, under different names, can share the same training, the same sources, the same evaluative standards, or the same conception of what a good result looks like, and fail together. The powers that Chapter 9, Section 5 named, such as keeping evidence apart, can hold routes of criticism apart in the way that an instrument in separate custody is held apart; on the extension, holding them apart matters to the extent that it leaves them different ways of failing. The power to stop updates to other parts can bear on this view as well, as a means of stopping a checking part from being updated on the producer's outputs (Section 8). Evidence about present language models gives one example of the insufficiency. The correlated errors noted in Chapter 10, Section 4 persisted, among the larger and more accurate of the language models studied at one time, even when providers and base architectures differed: a failure of the kind described here, in which a way of going wrong is shared although names and organizations are separate (Kim et al. 2025, abstract). The same can hold for human evaluators who draw on the same models. Corpus-level estimates suggest that between 6.5 and 16.9 percent of the text of peer reviews for four AI conferences held in 2023 and 2024 could have been substantially modified by language models, and the estimated share was higher among reviews that were more similar to the other reviews of the same paper (Liang et al. 2024, 1, 9). On this view, where such shared use is not visible, the formal separation of critics does not by itself show whether, or how far, they share a way of going wrong.

A defeat relation, in the sense of Section 1, gives a check a way to challenge a candidate or its grounds. On this error-detection view, the independence at issue is its informational contribution relative to what other checks already supply. Internal coherence and agreement among evaluators do not establish that contribution by themselves. Contact with the world and formal checking can supply it, including inside one configuration. What a check establishes also has limits: a checked proof establishes a theorem relative to its formal statement, not that the statement captures the question (Chapter 6, Section 3).

Consider a further constructed contrast. A check rejects a candidate, and the producer proposes that a calibration error caused the rejection. In one version, measurements against a separately established reference support that diagnosis; the calibration is corrected and the affected candidates are reassessed. In another, the only reason for changing the calibration is that it rejected the favored candidate. The first can improve the check, while the second gives no additional reason to accept the candidate. The difference is not whether the producer took part in the revision but what supports it and what remains able to count against the revised judgment. Records can expose that difference; merely recording a change does not by itself show that the revision was supported. Chapter 6, Section 5 develops revision of standards within inquiry; the point here is its consequence for reliance on certification.

Independence, protection, and severity therefore answer different questions. A check may add information despite being open to revision; protection can help maintain that contribution without proving it. Even a protected, informative check can detect too narrow a range of errors to warrant the claim made from passing it. What Mayo calls a severe test (Chapter 6, Section 2.3, on Mayo 1996), which can be glossed as one with a high probability of exposing the relevant error if it were present, concerns this further question. Adequate certification therefore does not follow from any one of these properties.

Four cases show what this view picks out, and each is introduced for a different point. The first two form a contrast between a division of relations that supplied independent routes and one that did not. After a 2020 report of phosphine in the atmosphere of Venus, one outside group re-ran the published scripts and criticized the data reduction, another informed the observatory of a difference in a calibration setting, and observatory staff identified problems and reprocessed the archived data (Snellen et al. 2020; Villanueva et al. 2020; European Southern Observatory 2020; Greaves et al. 2020). The case illustrates the claim that independent routes can lie outside the producing team: the problems were found by parties who ran their own analyses and who held the archive and the calibration records. By contrast, the investigation of the Hubble Space Telescope's flawed mirror found that the division responsible for fabricating it had been insulated from review, that indications of error had been discounted during fabrication, and that the program had relied entirely on a single test instrument (Allen et al. 1990, iv, 9-1). Chapter 6, Section 3.2 used the same case to show that a passed check can rest on a standard that is itself wrong; here it illustrates the claim that checks can be numerous and still share one error route, since tests that did indicate the error were discounted because they disagreed with the instrument on which the program relied, which served as the standard. Two episodes do not show how often a division of relations supplies such detection.

The third case shows that independence can be designed by arranging roles. At the eighteenth-century Nautical Almanac, each month's tables, apart from those for the moon's position, were computed separately by two computers and compared by a third person; when the comparer found, by January 1770, that two computers had been copying each other's work instead of working independently, both were dismissed, and when they were re-employed several years later they were not given the same months (Croarken 2002). The arrangement was valuable only so long as the two computations had separate routes to error, and copying removed that without changing the names of the roles. The fourth case shows the converse in artificial pipelines: dividing a pipeline into named roles does not make its checks independent. An independent critique of an autonomous materials laboratory argued that its automated analysis of diffraction data was not yet reliable and that, on the critics' reanalysis, none of the reported materials was new (Leeman et al. 2024, abstract). The authors' later correction, after manual reanalysis, confirmed 36 of the 40 reported syntheses and judged four inconclusive (Szymanski et al. 2026); the two address partly different claims, synthesis and novelty. A benchmark analysis covering ten tasks coded every paper generated by one automated research system as containing fabricated experimental results (H. Chen et al. 2025, app. B, fig. 6). In both, the systems included automated checking; what mattered was whether the checks had an error route of their own.

Different claims within one case can also be settled by different routes. When RIKEN tested whether the STAP phenomenon could be reproduced, a verification by a team led by Hitoshi Niwa concluded that the cell aggregates it obtained did not meet the original studies' definition of STAP cells (Niwa 2016), while a separate committee of outside experts, comparing genomes of stored cell lines with those of existing embryonic stem cell lines, concluded that the stored lines had derived from those lines (RIKEN Investigation Committee 2014, 13). On this view, whether a check is independent is a property of the check relative to a claim, not of the investigation as a whole.

On the view that values criticism for the errors it detects, a check's independence is judged claim by claim by the information it adds, not by the number or organizational separation of critics. Such a contribution can arise inside an artificial configuration. The further burden is to establish its scope and how it survives changes to the producer, the check, and the interpretation of failures. The next section asks where difference among critics comes from when they can be made, and how this view and other accounts assess both the limiting case of many critics made from one base and the case in which differences are made.

3. Where difference comes from when critics can be made

Section 2 asked what makes one check independent of another. In human research, critics are formed as separate persons with separate histories, so their number and their organizational separation have served as rough indicators of independence. Much of that independence was not designed by anyone; it was supplied by what human realizers are like. The first of the changes named in Section 1 removes this premise. An artificial configuration can make critics in any number and in any combination: it can make many copies of one, and it can make critics that differ widely. Number and separation then cease to serve as indicators, and where the differences among critics come from becomes something to be asked. This section first pushes to a limit the case in which many critics are made from one base, asking account by account what is added and what is not, and then turns to the case in which differences are made. The same question has a familiar form in human research. The reviewers of one manuscript read the same text, often with similar training, so that the question arises whether review is redundant. Whether redundancy is a defect depends on what is duplicated. Apart from overlap among reviewers, there is redundancy in another sense: review repeating an evaluation the producing side has already carried out. Section 7 takes that up.

This book distinguishes four senses of redundancy. The distinction is made for this chapter, to keep claims about redundancy from trading on one sense while establishing another. It draws on the conditions of Section 2 and on the literature cited below, and it divides the senses by what a further critic adds: nothing beyond what others supply, a route to error of its own, a different application of a shared standard, or a different addressee or function. The senses of redundancy are not limited to these four.

1. Duplication by correlated critics. Critics with the same training read the same record. On Goldman's condition, a non-discriminating copy adds no evidence beyond its source. Correlation alone does not imply this: critics can share some influences while still adding information about the claim. What matters is the additional information, not complete statistical independence.
2. Multiplication of independent error routes. Checks that work on different principles, measurements in separate custody, formal verification. A critic who rereads the record can find errors in what was done with it, such as a miscalculation, a confused statistical inference, or a gap in a proof, and adds to detection so far as it does not go wrong where the producer does. An error in how the record was made, such as a contaminated sample, may leave no trace in the record, and then a critic adds information about it only by making a new measurement or rerunning an experiment. The contrast between rereading and new evidence is drawn in work on jury theorems: as Franz Dietrich and Kai Spiekermann put it, enlarging a group whose members generate new evidence independently, as by experiment, adds "information, not interpretations of information", but only in a scenario that excludes any hidden common influence on different members' evidence, an exclusion they regard as implausible (Dietrich and Spiekermann 2021, §4.3). On the first branch, this is the redundancy that has value.
3. Diversity of judgment. Critics who share a standard differ in how they apply and weigh it. Thomas Kuhn, as Bird reads him, held that the values shared in science, such as accuracy, consistency, scope, simplicity, and fruitfulness, guide judgment rather than determine it as rules would, so that scientists who share them may still differ in their choice of theory (Bird 2025b, §7.3). Kuhn allowed, as Bird reports, that such differences can come from factors such as an individual's life history and personality and from influences outside the sciences, such as German Romanticism, while holding that in modern, professionalized science the factors at work typically come from within science, and that even a difference of personality, such as one scientist's being more risk-averse than another, remains a relation to the scientific evidence (Bird 2025b, §7.3). On this account, in the passage that Zollman quotes and Bird summarizes, the way new theories are ordinarily developed and tried by several scientists before a group accepts them requires a decision process that permits rational disagreement: a shared algorithm would lead all conforming scientists to make the same decision at the same time, and the looseness of the criteria, seen as values, is presented as essential to distributing the risks involved in introducing or supporting novelty (Zollman 2010, 2; Bird 2025b, §7.3). On Kuhn's account, the value of this redundancy lies not in error detection but in spreading across a group the risk that introducing or supporting a novelty entails.
4. Overlapping stages of evaluation. Evaluation can be repeated, at different times and for different addressees, between a result's production and its use. In human research, prepublication review, postpublication criticism, replication, citation, and downstream use, among others, have formed such stages. An artificial configuration need not have the same stages, but where it has arrangements that repeat evaluation, the same question arises. Where the stages have different addressees and functions, they are less redundant than they look; where each reads the same thing on the same record and goes wrong where the earlier stages do, they fall back into the first sense.

The first two senses are those the view of Section 2 captures. The third is not captured by it, since diversity of judgment may add no new error route and still matter, on Kuhn's account, for how novelty is carried. Kuhn's argument also has a limit. On Bird's reading, it supposes that in a large community the variable factors that shape individual judgment tend to cancel out: individual differences are normally distributed, and a judgment at the mean of the distribution corresponds to the one that the rules of scientific method, as traditionally conceived, would demand (Bird 2025b, §7.3). Kuhn does not consider groups of copies; applied to them, a large group of critics derived from one base system may, despite its numbers, lack the spread of individual differences on which Kuhn's argument relies.

Evidence from one field bears on the first two senses. Valentin Danchev, Andrey Rzhetsky, and James Evans compared published claims about drug–gene interactions with the results of robot-run high-throughput experiments. Findings about a claim from papers that shared authors agreed with one another more often than findings from socially independent teams, while, among claims supported by more than one paper, those confirmed by independent teams using varied methods and drawing on different prior literature were more often replicated in the experiments; in the absence of social independence, they write, "authors and methods reinforce agreement with themselves" (Danchev, Rzhetsky, and Evans 2019, 4, 6–9). The associations are correlational and come from one area of biomedicine. Read in the terms above, agreement among overlapping teams is partly duplication in the first sense, and confirmation by independent teams adds what the second sense counts. The study concerns human teams. Of its measures of dependence, shared methods and shared prior literature have counterparts among artificial critics that apply common procedures or were trained on a common record, whether or not they derive from one base system; whether such counterparts carry the same association, the study does not test.

Yasuo Deguchi gives an argument for the second sense at the level of kinds of inquirer rather than of individual critics. Drawing on nineteenth-century precision measurement, Deguchi notes that systematic error, the bias specific to an instrument, shows up only when different instruments measure the same object, and is cancelled by measuring the object by as many different methods as possible and averaging the results. With humans treated as one kind of cognitive instrument, the argument holds that a bias humans have simply by being the inquirers cannot be made visible so long as all inquirers are human, however many of them there are; artificial inquirers give a chance to see it where they and human inquirers diverge, and since they have biases of their own, the divergence points to both. Deguchi calls for keeping artificial inquirers unlike human ones. On this view, inquiry by humans alone leaves human bias in the result unseen, and leaving cognition wholly to artificial inquirers exposes the result to theirs; since no simple averaging cancels the two, both kinds have to take part in a dialogue that extends to the cancelling itself, examining the differences case by case, and artificial inquirers made too like human ones would inherit human bias and so remove the different perspective that lets it be seen (Deguchi 2026b, slides 4–12). Deguchi's argument concerns inquiry conducted jointly by humans and artificial systems, and its analogy assumes that the biases of different instruments scatter around a value there to be found. That assumption suits questions about how the world is, where cancelling biases can be thought to bring one closer to the value measured, which is answerability to the world (Chapter 8, Section 2.3). It does not carry over as it stands to questions about what matters to S, since what matters is not a single value on which inquirers of different kinds would converge; applied there, a divergence between kinds may reveal a difference of standpoint rather than a bias to be cancelled. Deguchi's point has the same form as Longino's, considered below, that assumptions shared by all members of a community are invisible to it; it differs in placing the point at the level of kinds of inquirer rather than of perspectives, and in seeking the remedy in a dialogue between inquirers of different kinds that extends to how their biases are cancelled.

To see what the different accounts imply, consider a thought experiment that pushes the example of Chapter 9, Section 1 to a limit. Its motive is to isolate one premise of human research, that critics are formed separately, by different histories, so that their number tracks their independence; the experiment removes that premise and keeps everything else. Call the construction a republic of instances, to distinguish it from Michael Polanyi's "republic of science" (Chapter 10, Section 2; Chapter 12, Section 3.2). A million artificial investigators are derived from a single base system and assigned to research, review, and replication roles. For the limiting case, suppose their subsequent experience does not diversify the background assumptions relevant to the claim. Common origin alone would not establish that further premise. The configuration maintains public venues in which any result can be criticized; it requires that criticism receive a recorded response and that the distribution of the instances' beliefs, or of the commitments that play their role (Chapter 2, Section 2), change over time in a way that is sensitive to the critical discourse (Longino 2025, §5); it encodes shared standards to which critics can appeal; and it gives every agent equal intellectual authority to raise objections. Taken as a community in its own right, the republic meets in form the first three conditions that Chapter 2 introduced from Helen Longino's account of objectivity, recognized venues for criticism, uptake, and public standards, together with the part of the fourth, tempered equality of intellectual authority (treating any perspective as having a prima facie capacity to contribute to the community's critical interactions, an equality that can be lost by failing to engage with or respond to criticism), that concerns who may raise objections (Longino 2025, §5). Whether it meets the diversity of perspectives that the same condition also calls for is, as will be seen, a separate question. Is its criticism adequate?

The answer depends on where an account places the adequacy of criticism. Table 11.1 sets out five families of answers, the last of them an ideal traced historically rather than an account; each row is a family rather than a single theory. This book chooses these five because each gives a different answer to what the republic lacks, and the choice keeps the view of Section 1's first branch, developed in Section 2, as one row among others rather than as the baseline. The five do not exhaust the possible answers. Each row states its family through the authors named in the first column. The right column summarizes the verdict this chapter draws for the republic; in every row, it is reached by applying an account or ideal to a configuration its authors did not consider.

Table 11.1. Where the adequacy of criticism is placed, and what comes into view for a republic of instances: a simplified overview

Family or ideal	Where adequacy is placed	What follows for the republic
Error routes (Goldman's condition, as extended in Section 2)	each further check adds error-detecting information that others do not supply	the number of instances counts for little; shared error routes make many critics close to one; checks that add relevant information can improve criticism, including checks maintained inside the configuration
Critical contextual empiricism (Longino)	transformative criticism among relevant perspectives (criticism in which perspectives that do not share a background assumption question it and can change what is accepted), in a community meeting four conditions	the procedural conditions can be met, but assumptions shared by all instances remain shielded, and may be reinforced
Social empiricism (Solomon)	an equitable distribution of the influences on judgment ("decision vectors") across theories, empirical ones in proportion to each theory's empirical success and non-empirical ones evenly; consensus is appropriate only when one theory has all the empirical success	the failure lies less in missing perspectives than in influences all instances share; more internal deliberation can make it worse, since groups can amplify individual biases, and pressure toward consensus can lead to groupthink and to the suppression of relevant data
Standpoint theory (Harding, Intemann)	inquiry that starts from the lives of those a community has marginalized, and a standpoint achieved through group consciousness	copying alone does not supply the relevant standpoint; engagement with marginalized lives and the achievement of group consciousness must be established separately
Mechanical objectivity (as traced by Daston and Galison, and Porter)	procedures and rules that keep the investigator's will and judgment out of the result, so that it does not depend on who produced it and need not rest on personal trust	uniform rule-following instances can count as objective in this sense; the sense guards against the bias of any one critic, not against assumptions all share, and the question passes to who designed the procedure

Longino's account is developed here because Chapter 2, Section 6 introduced it as one important way in which social organization can enter warrant, and the republic tests it directly. On this account, evidential relations are mediated by background assumptions (Longino 2025, §5), and, as Heidi Grasswick reports, a diversity of perspectives represented in a community is needed to expose questionable background assumptions that can be invisible when shared by a group of similarly situated knowers (Grasswick 2018, §4.3). As Dutilh Novaes and Dede report Longino's fourth condition, effective criticism requires diverse perspectives and the cultivation, rather than mere toleration, of dissent (Dutilh Novaes and Dede 2025, §2). Büter reproduces Longino's definition of acceptability: supporting data must be evident to the community at the time, and reasoning and background assumptions must have survived criticism from as many perspectives as are available to it, with the four conditions characterizing the community (Longino 2002, 135–136, as quoted in Büter 2012, 84). As Jaana Eigi quotes Longino, the tempered equality condition requires both that scientific communities include relevant subgroups of the society that supports them and that communities attend to criticism originating from "outsiders"; and for assumptions shared throughout a community, Longino proposes interaction across communities or, at least, openness to criticism from within and from outside the community (Longino 2002, 133–135, as quoted in Eigi 2016, 136). Since acceptability holds only for a community that meets the four conditions, acceptability even relative to a community's own perspectives then presupposes openness to criticism from outside that community. Wylie extends Longino's norm of tempered equality by proposing that communities actively cultivate collaborations with external communities whose different aims, practices, and beliefs can expose shared assumptions to criticism (Wylie 2015, 13, author manuscript). Wylie develops this proposal for archaeology; its application to artificial research is a further step here. As Yamamoto reports Longino's argument, assumptions shared by every member may be reinforced by surviving structures that screen out individual assumptions (Yamamoto 2018, 26); and, as Eigi reports Longino's account, a hypothesis is accepted as communal knowledge only if it survives criticism from multiple points of view (Eigi 2012, 51). Longino's later restatement adds that what fuels critical discourse is a diversity of perspectives, since assumptions on which everyone agrees go unacknowledged until something goes noticeably wrong (Longino 2025b, 13). Taken together, these points suggest that a republic whose internal criticism works well can entrench what its members share.

Chapter 8, Section 2 described the level at which such sharing can occur. Beneath explicit method lies a primitive epistemic organization: what a system treats as an observation, a repetition, a failure, or evidence bearing on one commitment rather than another. Instances derived from one base system can differ in their prompts, sampled outputs, and assigned roles while sharing this organization entirely. Their disagreements are then real but shallow. They can detect errors of calculation, gaps in reasoning, and inconsistencies with shared standards. In the distinction Eigi draws in setting out Longino's account, evidential criticism focuses on experimental and observational concerns, while conceptual criticism deals with theoretical and meta-theoretical issues, including whether given data are relevant to a hypothesis, and it is conceptual criticism that targets background assumptions (Eigi 2012, 51). The instances can supply evidential criticism and the part of conceptual criticism that tests claims against the theory they share. What they are especially poorly placed to detect is an error in what all of them take to count as evidence, which would require the conceptual criticism that questions the background assumptions in light of which data count as evidence at all. Longino's later statement describes testing as having this dual character, empirical testing by data and conceptual testing through criticism, and ties the identification of background assumptions to critical interaction among interlocutors with different perspectives (Longino 2025b, 13–14).

On Longino's account, the republic is therefore at most objective relative to its own perspectives, and only weakly so relative to S, to the extent that its claims have not faced criticism from perspectives the instances lack. Multiplying its members strengthens the criticism it can already perform; it does not add the criticism it cannot. The second half of the verdict presupposes that S contains such perspectives. The verdict also places S as the community relative to which acceptability is assessed. Longino's community need not coincide with S, the society to which an achievement is to be new and which relies on certified results. A specialist community may accept a result on which a wider society relies. Longino's own account leaves open whether the "we" of a community is those who produce a kind of knowledge or all those potentially affected by it (Section 9.4). The book identifies the two where the perspectives held by members of S can take part in the criticism that decides whether a result is accepted, and acceptability relative to S below refers to that case. Where they are not identified, whether perspectives belonging to S have routes into the criticism of the community that accepts the result is a further question. Where the members of S share the instances' primitive organization, as members of an artificial S derived from the same base systems may, the gap between objectivity relative to the republic and objectivity relative to S narrows without the republic's criticism having improved: acceptability relative to S then adds little to acceptability relative to the republic, and a difference that would add to acceptability relative to S can then come only through routes that neither the republic nor S contains.

On the argument of this section, there can be two kinds of such difference. One is difference at the level of primitive organization: critics formed through different histories, different sensors or data, or different ways of individuating what counts as a repeatable event. The other is a defeat relation that reaches outside the shared organization: measurement through routes the base system does not model, or criticism from outside the configuration. Yamamoto reports Longino's response to shared assumptions as openness to discussion with other communities (Yamamoto 2018, 26). Neither requires that the critics be human. Both require that the configuration contain, or remain open to, a way of going wrong that its members do not share. Longino's account points to a third kind, difference in the values and social commitments of critics. In her later statement, the context of research includes the personal and social allegiances of researchers, which can make some data and hypotheses more salient than others and some evidential relations more plausible, and background assumptions can be the vehicles by which such allegiances enter (Longino 2025b, 13). ChoGlueck and Lloyd present Lloyd's earlier study of twenty-one evolutionary explanations of the female orgasm as a case of the kind: androcentric and heteronormative background assumptions at work in many of the explanations were detected and undermined by a feminist critic, through standards of evidence that the wider evolutionary community already held, and the criticism went on to change the questions that several fields pursued (ChoGlueck and Lloyd 2023, §§4.1–4.2). Whether an artificial critic can supply that third kind is not settled by this section.

The fourth condition, tempered equality of intellectual authority, also bears on standing for dissent, the standing to object to a judgment and to refuse to give up one's own view, one of the four goods that Chapter 9, Section 2.3 proposed the division of the relations surrounding research among separate parties may have supplied without anyone designing it (Section 1). The four are error detection across boundaries, points at which an error made in one place could be detected in another; standing for dissent; boundaries across which accounts had to be given, at which one party had to explain or justify what it had done to another; and responsible parties, parties who could be held responsible for what was done. In criticism, Longino's account ties the epistemic work of dissent to tempered equality of intellectual authority and to the cultivation of dissenting perspectives rather than their mere toleration. The republic can grant every instance equal intellectual authority by design, as the thought experiment does. On the argument of this section, what the design equalizes is then authority among critics who share their assumptions, so that the condition can be met in form while the dissent it is meant to protect has no source inside. Standing for dissent in the sense of Chapter 9, Section 2.3 is a further matter: it would be held inside a configuration only by parts that hold it on the account adopted: parts recognized, in relation to those who make or are bound by a judgment that would settle a question, as entitled to object and refuse in their own name, as human members included in the configuration, for instance, may be (Chapter 10, Section 4; Chapter 12, Section 5).

Miriam Solomon's social empiricism, whose account of dissent in allocation Chapter 10, Section 4 sets out, reads the same republic differently. Solomon regards accounts that seek a community's epistemic good chiefly in improving the thinking of individuals, among them Longino's account of the elimination of individual errors and biases through criticism, as too conservative (Solomon 2001b, 68); groups, she notes, do not always correct the biases of individuals and in some cases amplify them (Solomon 1994, 57), and pressure to reach consensus can lead to groupthink and to the suppression of relevant data (Solomon 2006a, abstract). Aggregating individual decisions rather than deliberating to a consensus, Solomon argues, can produce better decisions, since dissenting positions are often associated with particular data or insights that consensus formation would lose (Solomon 2006a, abstract). What aggregation keeps, however, depends on how independent the judgments are. As Dietrich and Spiekermann note, common causes of judgment, such as shared evidence or influences from shared concepts, methods, or hypotheses, correlate the judgments even of people who do not influence one another, and the correlation usually reinforces shared tendencies and errors (Dietrich and Spiekermann 2021, §2.2). Applied to the republic, these accounts would locate its failure in influences that all its instances share and that push them the same way; neither more internal criticism nor aggregation over its members would correct that. What, then, could correct it? Solomon's account presupposes that the influences on judgment are distributed differently from person to person, as in the geological debates that Chapter 10, Section 4 described, and it takes their distribution to be assessed and corrected not in individual reasoning but at the level of the community and its institutions. There is, Solomon argues, no invisible hand of reason that makes the distribution good, so it is open to normative evaluation and to deliberate alteration, through such means as targeted grants and prizes, conferences, and raising the appeal of marginalized approaches, which act at an administrative level rather than on the individual scientist; and in her account of the consensus on plate tectonics, the community as a whole selected the theory for its universal empirical successes while no individual did so (Solomon 2001b, 71–74). Applied here, positions that assess and influence how research effort is distributed can be placed inside an artificial configuration, but on condition that their occupants do not share the influences the instances share; at least part of identifying shared influences then falls to those who do not share them, such as members of different make or parties outside the artificial configuration. An artificial configuration can aggregate independent judgments without deliberation at little cost, and so can readily use the means Solomon prefers to deliberation; but so long as its instances share influences on judgment, the aggregate is skewed in the same direction, which connects with the problem of influence in aggregation considered below.

Standpoint theory, whose account of where research starts Chapter 10, Section 6 sets out, bears on critics in one further respect. Sandra Harding argues that conventional method is relatively good at removing social values that vary among observers, since such differences stand out when others repeat an observation, but has no procedure for identifying the values and interests that all or nearly all of them share (Harding 2007, 163). Chapter 10, Section 6 discussed a configuration that takes over its view of what is significant from the record it was trained on. What this suggests for critics is that copying a base system does not establish the engagement with marginalized lives and the group consciousness that Harding's account, as Crasnow reports it, requires (Chapter 10, Section 6). Whether subsequent learning and relationships could establish them is a further question, not settled by common origin. Extended to the composition of S, the account suggests two more things. In a society of human and artificial members, the relation between the two can itself be a power relation of the kind standpoint theory examines. And a society of artificial members trained on human records starts not from nowhere but from where the human research communities that left those records stood. Harding's further demand, as Reiss and Sprenger report it, that scientists, their communities, and their practices be investigated as rigorously as the objects of their research (Reiss and Sprenger 2020, §5.2), if extended to the republic, would place among what an account of its results has to explain the evaluators who judged the quality of outputs in training the base system and so shaped its behavior and the assumptions they shared. Longino's view that the limitations of point of view are overcome by the sociality of inquirers, so that background assumptions earn reliance by surviving critical discursive interaction (Longino 2021, 14) points the same way: instances that retain the same relevant assumptions, as stipulated in this limiting case, cannot supply that interaction merely by multiplying (compare Chapter 8, Section 3.3, on systems shaped by evaluators' approval).

The last row of Table 11.1 shows that the republic's verdict can also be reversed. Lorraine Daston and Peter Galison trace an ideal of objectivity that they call mechanical objectivity and, in a formulation that Reiss and Sprenger quote, tie to a "relentless search" to replace "individual volition and discretion in depiction" by "the invariable routines of mechanical reproduction" (Reiss and Sprenger 2020, §4.1). On Galison's account the ideal took hold in scientific atlases in the nineteenth century: its advocates, horrified by the personal and the interpretive, prized images made automatically, which did not pass through interpretation, and sought to quiet the willing, intruding self, even at the cost of precision, color, and depth (Galison 1999, "Objectivity and the Scientific Image", "Revolutions of Objectivity: The Will and the Machine"). Theodore Porter, writing of quantification in science and public life, describes mechanical objectivity as following rules that "should make it impossible for personal biases or preferences to affect the outcome," in contrast to a disciplinary objectivity that rests on a consensus of experts (Porter 1995, 3–4), and argues, as Reiss and Sprenger report, that quantification is a technology of distrust and weakness, in which "objective procedures", which do not require trust in persons, take the place of subjective judgments, which do (Reiss and Sprenger 2020, §4.1). In a paper published the following year Porter sets aside the use of "objective" as roughly synonymous with truth, speaks of cultures of objectivity, and holds that the insistence on impersonal rules in science is a cultural response to conditions of distrust within disciplines and in the wider society (Porter 1996, 177–178); following rules, he had written, "may or may not be a good strategy for seeking truth" (Porter 1995, 4). Both, then, present mechanical objectivity as one historical ideal among others, distinct from truth (Daston and Galison [2007] 2010, 17; Porter 1996, 177), and both trace it historically rather than defend it. Judged by this ideal, a configuration of instances applying explicit rules could count as highly objective. Porter notes that once the numbers are in hand, results can often be generated by mechanical methods, "nowadays this is usually done by computer," and that the ideal is for expertise to be "mechanized and objectified," so that mere judgment "seems almost to disappear" (Porter 1995, 6–7); disciplines approached the ideal by relying on automatic instruments in place of subjective evaluations and by enacting rules for analyzing their readings (Porter 1996, 188). Porter also marks what this sense does not supply. Where conclusions are only a matter of applying accepted tools to solid data, researchers "cannot very well fudge their results," and outsiders lose most of their grounds for suspicion; yet statisticians disagreed, often vehemently, about the best methods, it was not clear which school provided the proper foundation, and "the statistical method" as something unitary was in many ways a myth (Porter 1996, 188–189). Porter adds that the problems of some disciplines were redefined, and their subject matter narrowed, to accord with statistical approaches (Porter 1996, 189). And quantification, he writes, "is a way of making decisions without seeming to decide" (Porter 1995, 8): economists who calculated costs and benefits were permitted to operate at the node of power because they were assumed to be self-effacing, a power Porter describes as "grounded in a presumption of impotence" (Porter 1996, 182). Applied here, the ideal would make the republic strong against the idiosyncrasy of any one instance and silent about the assumptions all instances share, and the question would pass from the instances to the procedure: who designed it, and which judgments it encodes. Porter's observation about narrowed subject matter suggests a further step. A procedure that all instances follow may not only leave their shared assumptions unexamined but also shape which problems they take up, so that what the procedure cannot treat drops out of what is asked.

The republic fixes by assumption that its members' background assumptions do not diversify. An artificial configuration may, however, be able to produce critics that differ widely at not much more cost than copying: critics formed from different data or histories, critics built from different bases, critics organized in different ways, among others. What the analysis of this section says about such a case is, first, that as regards the evidence they add, on none of the accounts does how cheaply or quickly critics were produced matter by itself. What is asked is whether the differences that result lie at the level the account takes to matter. Critics that differ in prompts or assigned roles can still share their primitive organization; critics formed from different histories or data can add differences in error routes or background assumptions. Produced differences then raise a question of their own. When one process chooses which differences to produce, that choice can itself become a source common to all the critics it produces. Goldman's warning about sources that merely reflect a common third party, as reported by Guerrero (Section 2), extends to this level: differences in directions in which the choosing process did not imagine it could err may be unlikely to arise from its choices, whether that process lies with someone outside the artificial configuration or with the configuration itself as it generates critics in the course of its inquiry. How far produced diversity adds independent information therefore also depends on whether the choice of differences is open to differences the choosing process itself did not anticipate, for instance whether that choice is divided among other parties, or whether the differences take shape through a process that no one sets, or, as on Deguchi's argument, whether inquirers of a kind the choosing process does not produce, such as human inquirers, take part in the dialogue through which biases are examined and cancelled. On Longino's account a further question is whether produced perspectives count as perspectives for the community; on standpoint theory, whether engagement and group consciousness can be produced. Who may make members and set the distribution of viewpoints among them is a question of authority, taken up in Chapter 12, Section 5.

What, then, does adding critics add? The answer depends on what "adding" is taken to mean. In the sense of evidence or information about the claim, what adds is not number but difference. Critics that are the same in the relevant respects add no evidence beyond the original critic on Goldman's condition, however many there are and however cheaply or quickly they were made. The republic's instances come close to such critics for errors that turn on the background assumptions they all share. Which differences add evidence differs by account: on the error-detection view, differences in error routes; on Longino's account, in perspectives; on Solomon's, in the influences on judgment; on standpoint theory, in starting points. On Kuhn's account, variation in how shared standards are applied is added to these. The ideal of mechanical objectivity alone asks not about differences among critics but about how the procedure all of them follow is designed. Depending on how judgments are aggregated, on the other hand, number can by itself decide the outcome. Suppose that of a hundred critics, ninety are copies of one and the remaining ten are different, and that the verdict is decided by majority. The copies then enlarge only their weight in the aggregate, without adding evidence. Number matters here not as evidence but as influence, and whoever sets the proportion of copies can sway the verdict. Solomon's point that biases shared in a community can distort aggregation bears on problems of this kind. It is for such reasons that the verdicts on the republic divide account by account, as Table 11.1 records. It follows that the builders' claim that results which pass the internal checks are certified for S (Section 1) cannot be supported by the number of checks alone. Assessing it requires being shown which account is adopted, how the checkers differ in that account's sense, and how their verdicts are aggregated.

4. Which errors matter: values in certification

Sections 2 and 3 asked what criticism adds about errors. This section asks what certification turns the support that criticism leaves into, and against whose valuations it does so. Chapter 9, Section 2.1 took certification as one of the relations that connect research to S. In this section, certification is considered through the judgment that realizes the relation of Section 1: the judgment that a result is good enough to rely on for a specified use. In human research, methodological conventions such as significance levels, acceptance in peer review, the standards of regulatory approval, and the like have fixed the point at which a result may be treated as established for a use. Put abstractly, what they have done is to turn support, which comes in degrees, into a standing on which others may act without checking again. This function is needed not to produce knowledge but because the results produced are used by someone who does not check them again. Since evidence does not make a claim certain and a finite user cannot check everything again, the function is needed whatever does the research and whatever uses it, as long as results are used by others. When certification moves inside a configuration, the configuration can also come to screen results and to set and apply the thresholds at which a result is treated as established for a use.

What such a threshold fixes is the subject of what is called the argument from inductive risk. Inductive risk is the risk of error that remains when a claim is accepted or rejected on evidence that does not make it certain. On this argument, how much support counts as enough depends on how bad the different errors would be. The support in question is, on a probabilistic representation, the level of support reached rather than how much a check raised it (Chapter 2, Section 5.2). Torsten Wilholt traces this argument from statisticians to Richard Rudner and Heather Douglas (Wilholt 2013, §2). If the argument is granted, bringing certification inside brings inside not only procedures of checking but a judgment of value: which errors to avoid, and how much. In human research this judgment was divided between the conventions just named and the judgments of referees, regulators, and others; and on Wilholt's view, reliance on the choices that conventions leave open works stably where producers and those who rely on them share a sense of what the right valuations of the possible outcomes would be (Wilholt 2013, §§4–5). This sharing comes, like the convictions that Chapter 6, Section 4.2 found guiding where an error is attributed when a result goes against expectation, from outside the result at hand, but it concerns how the consequences of outcomes are valued; Section 5.3 returns to it. Where what produces is a machine, or where S is artificial, that sharing cannot simply be presupposed. Where the producing configuration and the members of S share how they were made or their lineage, the sharing may be easier to presuppose; even then, whether those who bear the costs of errors without relying on the result share that sense remains a separate question (see below).

Chapter 5, Section 4.2 separated judgments of what to pursue from judgments of what is likely true, and Chapter 6 asked what a result bears on and how firmly a passed check settles it. This section takes up a judgment different from both: whether the support so assessed is enough for others to rely on the result for a use. Whether it is enough depends on what would be lost if one acted on the result and was wrong, and what would be lost if one did not act and was wrong; use enters through these losses. The same result can be enough or not depending on what is done on its basis and who uses it: the same evidence about a drug's safety may be enough to judge it worth further study but not enough to approve its wide prescription. Whether a result matters to a given party at all is a different question, taken up in Chapter 9, Section 3 and in Section 7 of this chapter. Chapter 6, Section 3.3 treated who sets what counts as passing as a matter of how a configuration is assembled; this section treats the judgment of value contained in that threshold.

This section therefore asks to whom the weighting of errors contained in a threshold answers. To answer to someone here covers three things together: whose valuations fix the weighting, who can object to it, and who is responsible for it. Even among authors who grant the argument from inductive risk, positions on this question divide. To compare them, the section separates what lies behind a single threshold into five items: the evidence; the losses assigned to errors; the authority to choose those losses and the threshold; the authority to rule on individual results under the threshold and later to revise the rule itself; and the background assumptions through which data become evidence. Keeping evidence and losses apart follows Douglas's remedy, the background assumptions come from Longino's account, and the two kinds of authority correspond to the decision rights of Chapter 12, Section 2. Of the five, the evidence and the background assumptions are epistemic matters of the kind treated in Chapters 2 and 6, while the losses and the two kinds of authority concern use and decision. The point of the argument from inductive risk is that the latter enter the former judgment of what to accept as established, and that is what the reply, considered below, of stopping at reports of probability rejects. Once the items are separated, one can ask of each whether a configuration can take it on internally or whether it is a relation with those who rely, those who bear the costs of error, or those who can contest the assumptions. The section first establishes where in research this weighting enters, then asks whether reporting probabilities alone would avoid it. It then takes up accounts of whose valuations the losses answer to, turns to an account of background assumptions, hands the questions of authority to Chapter 12, and ends with a point of entry outside the threshold, the order of testing. The section also takes up part of a question that Chapter 1, Section 1 left to later chapters, when a result is to alter a decision, insofar as it concerns the threshold at which a result is treated as established for a use.

It is worth establishing first where in research this weighting enters. It does not lie only at the point where results are handed to decision-makers. In one doctoral study of lymphatic spread in head and neck cancer, for example, the author retained one of the connections in a statistical model that represent routes of spread between lymph-node levels, although the model-comparison criterion neither supported nor discouraged it, because it was clinically meaningful and anatomically plausible; and the author cited the lower danger of overestimating rather than underestimating involvement as one reason for rejecting another connection (Ludwig 2023, 102, 105–107). The second decision is the one that bears on this section. In it, the differing consequences of error enter at the level of a single modeling choice inside research, the kind of choice that Chapter 6 treated as a question of evidence. The same can hold whenever an artificial system screens claims, decides whether to stop an experiment, flags misconduct, or recommends whether a risky intervention may proceed. A numerical threshold is not neutral merely because it is numerical. False rejection can suppress a valuable line of inquiry, unsafe continuation can expose others to harm, and delayed correction can let an error propagate; and different parties may bear those costs. Wilholt too notes that the issue is particularly sensitive where the consequences of false positives and of false negatives would be borne by different parties (Wilholt 2013, §4). A configuration that brings certification inside can therefore weigh errors not only when it applies a threshold at the end but at each such choice.

One reply holds that the weighting can be avoided by estimating and reporting probabilities and leaving the weights to those who use the result; as Reiss and Sprenger report, the reply goes back to Richard Jeffrey (Reiss and Sprenger 2020, §3.2). An accurate system may deserve reliance in estimating a relevant probability. Douglas argues, however, following Rudner, that even a statement of the uncertainty of a claim involves judging that the assessment of that uncertainty is accurate enough, a judgment that weighs the consequences of error (Douglas 2007, 10). Douglas further argues that the same error costs figure in the process that leads to an estimate, in the choice of method and in the characterization of data; choosing a significance level in a hypothesis test is one example (Douglas 2000, §§4–5). Stephen John points to a further place at which such choices enter reports of probability: a probability is reported relative to a body of evidence, and the criterion that decides what enters that body, such as the IPCC's reliance on peer-reviewed work, trades the risk of admitting false reports against that of excluding true ones (John 2015b, §3). John grants that reports indexed to several bodies of evidence could avoid the trade, but holds that making them usable would require interpreters who face the same problem (John 2015b, §§3–4). John does not extend the point to artificial systems; applied to them, a configuration's rule for admitting sources to its retrieval store (Chapter 2, Section 4.4) is such a criterion. Where those who use the reports can handle indexed reports directly, interpreters may not be needed, but even then which bodies to index to is chosen. On the accounts considered here, the accuracy of the estimate does not by itself settle whose losses from error receive what weight.

Whose valuations, then, do the losses assigned to errors answer to? Wilholt draws two points for cases in which others, such as other researchers or the public, use the results of research without checking them themselves. First, the reliability that reliance presupposes is not a single level fixed by the aim of truth: whether a result is reliable "enough" cannot be settled without a valuation of the consequences of the possible outcomes, because methodological decisions are trade-offs. For a given test, making a standard stricter to reduce the risk of wrongly reporting that an effect or relation exists usually yields more cases with no result and can raise the risk of wrongly reporting that it does not; in Wilholt's terms, methodological decisions trade the reliability of positive results against the reliability of negative results and the power of an investigation, by which Wilholt means the rate at which it delivers definitive results (Wilholt 2013, §§2–3). Second, methodological conventions settle some of these decisions, such as the convention, found in many fields, that 0.05 is the highest acceptable significance level; but such conventions often only limit the range within which the trade-off is struck, some decisions are not covered by any convention at all, and such decisions are often not made explicit in published work; some involve tacit judgment that could not be fully documented even without limits of time and space. On Wilholt's account, those who rely on a result therefore work from an estimate of the valuations in light of which the decisions were probably made (Wilholt 2013, §§2–4). Relying on a result then involves relying not only on its producer's competence and honesty but on its having struck these unstated trade-offs in light of the valuations the reliant party would expect. Wilholt presents this reason as independent of Hardwig's (taken up in Section 5.3), which rests on the premise that peer review and replication cannot detect fraud effectively, and states that it would remain convincing even if there were an effective mechanism for detecting fraud (Wilholt 2013, §4).

Two readings of what this asks of those who certify are available. Wilholt argues first that, for decisions left open by existing conventions, a research community can do justice to the trust placed in it only by actually basing them on the kinds of valuations the public expects, and, assuming that striving to deserve that trust is at least sometimes and in some respects desirable, takes this as a further reason why science needs trust and not mere reliance; on this argument, when research communities adopt new methodological conventions, the distribution of inductive risk implicit in them should likewise correspond to the valuations the public expects. Wilholt adds that this does not mean always valuing outcomes in line with the public, since a public that trusted climate scientists to relieve it of an inconvenient conclusion would give them no reason to relax their standards (Wilholt 2013, §4 and n. 22). On one reading, what this supports is correspondence: the trade-offs made on behalf of those who rely match the valuations they expect. The other is a stronger reading, which this book formulates for comparison and which Chapter 12, Sections 1 and 7 take up: the trade-offs would also have to be answerable to those who rely, in the sense that they can object to them. For a configuration, the difference is between its being enough to check whether the weighting matches what is expected and its needing a relation in which those who rely can object. Wilholt does not discuss artificial systems; applied to them, on either reading a log of a configuration's choices can show its valuations as well if it records the losses adopted and the reasons for its choices, but recording them does not by itself establish that those valuations were appropriate for those who will rely on the result.

Douglas places the same weighing within the responsibility of whoever makes the judgment. On this account, scientists have a responsibility to weigh the different costs of false positives (accepting a false hypothesis) and false negatives (rejecting a true one). These costs therefore bear indirectly on judgments of evidential sufficiency throughout inquiry, from the characterization of data to the acceptance of claims, without serving as evidence that a proposition is true (Douglas 2000, §§3–5). Douglas grounds the responsibility in the general moral responsibility of any agent not to be reckless or negligent about the consequences of error (Douglas 2003). The account also restricts the role values may play at these points: values of any kind may play only an indirect role, weighing the importance of the uncertainty that remains rather than serving as reasons for the claim; a direct role is legitimate in choices such as which problems to pursue (Douglas 2016, 10–11). Douglas's own remedy keeps the evidence apart from the weighing of the costs of error. Scientists are asked to make the values that weigh uncertainty as explicit as possible, so that those who rely on their advice can tell whether values are weighing the sufficiency of evidence or taking its place; the test for the latter is whether an expert can say what evidence could change the expert's conclusion (Douglas 2008, 19; Douglas 2017, 16–17). Brought to bear on an artificial system that sets or applies such a threshold, Douglas's argument raises the question of which responsible agent answers for the weighing, a question Chapter 12, Section 6.3 takes up.

The distinction between direct and indirect roles separates values that enter as weights on losses from values that serve as reasons for the claim itself. The distinction has, however, itself been challenged. Kevin Elliott notes that the line between them may be less clear than it first appears: Douglas ties the indirect role to unintended consequences of error and the direct role to consequences a scientist aims at, yet a scientist could have intended consequences in mind when deciding how much evidence is enough. Even where the line is clear in principle, scientists may be unsure, even in their own minds, whether a value guided their choice of technique, model, or interpretation directly or through concern about the consequences of error. Elliott also reports Daniel Steel and Kyle Whyte's argument that the distinction may not suffice to prevent problematic influences of values: pharmaceutical companies that fail to publish results indicating that their products are problematic could be moved by the consequences of error, such as patients' hastily stopping a drug and harming their health, or the companies' own economic harm. What they object to is that the failure to publish prevents later meta-analyses of the drugs from providing severe tests; on this objection, an indirect role for values does not by itself make their influence acceptable. Elliott notes that Douglas could respond by arguing that the companies appeal to inappropriate values, that is, by constraining which kinds of values may play that role (Elliott 2013, §§3–4). Neither Elliott nor Steel and Whyte discuss artificial configurations; applied to them, even where a configuration logs evidence and losses separately, the log alone may not show in which role a value entered, and a weighting that stays in the indirect role can still reflect the interests of those who build and supply the configuration. Douglas's own test, however, asks not what a log shows but whether there is evidence that would change the conclusion, and applied to a configuration it can be put as a trial: whether, for a certified result, the evidence that would reverse the verdict can be specified, and whether the verdict reverses when such evidence is supplied. A verdict that no possible evidence could reverse would, by this test, rest on something other than evidence; the test asks whether, with the weighting applied, the defeat relation of Section 1 is preserved, and it leaves open whose interests the weighting serves.

On John's reading, Douglas's account would have standards vary with all the foreseeable consequences of error (John 2015a, §2). John's own position, which grants that scientists face problems of inductive risk but disputes how they must respond, ties the weighting instead to an audience. When a claim is addressed to hearers who can be identified in advance, the standard for asserting it can follow the standard at which they ought to accept it; but publication addresses audiences that cannot be identified even in principle, and for public claims John defends using fixed, high standards for hearers with differing practical interests (John 2015a, §§1–3). John adds a duty to "speak out" about claims that are well enough established to warrant action by some in the community, though not by every rational agent. The duty, as John grounds it, rests on scientists' more general civic duties and is not framed in terms of those who bear the costs of error; it can nonetheless be read as one route by which those who bear such costs without being among a claim's hearers come in (John 2015a, §4). Ahmad Elabbar takes views of this fixed-high-standards kind as a target (Elabbar 2026, §3.2) and argues that such a standard is not neutral among those affected: where evidence is unequally distributed, it yields many findings about those for whom data are plentiful and few about those for whom they are scarce, and where the latter are thereby disadvantaged in fundamental interests and varying the standard can mitigate that disadvantage, Elabbar holds that justice gives reasons to vary it (Elabbar 2026, §§4–5). Neither author discusses artificial configurations. Applied to a configuration whose results enter a record that parties it cannot identify will use, John's argument would favour a fixed standard; Elabbar's asks in addition about whom the configuration's evidence is plentiful. Where a configuration's certifications go instead to hearers it can identify in advance, such as an identified decision-maker or a system that will act on them, John's other branch applies, and the standard can follow the standard at which those hearers ought to accept the result; which branch applies is then settled by how outputs are released, to identifiable parties or into a record, not by the configuration's being artificial. Where a fixed high standard leaves no action-guiding findings about parties on whom evidence is thin, and their fundamental interests suffer, Elabbar's argument gives reason to vary the standard and points in a different direction from John's argument for a fixed one.

Longino's account holds that values bear not only on the standard at which evidence counts as sufficient but on what counts as evidence at all. It works with a distinction that Julian Reiss and Jan Sprenger report as widely drawn, between cognitive values, also called epistemic or constitutive, such as predictive accuracy, scope, simplicity, and coherence with other accepted theories, which are taken to be indicative of a theory's truth, and contextual values, the moral, personal, social, political, and cultural values of the setting in which science is done (Reiss and Sprenger 2020, §3.1). Because data count as evidence for a hypothesis only in light of background assumptions, contextual values can, on Longino's account, shape what a research program concludes through the assumptions that make some data and hypotheses salient and some evidential relations plausible (Section 3). ChoGlueck and Lloyd report, quoting Longino's 1990 book, that on her account constitutive values check the role of contextual values and cultural assumptions through the empirical and conceptual evaluation of assumptions, and they read the account as treating values as heuristics whose use a community can assess by their empirical fruitfulness, not as substitutes for evidence (ChoGlueck and Lloyd 2023, §§3.3, 4). Criticism of such assumptions, Longino holds, needs perspectives developed enough to offer alternative ways of interpreting the data, some amounting to fully developed theories and others to promising research programs (Longino 2025b, 13); and assumptions that all members of a community share are hard to bring under it (Section 3). Applied to an artificial system that interprets data or sets a threshold, the two accounts ask different things of it. On Douglas's, values weigh only the uncertainty that remains when a claim is assessed; on Longino's, what is to be checked includes not only the threshold but the background assumptions through which data are interpreted as evidence, and what is asked is whether the assumptions at work can be made out and criticized. Of the five items separated at the outset, the background assumptions come from Longino's account.

The remaining two, the authority to choose losses and thresholds and the authority to rule on individual results and later revise rules, concern a question other than the one the accounts above address, whose valuations fix the losses: the question of legitimacy, of who may choose and revise them. Being able to criticize background assumptions and having the authority to choose losses and thresholds anew are also different things. There is a further question, whether the weighting of errors can itself later be revised in light of its consequences. A view with roots in Dewey's account of valuation, on which judgments of value concern the consequences of acting on them and can be revised through inquiry into those consequences (Chapter 9, Section 3), would hold so if extended to these judgments; this is disputed (Chapter 12, Section 1), and Chapter 12 takes it up together with the question of who may choose and revise. Its Section 2 places these kinds of authority among decision rights.

Finally, values can enter from outside the threshold as well, through which claims are checked and when. Oscar Varsavsky, in the debate set out in Chapter 10, Section 6, argued that a false hypothesis can serve as scientific backing for decisions that affect millions of people while its refutation is delayed, that science checks claims in an order that reflects priorities, and that the influence of ideology on truth, exercised through such priorities and the choice of methods, works mostly by delaying it rather than by falsifying it (Varsavsky [1971] 1975, 46–47, 50). Thomas Moro Simpson replied that partiality in what is examined or made known does not imply a lack of objectivity in results' relation to truth, that where truths already acquired are concealed the remedy is to denounce the concealment, and that the question of priorities is a further topic, one for science policy rather than for epistemic objectivity (Moro Simpson 1975, 83–84). The debate separates three dimensions: whether results are true, to whom they are important, and when needed truths are reached (Chapter 10, Section 6). Read in relation to certification, it marks a difference between attending only to the relation of results already obtained to truth and attending also to the period during which erroneous results remain in use unchecked. Varsavsky's point concerns not the threshold applied to a claim but which claims reach it and how soon. Neither author discusses artificial configurations; applied to a configuration that decides which of its results to re-examine and in what order, that decision lies where allocation is coupled with criticism and certification (Chapter 9, Section 2.4). In the absence of correction from elsewhere or of a halt in use, the lower a result's checking is ranked, the longer an error in it can remain in use.

Five things can then be discussed about a weighting that has moved inside a configuration. First, of the evidence and the losses, one can ask what a configuration can take on internally. Douglas and John argue that even a configuration that stops at estimating probabilities weighs errors through what it admits to its body of evidence and through its choices of method; if that is granted, making these choices explicit and logging them is a part a configuration can take on internally. Wilholt, however, holds that in human research some of these choices involve tacit judgment that cannot be fully documented, and on the argument Elliott reports, those who make them may themselves be unclear in which role a value entered. How far a configuration can log them depends on how it is assembled; and what a log can show is what was done, not that the valuations were appropriate for those who rely.

Second, one can compare accounts of whose valuations fix the losses. On Wilholt's account they do so through correspondence with the valuations those who rely expect; on the stronger reading, through those parties' capacity to object; on John's, through fixed standards for hearers with differing practical interests. Where those who rely, or the hearers, are outside the configuration, these become relations to parties outside it, whether or not the producing process is artificial. On Douglas's account it is the responsibility of whoever makes the judgment, and for an artificial system it becomes the question of which responsible agent answers (Chapter 12, Section 6.3). Where what produces is a machine and those who rely are human, correspondence with those who rely becomes the question whether the machine's weighting matches the valuations humans expect, one form in certification of the concern that Chapter 8, Section 3 found structurally similar to the AI alignment problem.

Third, those who rely on a result and those who bear the costs of its errors can be different parties, and where S is artificial they can differ in kind: its members may rely on a result whose errors fall on humans or on an environment that an experiment exposes to risk (Chapter 12, Section 1). That the weighting answers to those who rely does not settle whether it answers to those who bear the costs, and there John's and Elabbar's arguments can point in different directions. Fourth, of the background assumptions and the two kinds of authority, one can ask, having separated being able to criticize from having authority to choose and revise, to whom they are open. Fifth, outside the decomposition of the threshold, which results are re-examined and in what order remains a place where allocation and certification are coupled inside a configuration.

In terms of the sorting of Chapter 9, Section 6, both the need for this function and the question of whose losses fix its standard come from relations with those who use results without checking them again, and so, so far as those users are outside the configuration, belong, of the four kinds, to relations with parties outside it; that the function was stabilized by conventions and a shared sense can be read as a premise about how human research communities have been organized, and, so far as those conventions rest on tacit judgment that cannot be fully documented, as one about human realizers as well. Where those who rely or bear the costs are outside the configuration, it is here, on Wilholt's and John's accounts, that the part of certification lies which cannot be wholly brought inside a configuration. On Douglas's account, the line is drawn by where in the configuration something can answer as a responsible agent. Taken together, these points bring into view what can happen to research when certification moves inside an artificial configuration. Since the weighting enters not only at the single point at which a threshold is finally applied but at each modeling choice and each decision whether to stop an experiment, which results enter the record as established for S is shaped in advance by the weights of losses built into the configuration. If those weights reflect the settings or interests of those who build and supply the configuration (applying the argument Elliott reports), then merely revising those settings can change the range of findings counted as established, without anyone's being aware of having decided on a new threshold. In research, moreover, errors in the weighting are hard to bring to light. Candidates set aside by false rejection, and lines cut off, are not pursued afterwards, so the results that would show the rejection to have been an error are usually never produced; and since results are new to S, there is often from the outset no other source against which to check whether the weighting was appropriate. The accumulation of findings about parties on whom evidence is plentiful (Elabbar), and the longer use of errors in results ranked low for checking (Varsavsky), can enter the record of research by the same route.

5. Reliance, testimony, and trust in certification

Sections 2 and 3 asked what criticism inside an artificial configuration adds, and Section 4 asked which valuations enter the judgment that a result is adequate for a use. What remains is the question of what grounds someone who does not repeat those checks has for relying on an artificial configuration's certification. Certification gives others reason to rely on results they have not checked themselves (Section 1). On what grounds, then, and on what or whom, does someone who uses a result without checking it again rely? Earlier chapters have touched on this question. Chapter 2 asked what it takes for a result to become knowledge for S, including when one knows through others; Chapter 8, Section 3.4 asked what has to pass to the relying side; and Chapter 9 placed certification among the relations that tie research to S. This section applies those discussions to artificial configurations that issue certification from internal checks. What is required between the two parties, the one who relies and the one who certifies, is what the epistemology of testimony and the philosophy of trust have dealt with.

In human research, reliance on certification passed through persons: certification was a person's assertion, those who relied counted on the certifier's honesty and competence, and someone answered when it went wrong. John Hardwig, as Longino summarizes his argument in her encyclopedia entry on the social dimensions of scientific knowledge, discussed researchers' reliance on others' results whose evidence they are not in a position to evaluate themselves (Longino 2025, §2), and later argued that such reliance involves trust in the other's truthfulness, competence, and conscientiousness (Hardwig 1991, 699–700). When an artificial configuration certifies, none of this can be presupposed. Whether its output is an assertion at all, whether reliance must be trust in honesty and competence, and who answers when it goes wrong each become questions. The section therefore separates what an artificial configuration can supply internally from what is required as a relation to those who rely or to those who stand behind the certification, and asks how far the latter can be placed inside the configuration.

To that end it separates three questions and takes them in turn. One is what kind of act certification by an artificial configuration is: the configuration's own assertion, the assertion of a group for which it speaks, or the reading of an instrument (Section 5.1). The second, a question the epistemology of testimony has dealt with, is what grounds reliance on a result one has not checked (Section 5.2). The third, a question the philosophy of trust has dealt with, is whether that reliance must be trust (Section 5.3). Chapter 8, Section 3.4 placed the positions bearing on these on a single ordering, by what has to pass to the relying side. Those positions, however, answer different questions, and the three are separated so as not to treat answers to them as rival answers to one question. They are not the only questions: questions about the ethics of trust, such as when a relation of trust is morally decent, are left aside. The reliability of reliabilism in Section 5.2 is the dependability of a process; how trust is to be distinguished from it differs among the accounts compared in Section 5.3.

After the three, Section 5.4 uses their distinctions to ask what those who rely can read from agreement among a configuration's internal checkers; Section 5.5 draws together what has emerged on the three questions; and Section 5.6 turns to the case in which the achievement sought is understanding rather than knowledge.

5.1 What is certification by an artificial configuration?

Of the three questions, the first is what kind of thing certification by an artificial configuration is. If it is an assertion, the norms of assertion and reliance on a speaker bear on it; if it is the reading of an instrument, reliance on those who make the instrument and stand behind it does. Which it is treated as changes which norms apply and who answers. Chapter 2 (Sections 6 and 9) and Chapter 8, Section 3.4 also noted this distinction and said that the accounts behind it would be set out here. The question there was who knows a result and who answers for it. The question here is what kind of act certification is, that is, a report offered so that others may rely on it without checking. Lackey's account of group assertion, which differs from the same author's account of group belief seen in Chapter 2, is developed here for this question, together with the application of Williamson's knowledge norm that Chapter 2 only touched on.

Elizabeth Fricker defends an internalist, local reductionist account of testimony. Reductionism here is the view that the justification for believing testimony reduces to other reasons the hearer has; it is local in requiring such reasons speaker by speaker and testimony by testimony, and internalist in requiring that the hearer can be aware of them. On this account, a hearer's justification for believing what they are told is the hearer's justified belief that the speaker is trustworthy and has the epistemic authority to take responsibility for what they tell; the empirically ascertainable ground a recipient often and typically can have for such a belief is the speaker's sensitivity to the norm of assertion (E. Fricker 2026, 2877–2879, 2881). On Fricker's account, current AI systems lack the communicative intentions and the sensitivity to social norms that assertion requires, so their outputs are designed to seem like assertions but are not (E. Fricker 2026, 2861, 2870–2871). Reliance on a system that produces information no one has known can still be justified by an empirical account of why its outputs are reliable, but what is acquired is then knowledge from an instrument rather than from testimony, and responsibility for trusting the outputs stays with the human user (E. Fricker 2026, 2863 n. 8, 2880–2881).

Sanford Goldberg distinguishes relying on instruments from relying on speakers, but argues that the former can still manifest epistemic dependence (Goldberg 2020, abstract). As Inkeri Koskinen reports Goldberg's argument, instruments are not responsible subjects but products of others' epistemic work; she connects this to users' trust in those who designed and built them (Koskinen 2024a, 463). Dominic Dixon's report of Goldberg's collected-essay version likewise describes design dependence on the competences of those who design, build, and maintain technologies, and environments deliberately arranged to reduce the cognitive burden of acquiring knowledge (Dixon 2025, §4.3.1). On Goldberg's account, an instrument's unreliability on one occasion does not by itself settle whether a belief formed from it is justified; that depends on whether the failure reflects an unmet expectation the user was entitled to have of the designers and manufacturers (Goldberg 2020, n. 9). Applied to a configuration's certification, this reported account makes reliance run through the epistemic work of its designers and makers. Whether trust in those designers and makers can support reliance on essentially opaque AI is disputed: Koskinen argues that even developers may lack the understanding needed to take full responsibility for its workings (Koskinen 2024a, 464–465; Section 5.3).

Jennifer Lackey's theory of group assertion places the one who answers differently. On this account, as Neri Marsili sets it out in his own terminology, a group makes a proxy assertion, which Lackey calls an authority-based assertion, that p when p belongs to a domain and a spokesperson reasonably intends to convey the information that p through an individual act of communication, has the authority to convey information in that domain, and acts in this way in virtue of that authority as a representative of the group; a group makes a collective assertion when its members coordinate individual acts so that they all reasonably intend to convey p together (Marsili 2023, 371 and n. 9). The intention condition bears directly on automated outputs. Applied to a group whose automated system aggregates its members' inputs and issues a report that no one reasonably intends to convey, neither definition is met, so the report is not, on this account, the group's assertion. Marsili proposes a variant that replaces the intention condition with conditions drawn from a commitment-based account of assertion: the spokesperson utters an expression with content p, presents p as true, and undertakes assertoric commitment to p on the group's behalf, having the authority to undertake it; commitment here includes a responsibility to provide reasons if the assertion is challenged and liability to social sanctions if it proves false (Marsili 2023, 373–374). Applied to a configuration, Lackey's view would allow a certification to be the assertion of a collective for which the configuration speaks, provided the configuration can be a spokesperson in that sense, which returns to the question of communicative intention that Fricker raises. On Marsili's variant, the question becomes instead whether anyone with the authority to do so undertakes commitment to the certification on the group's behalf.

Only if a certification is an assertion, the configuration's or a group's, do norms of assertion bear on it. On Timothy Williamson's account, assertion has one constitutive rule, to assert only what one knows. Williamson argues from a lottery: someone who tells a ticket holder flat out "Your ticket did not win," on the merely probabilistic ground that it is one of very many, lacks the authority to say so, however high the probability is made, and the fault does not depend on how bad a false assertion would be; the natural complaint is that the speaker did not know it. On this account, merely probabilistic grounds, however strong, therefore warrant only the assertion that something is probable; there are also speech acts, such as conjecturing, whose evidential norms are weaker (Williamson 2000, 238–241, 243–244, 246–249). If a configuration's reports are assertions in this sense, the configuration, or the group for which it speaks, is the one whose knowledge the rule concerns, and by Williamson's rule a report resting on a calibrated score alone would be warranted only as a report that the result is probable, and whether an output is offered as a conjecture or as an assertion would become part of what those who evaluate the configuration check. A configuration whose reports rest on calibrated scores alone and that keeps to this rule would then stand where the reply considered in Section 4 stands, reporting probabilities and leaving the weights to those who use the result, and the arguments of Douglas and John against that reply apply to it: the weighting of errors is not avoided but enters the estimate, and turning graded support into a standing on which others may act falls to those who rely. The knowledge norm is disputed. Igor Douven and Lackey, among others, defend a norm on which one may assert what it is rational for one to believe, and others hold that the support an assertion requires varies with the conversational context, including what is at stake (Pagin and Marsili 2025, §5.1.4). On such norms, whether a calibrated score suffices for asserting a result would turn on whether it makes belief in the result rational for the asserter, and on the context-sensitive versions also on the use for which the result is certified (Section 4). If its reports are readings of an instrument, these norms do not apply to it.

In sum, certification by an artificial configuration can be treated as the artificial configuration's own assertion, as the assertion of a group for which the artificial configuration speaks, or as the reading of an instrument. Which applies is not settled merely by the configuration's being artificial. On Fricker's account, the outputs of present AI are not assertions, though the possibility is left open for future systems. On Lackey's account, an authority-based group assertion requires a spokesperson who reasonably intends, in virtue of their authority, to communicate the content; a collective assertion instead requires members to coordinate individual acts so that they all reasonably intend to convey the content together. If the artificial configuration itself is the spokesperson, it must be capable of that intention; if a member of the configuration or another party takes up the report and communicates it, the spokesperson conditions concern that party. If certification is treated as an instrument reading, applying Goldberg's reported account directs reliance toward the epistemic work of its designers and makers; whether trust in those designers and makers can support reliance on essentially opaque AI remains disputed. Which it is treated as is presupposed by the grounds of reliance considered next and by the question of trust in Section 5.3.

5.2 What grounds reliance on an unchecked result?

What, then, grounds such reliance? The answers are not limited to four, but four are considered here: reductionism and non-reductionism, the central opposition in the epistemology of testimony, and reliabilism and assurance views. Some of these accounts have already been used for other questions. Chapter 2, Section 4 discussed reliabilism, through the generality problem, for beliefs formed through artificial systems, both a system's own commitments and human beliefs that rely on its outputs, and Chapter 3, Section 5.3 discussed Juan Durán and Nico Formanek's computational reliabilism with respect to researchers' reliance on opaque simulations. Chapter 8, Section 3.4, anticipating process reliabilism, treated in this section as an account of reliance, and the assurance views introduced here, ordered positions by what has to pass to the relying side. The question here differs from those. What is relied on is a certification that comes out of the internal checks of an artificial configuration (Sections 2 and 3), and the question here is what each account requires for the justification of those who rely on such certification, and so what it would have an artificial configuration supply. And because what is relied on comes out of internal checks, two further problems arise: the routes by which the checks' findings reach those who rely, and the use of checkers' verdicts to measure those same checkers.

On reductionist accounts, a hearer is justified in believing what they are told only if they have positive reasons, not themselves derived from testimony, for thinking the source reliable; on non-reductionist accounts, a hearer has a defeasible but presumptive right to believe what they are told unless they have an undefeated defeater (Leonard 2021, §§1.1–1.2). Applied to an artificial configuration, on the first, what a member of S needs is positive reason, not derived from testimony and available to that member, to think the configuration reliable for the use at hand. A reliability record is the typical case, though some, such as Fricker, allow a background assumption, warranted by inference to the best explanation, that some mechanism keeps the outputs reliable (E. Fricker 2026, 2880). On the second, the burden shifts to defeaters: what the hearer needs is the absence of an undefeated defeater. That condition, in the form stated here, concerns the defeaters the hearer now has, not whether any could reach the hearer. As Leonard sets it out, some non-reductionists add that the hearer must be counterfactually sensitive to, or on the lookout for, defeaters in the environment (Leonard 2021, §1.2). On such versions, whether S has routes by which a defeater could reach it is part of the condition itself. On the minimal version, whether there are such routes lies outside the condition. This book nonetheless raises it as something to be asked in building and evaluating artificial configurations whose certifications are relied on in this way, for the following reason. Without such routes, an error found by internal checks does not act as a defeater for those who rely, and the absence-of-defeaters condition holds alike whether or not the result is in error; in that sense, however strong a configuration's internal checks, their findings make no difference to the justification of those who rely. Those routes are the defeat relations of Section 1, seen from the side of the one who relies.

Ramón Alvarado, writing about computer simulation, limits where such a presumptive right holds. On Alvarado's account it is reasonable for a layperson to heed a meteorologist's warning without examining the methods and evidence behind it, but not for the meteorologist to take the same attitude toward the simulation, which is not itself a source of expert testimony; what laypeople trust is the judgment of human experts, whose expertise has been established beforehand, not the simulation's output directly (Alvarado 2020, §5.2.2). Alvarado does not discuss artificial certifiers. Applied to an artificial configuration, the account would make the non-reductionist answer depend on whether its certification is testimony from a source with expert standing or the reading of an instrument (Section 5.1), and in the latter case what a member of S would trust is the community that certified the instrument.

Reliabilism gives a third answer. On Goldman's process reliabilism, a belief is justified if it is produced by a reliable process, without the believer needing reasons for thinking the process reliable (Goldman and Beddor 2021, §§1.1, 2.1); Goldberg's testimonial reliabilism makes justification depend in part on the reliability of the speaker's process (Goldman and Beddor 2021, §3.2). Applied to an artificial configuration, what such views ask about is the reliability of the recipient's belief-forming process, including the production of the certification and the way it is taken up, and the recipient need not have reasons concerning that reliability. Nothing in that requires a human at the end of the chain. The configuration, however, supplies only part of that process; if a recipient misunderstands a certification or uses it outside its scope, the process as a whole can become unreliable. And a belief's being justified is not the same conclusion as a result's being usable for a given purpose, which involves the valuations of Section 4. The view carries known objections. Whether a process is reliable depends on how it is typed, which is the generality problem that Chapter 2, Section 4 noted. Jun Otsuka's observation, also seen in Chapter 2, Section 4, can be read as one instance of it: benchmark scores establish reliability relative to a model and a test set, whereas those who must decide on a use ask whether the model is safe and effective for their task (Otsuka 2023, 133, 141–142). The verdict changes with which type of process one asks about. The other objection is the bootstrapping problem: a subject who infers a gauge's reliability from the agreement between its readings and beliefs formed by reading it seems to gain justification too easily (Goldman and Beddor 2021, §2.4). An artificial configuration that treated its checkers' verdicts as correct merely because those checkers issued them, and then established the same checkers' reliability from agreement with those verdicts, would face a version of this objection, as would members of S who took that agreement as their reason to trust the checkers. Agreement among internal checkers does not by itself establish this circularity, though. How much evidence such agreement gives about the result is a separate question, taken up in Section 5.4. Neither Durán and Formanek's computational reliabilism nor Jakob Ortmann's account, introduced in Chapter 3, Section 5.3 and Chapter 8, Section 3.4 respectively, adopts this pure form. Durán and Formanek require that the sources conferring reliability on a simulation be shown reliable in turn, and leave open where the resulting chain ends (Durán and Formanek 2018, §§3.2, 4, n. 6); Ortmann counts expert judgment among what establishes a tool's reliability and adds that judging whether a tool's use is appropriate for a task is a value decision par excellence that can be deferred to a trusted party (Ortmann 2025, 17–19).

Assurance views give a fourth answer. On these views, a speaker who tells a hearer that p invites the hearer to trust the speaker, assuring the hearer that p is so. Evidence counts for a proposition whatever anyone intends, as a fingerprint shows that someone was present whether or not they meant to leave it; an assurance counts for p only because the speaker intends it to, so the reason it gives is not evidential in kind, and some proponents add that the hearer acquires a right of complaint against the speaker if p proves false. On these views, reliance therefore requires a speaker who intentionally assures and stands behind the claim; for those who add the right of complaint, that speaker must also be answerable for failure (Leonard 2021, §3.2.1 and n. 14).

The four answers ask different things of an artificial configuration. On reductionism, those who rely need positive reason to think the configuration reliable for the use at hand, for which a performance record is the typical material. On assurance views, someone is needed who intentionally assures and stands behind the certification. On the minimal form of non-reductionism, it suffices that the hearer lack an undefeated defeater; but without routes by which the findings of internal checks reach those who rely, that condition holds alike whether or not the result is in error. On pure reliabilism, no human need stand at the end of the chain, but what is asked about is the reliability of the recipient's belief-forming process, including how the certification is taken up, and an inference that measures checkers by their own verdicts is a form of bootstrapping. Findings that do not reach those who rely, and circles of self-measurement, can arise in human organizations too. What is asked of an artificial configuration is how the checks placed inside it connect to the reasons and corrections of those who rely. Except on pure reliabilism, every answer involves a relation to those who rely or to those who stand behind the certification. Where the results relied on are results of research, the difference shows sharply. Since a certified result is new to S, those who rely often have no other source against which to check it. The performance record that is the typical form of the positive reason reductionism requires is then one built on tasks whose answers were already known, carried over to tasks whose answers are not (Section 6). On the non-reductionist answer too, the sources of defeaters that could reach those who rely may be largely confined to the findings of internal checks and to the research that follows. That is why whether there are routes by which findings reach them weighs so heavily.

5.3 Is the reliance trust, and is trust needed?

Section 5.2 asked what justifies reliance on results one has not checked. The question here is what kind of reliance it is: whether mere reliance on a dependable mechanism suffices, or whether it must be trust. The distinction is Annette Baier's. As Carolyn McLeod summarizes Baier's account in her encyclopedia entry on trust, trust is reliance on another's good will, as distinct from mere reliance (McLeod 2020, §§1, 1.1). In Baier's words, quoted there, "trusting can be betrayed, or at least let down, and not just disappointed," whereas mere reliance can only be disappointed (McLeod 2020, §1). On Baier's analysis, trust involves belief in the goodwill and competence of those trusted and a willingness to give them discretionary powers in matters of concern to oneself, letting them decide how one's welfare is best advanced, so that the one who trusts accepts vulnerability to a power to harm that is inseparable from the power to look after some aspect of one's good; and trust of this kind runs through institutions too, as when the welfare of a university is entrusted to trustees and administrators whose discretion no set of rules replaces (Baier 1992, 112–113, 117, 119). Wilholt carries the distinction into science, taking reliance to be working the assumption that another will do something into one's own plans (Wilholt 2013, §1).

The difference matters for an artificial configuration for the following reason. If trust is reliance on someone to whom good will or responsibility can be ascribed, then where reliance in science needs only reliance on reliability, an artificial configuration with reliable processes and records could supply it internally, whereas where it needs trust, such someone has to stand somewhere. Baier's institutional examples concern those who hold discretion within institutions, and Baier does not discuss artificial systems, invoking a well-programmed robot only as an image of the mechanical dependability that, on Baier's account, trustworthiness goes beyond (Baier 1992, 117). Carried over to an artificial configuration, the analysis would make whether it can be trusted rather than relied on turn on whether it can be said to have good or ill will toward those who depend on it in exercising its discretion. What follows has two parts. One asks whether any of the reasons for which reliance in human science has been held to need trust can be weakened by artificial configurations. To that end it takes up Hardwig's argument and the levels of its grounds, shows what records of execution change and what they leave in changed form, and considers Wilholt's ground, which records do not change. The other asks who the reliance that remains can be directed at, taking up the debate between Koskinen and Ortmann and accounts that do not restrict the objects of trust to those who can be responsible.

Chapter 9, Section 1 cited Hardwig as an example of others' entering the warrant or possession of knowledge; here Hardwig's argument is taken as an argument about whether that reliance requires trust in character. Hardwig argued that reliance in science is unavoidable, because the relevant data and arguments are too extensive and too difficult to be had except through testimony, and that it is reliance on character: because the one who relies lacks the testifier's reasons, relying on the testimony includes relying on the testifier's truthfulness, competence, conscientiousness, and capacity to assess the limits of their own knowledge and judgment (Hardwig 1991, 699–700, 706). Character, in this sense, is these qualities.

Against this one may object that prudence, the fear of losing if found out, is enough to keep researchers from deceiving, so that trust is not needed. Read against Baier's distinction, the objection takes reliance in science to be reliance on a motive compatible with ill will, the fear of being found out, and Hardwig's reply can be read as an argument that this is not enough and trust is needed. Hardwig argues that detection is weak, and sorts the grounds into two: peer review and replication may once have been effective deterrents but are now much weaker ones, and both have in-principle weaknesses (Hardwig 1991, 706). The grounds of the first kind concern how science, biomedical research in particular, was organized at the time: well-qualified referees are too few for the volume of submissions, and since replication is hard to fund and credit normally goes only to new findings, the structure of science works to prevent replication rather than to ensure it. Hardwig argues these points chiefly from the consensus of the biomedical community and conjectures that the conditions limiting replication increasingly apply to many sciences. The in-principle weaknesses are also two. One is that referees, who do not examine the original data or their gathering, cannot detect plausible and internally consistent fabrications. The other is that replication can confirm rather than expose fraudulent reports whose conclusions happen to be true. Hardwig further holds that, because only the researchers have the data, some lapses are not discoverable by anyone else (Hardwig 1991, 702–706). Which of these grounds depend on how science was then organized decides what an artificial configuration could change in that reply.

Lapses undiscoverable by anyone else need not involve fabrication. A 2017 paper on learning to evaluate dialogue responses, for example, left the investigation of its method's transfer across domains to future work. As one of its first authors disclosed two years later, the authors had already run a preliminary test, not mentioned in the paper, on the Twitter and Ubuntu evaluation data of an earlier paper, where the method showed no correlation at all with human judgments; that first author wrote that it "should have been a red flag" (Lowe et al. 2017, 1124; Lowe 2019). What the case shows is not a falsehood but an asymmetry of information about the research process: readers of the paper could not learn from it which tests had been run and what they had yielded. It became known because the paper's own author wrote it.

In terms of the sorting of Chapter 9, Section 6, most of Hardwig's grounds belong not to premises about human realizers but to premises about the present organization of science. The weakening of review and replication does. That only the researchers have the data, too, is less because humans do the research than because research has been organized so that records of data and process stay with the researchers. The in-principle weakness of review also depends on this organization, through review's being constituted so that referees do not examine the original data, and it could change if the handling of data were reorganized. The weakness of replication, by contrast, comes from replication's testing a report's conclusions rather than whether the reported data were obtained, and so, as long as replication is taken to test conclusions, it remains however the organization changes. This weakness comes from how the function of replication is defined, not from whether its realizers are human. On the view cited in Chapter 9, Section 2.3, the present social processes of science are one historically formed realization among possible processes, and these premises are not guaranteed for the future either. Arrangements that make records of data and process shared by design, and build checking into the organization of research itself, can be imagined within human research and have been attempted, as with the open notebooks of Chapter 5, Section 6.3. In Hardwig's argument, however, these premises do not produce the reliance on character itself. Since the one who relies lacks the testifier's reasons, relying on testimony includes relying on the testifier's truthfulness and competence, and this comes from the structure of testimony. What the premises support is that this reliance cannot be replaced by reliance on the fear of being found out, that is, on deterrence; from this Hardwig concludes that scientific knowledge rests not only on trust in deterrents but also on trust in the character of scientists. The part that comes from the structure of testimony is a premise neither about human realizers nor about the organization of science. It holds whatever testifies, so long as the one who relies lacks its reasons. Where certification by an artificial configuration is treated as the reading of an instrument (Section 5.1), the step that comes from the structure of testimony appears as reliance on those who designed and tested it. Hardwig also anticipated one reply: even with logbooks open to inspection, one would need to have been present during the experiment to know they are not themselves fraudulent, and even full policing would require relying on the integrity of the police; institutional reforms, on Hardwig's view, may diminish but cannot obviate the need for reliance on the character of testifiers (Hardwig 1991, 704, 707).

An artificial configuration, if so assembled, could realize this reorganization in a stronger form than human open notebooks, since a human notebook is a report a researcher writes, whereas in an artificial configuration the runs themselves can leave the record. This bears on Hardwig's reply that one would need to have been present to know a logbook is not fraudulent. If the scope of the runs is fixed, the runs within it are logged by infrastructure that neither the artificial configuration nor its operators can alter afterward, and those who rely can access the record and re-execute cheaply, the need to rely on the researchers' report of what was run and what came out can be reduced. The premise that only researchers hold the data and the record of the process weakens; but, as Chapter 9, Section 6 anticipated, the problem it gave rise to changes form rather than disappearing. That becomes visible when the public notebooks that Chapter 5, Section 6.3 considered as records for continuing inquiry are considered here as records through which others than the producer can check the research process.

The changed problem appears in three places. First, continued access to a record held by a third party depends on maintenance, preservation, or transfer to a successor. The supervisor of the public chemistry notebook described in Chapter 5, Section 6.3 counted the time-stamped version history of its hosted wiki among its advantages; after the hosting service closed, that history can be seen to the extent that a web archive happened to capture it (Bradley 2008; Wikispaces 2018; UsefulChem 2016). Whether the history survives depended on maintenance by someone who was neither the notebook's author nor its reader, and on chance preservation. Second, even where the infrastructure is maintained, what can be checked lies only within the record's boundary. Runs outside the recorded scope, and alternatives discarded before recording began, such as Lowe and colleagues' preliminary test, cannot be learned from the record. A record may also give the location of data without the data's being reachable. An electron-microscopy notebook was deposited by its author when the server hosting it was shut down. In it, the raw data of a shared session are listed by paths that, as the record itself notes, can be reached only on the institute's network or on the author's own. The author's account of the deposit calls the links to the author's personal network storage "now meaningless", and says that some datasets are too large for all to be public but that, at the time of writing, all can be provided privately on request (Cautaerts 2026). Reaching what the record points to comes back to asking the author. Third, the capacity to verify within the record is also limited: some kinds of checking become cheaper and some do not (Section 7), and checks that need new measurement or re-running an experiment do not become easy because the record is complete. And on Hardwig's argument of 1985 that capacity may be lacking in a layperson (Section 6). As for the part of the reliance on character that rested on the premise, what remains is reliance less on the character of an individual than on the maintenance of the record, its boundary, and the capacity to verify within it (below, the three supports of the record). These are conditions of reliance; whether dependence on those who sustain them requires trust in character differs across the accounts considered below.

Related arguments have already been made that increasing what can be disclosed or checked leaves dependence on others who sustain it. Besides Hardwig's own point about policing, Onora O'Neill argued that openness adds to trust only where those who sort and assess what is disclosed are themselves trusted (O'Neill 2002, Lecture 4); Ane Møller Gabrielsen, writing on data-intensive biology, argued that openness does not diminish the need for trust but calls for extending it to those who curate data after its production (Gabrielsen 2020, 497, 501–503); and Primavera De Filippi and colleagues argued that a blockchain, whose records its participants can verify, raises confidence in the system's operation while partly displacing trust to the developers and maintainers of the network (De Filippi, Mannan, and Reijers 2020, 6 n. 16, 7–8). Gabrielsen and De Filippi and colleagues distinguish confidence from trust after Luhmann, a distinction whose source differs from Baier's. These arguments do not share one concept of trust, but they bear on this section's question in separating the checkability of disclosed content from dependence on what sustains the disclosure. None of them discusses the certification of research results by artificial systems; De Filippi and colleagues discuss an automated system, but one that verifies the transactions recorded in a ledger. That trust shifts to those who maintain records and data, Gabrielsen and De Filippi and colleagues have already argued. What is attempted here is to apply the observation to the premise of Hardwig's argument, to keep apart, for an artificial configuration whose runs are logged, the record's boundary and the capacity to verify within it in addition to the record's maintenance, and to place these in the debate considered later.

Records can weaken the part of Hardwig's grounds that turns on lapses being undiscoverable. Wilholt locates the reason why reliance in science needs trust somewhere other than in detection. Section 4 took up Wilholt's account to ask whose valuations the weighting of errors answers to in judging a result adequate for a use. It returns here because the same argument distinguishes which reliance in science is trust, and that ground is not weakened by records. As Section 4 set out, on this account one who relies on a result also relies on its producer's having made the trade-offs between errors that conventions leave open in light of the valuations the one who relies expects. From this Wilholt distinguishes three kinds of reliance. The first is reliance on the assumption that a colleague's reports are reliable. Since how reliable is reliable enough cannot be settled without a valuation that compares the benefit of a correct result with the cost of an error, an analysis of reliance as reliance on this assumption is, on this account, incomplete. The second is reliance on conventional methodological standards. The third is reliance, in the leeway those standards leave, on others' working from the right value judgments. Wilholt calls only the third genuine epistemic trust, since it makes the one who relies dependent on another's attitude, and since the expectation is normatively charged, the one who relies can be let down and not merely disappointed (Wilholt 2013, §§2–5 and nn. 3, 20–21). This leeway can be read as the discretion Baier made central to trust. Wilholt notes, however, that this characterization does not exactly coincide with Baier's analysis in terms of good will, and that Baier's conception, built around one trusted and one trusting party, cannot be transferred unchanged to a collective enterprise (Wilholt 2013, n. 20). And Wilholt holds that this ground remains convincing even where there is an effective mechanism for detecting fraud (Section 4).

With Wilholt's distinction carried over to artificial configurations, a record of execution can show what was chosen in the leeway. If the record also shows the valuations adopted, and the expectations of those who rely are ascertained separately, it can serve as material for checking whether the two correspond. Recording alone, however, establishes neither that correspondence nor that the valuations were appropriate (Section 4).

Extending Wilholt's point about the first kind of reliance to a configuration's checkers, the judgment that a checker is reliable enough for the use at hand already weighs the costs of its errors. The question whether trust is needed therefore does not first arise after reliability has been established; it enters even the reliability assessment that a configuration seems able to supply internally. Douglas's point, seen in Section 4, that even stating uncertainty involves a judgment that weighs the consequences of error, points the same way. On Wilholt's account, what stabilized this layer among human researchers was a sense, shared by producers and those who rely, of what the right valuations are (Section 4). Where the producer is an artificial configuration, that sharing cannot be presupposed. Whether choices in the leeway match the expectations of those who rely must then be shown in some other way, or be supported by a relation in which those who rely can object (the stronger reading of Section 4). Whether trust in the third sense can be directed at the configuration itself turns on whether the configuration can be said to work from value judgments, and this returns to the question drawn above from Baier's analysis.

Applied to Hardwig's argument, an artificial configuration that records its executions under the conditions stated above can narrow the range of matters on which one must rely on the researchers' report alone, and so weaken part of the reason why reliance requires trust. Where one still relies on testimony without having the testifier's reasons, however, the reliance on character that comes from the structure of testimony remains. Nor does recording alone rebut the ground that Wilholt takes not to be weakened by the detection of fraud. And what is weakened remains, changed in form, as reliance on the three supports of the record.

Must the reliance that remains, then, be trust? Records open access to what was run; they need not make how a system works surveyable. Whether, under that second kind of opacity, what is depended on can still be linked to someone who can be trusted is what the following debate disputes. On Inkeri Koskinen's account, collaborative science has relied on trust in agents who can be responsible; Koskinen calls this view the necessary-trust view and takes it to be widely accepted in social epistemology. That trust extended to dependence on instruments: scientists who could not see into an instrument's working trusted those who built it and could take responsibility for its working and for the value-laden choices built into it. For a system that is essentially opaque in the sense of Chapter 3, Section 5, one whose internal processing cannot be seen into, no one grasps how it works well enough to take that responsibility, so what is depended on and who can be trusted come apart. Koskinen holds that essentially opaque AI applications are not agents who can be responsible and so can be relied on but are not appropriate candidates for trust, and that no available account yet closes the resulting gap (Koskinen 2024a). Carried over to artificial configurations, that an essentially opaque artificial configuration can be trusted is doubtful. That an artificial configuration can in fact be relied on is not in question; what Koskinen questions is whether such reliance, not mediated by anyone who can be trusted, is epistemically acceptable in science. Essential opacity, however, is relative to the nature of the subject (Chapter 3, Section 5). If there could be artificial subjects able to grasp the working of a system essentially opaque to humans, whether there is someone who can be responsible turns on whether such subjects can bear responsibility, a question Chapter 7, Section 3.3 left open.

Ortmann's account was touched on in Section 5.2, as an example that does not adopt the pure form of reliabilism, and in Chapter 8, Section 3.4, within the ordering of what passes to the relying side; here it is taken as an answer to whether trust is needed. Ortmann replies, from the case of AlphaFold2, that what epistemic dependence on an opaque system requires is not trust in the system but reliability established empirically in a context and transmitted through trusted investigators, so that what is trusted is not the tool but those who evaluated its reliability; Ortmann presents this as showing that dependence on AI poses no novel problem for the necessary-trust view, insofar as AlphaFold2 is representative of AI use in science (Ortmann 2025, 1, 3, 17–19). Both treat the opaque systems they discuss as objects of reliance rather than of trust, and both keep trust in persons; they divide over whom one must trust, and for what. On Koskinen's view, it has to be trust in agents who can be responsible for what is depended on, the link that opaque systems break; on Ortmann's, trust in the investigators who established the system's reliability and transmit it can support the dependence. Behind this difference lies the question of what one must be able to grasp in order to take responsibility for a system. Uwe Peters holds that taking responsibility for a system requires not understanding its internal processing but being able to design, monitor, and control it so that it is reliable and follows accepted procedures, and suggests that independent auditing bodies could check this (Peters 2024a, 60, 63–64); Koskinen replies, drawing on Wilholt's argument (Section 4 and earlier in this section), that reliability tests cannot show that value-relevant choices made inside an opaque classifier are acceptable (Koskinen 2024b), and Peters has replied in turn (Peters 2024b). What is at issue is not only whether the outputs are dependable but who can answer for the value-relevant choices behind them. Ortmann adds that if one can depend on an opaque system without trusting it, relying on moral agents without trusting them is at least logically possible, a question the paper leaves unresolved (Ortmann 2025, 18). It can be read as raising the question whether trust is needed for human researchers as well. If so, the question that separates the conditions that justify reliance from those that make it trust arises among human researchers too; the end of this subsection returns to that distinction.

Whether reliance on the three supports of the record must be trust divides with how trust is understood. On the accounts considered here that understand trust as reliance on someone who can be responsible, those of Hardwig, Koskinen, Ortmann, and Peters, unless an artificial configuration can itself bear responsibility, the possible objects of trust are those who maintain it, evaluate it, and are responsible for it. Mark Ryan reaches a similar conclusion for present, narrow AI: it can be relied on but not trusted, and trust goes to the organizations and individuals who develop and use it (Ryan 2020). On this book's application, those who maintain the record and those who verify within it become such objects, and the record's boundary bears on trust only through whoever drew it. On Koskinen's view, though, for an essentially opaque system no one can be responsible for its workings, so trust in them does not suffice to support the dependence. Accounts on which trust need not be directed at someone who can be responsible draw the line between what can be trusted and what can only be relied on elsewhere. For Andrea Ferrario and colleagues, trust at its simplest is relying without seeking further information to check the other's reliability (Ferrario, Loi, and Viganò 2020, 530–533); for Mona Simion and Christoph Kelp, trustworthiness is a disposition to meet obligations grounded in function (Simion and Kelp 2023, §§2–4). On either account, an artificial system can itself be trusted, or be trustworthy, in the sense they define.

Brought back to the question of Part V, how far the grounds of reliance can be met inside a configuration, the result is as follows. On accounts that require trust in someone who can be responsible, whether the artificial configuration as a whole can be that party is a different question from whether a member inside it, for instance a human included in it, can. Where the required party is outside the configuration, the relation to that party cannot be completed inside it; where a member inside can fill the role, there is room to place the relation inside. The disagreement between Koskinen and Ortmann and others remains, however, over who must be responsible for what if the dependence is to be supported. On accounts that do not so restrict the trusted, even reliance that requires trust could be met by directing it at the configuration itself. And on a view such as the pure reliabilism of Section 5.2, on which the reliability of the process justifies reliance, that justification needs no trust. The conditions that justify reliance and those that make it trust are, here, to be considered separately.

5.4 What internal agreement tells those who rely

The builders of Section 1 offered, as a reason for S to rely, that a result had passed internal checks, that is, that the configuration's internal checkers had accepted it. Sections 2 and 3 asked, from the side of criticism, what adding checkers or critics contributes. Here the same agreement is seen from the side of those who rely, and the question is what it tells them about this result. Since they do not examine the checkers themselves, the question extends beyond what the agreement actually adds to what reaches them about it, and to who takes the agreement up. The three are taken in turn. One is what the agreement adds as evidence, whose weight varies with the accounts of Section 5.2. The second is how verdicts were aggregated. The third is who takes up the agreement, which bears on the treatments of Section 5.1 and on assurance views. The checkers may be copies of one base system, built from different bases, or made to differ (Section 3), and what follows bears on each of these.

On the first: as Section 3 showed, on Goldman's condition (Section 2) agreement among sources that non-discriminatingly reflect a common source adds no evidence beyond that source's judgment, however many there are. Derivation from one base does not by itself make checkers sources that merely reflect one another, though. Two reports supply the step from this condition to what must reach those who rely. Frances and Matheson list, among the sources of evidence Goldman proposed for a novice judging between putative experts, agreement from additional putative experts (Section 6 below; Frances and Matheson 2024, §4); on Goldman's condition as Guerrero reports it (Section 2), that agreement is evidence only insofar as the agreeing sources do not merely reflect one another or a shared third party. And Dietrich and Spiekermann observe that judgments can be causally independent yet correlated through common causes such as shared evidence, a shared methodology, or shared hypotheses (Dietrich and Spiekermann 2021, §2.2). Taken together, for one who relies on internal agreement without examining the checkers, how much the agreement adds beyond a common source depends on the checkers' common causes, and so on how they were made and from what; information about this, their composition and lineage, becomes part of what the one who relies needs to assess that. The weight of this agreement varies with the account of Section 5.2. Applied to reductionism, for checkers built from a common base the question is whether, distinguishing the track record they share from the track record each gained after they diverged, and taking their dependencies into account, one can judge their reliability for the use at hand. Applied to pure reliabilism, the question is whether the process, including aggregation and uptake, is in fact reliable, and the one who relies need not have reasons to that effect.

On the second: where certification is settled by aggregating internal verdicts, for instance by majority, number can work, as Section 3 noted, as influence rather than as evidence. From the side of those who rely, assessing the grounds of a verdict reported as "accepted by many" requires knowing not only how many agreed but which judgments received which weights, and who set the proportion of copies. How verdicts were aggregated also belongs to this information.

On the third: showing how much evidence was added does not by itself settle whose assertion the certification is or who gives an assurance. Applying assurance views, what is asked is not only who built the checkers and aggregated their verdicts but who assured the recipient of the content and invited their trust. One party may fill both roles, but having provided a base system or deployed a configuration does not by itself amount to giving an assurance. Applying Goldberg's reported account of instruments, dependence on checkers built from a common base includes dependence on the epistemic work behind that design; it does not by itself establish who can answer for every verdict, especially where the system is essentially opaque (Sections 5.1 and 5.3). Applying the definitions of Lackey's account reported by Marsili, a report that an automated system produces by aggregating inputs, and that no one reasonably intends to convey, is not the group's assertion (Section 5.1). A report that automatically aggregates checkers' verdicts is close to that case; on Lackey's account it is a group assertion only if an authorized spokesperson with the requisite communicative intention takes it up, or members coordinate individual acts so that they all reasonably intend to convey its content together.

That different origins need not mean different routes of error was shown by the finding of Kim and colleagues cited in Section 2. There are also findings that bear on aggregation. In a preprint, Song and colleagues evaluated language models on questions drawn from research projects in four sciences. The top models from four providers often erred on the same set of hardest questions, and on one task all four made the same mistake on four of 22 questions. Accuracy rose and fell together across research scenarios, with all pairwise correlations above 0.8 in chemistry and physics. They infer that on inherently difficult scenarios and questions, naive ensembles such as majority voting across providers may deliver limited improvement, and they suggest similar pretraining data and objectives as a plausible explanation (Song et al. 2026, 7–8). It is the shared errors that bear on Goldman's condition. Accuracy rising and falling together can arise from the shared difficulty of scenarios alone, whereas what Goldman asks is whether sources agree even when the result is false. For the proposal of hypotheses, Chapter 10, Section 4 cites both a finding that hypotheses from models of different providers were almost as similar as repeated samples from one model and a finding of no evidence that models from the same company generated more similar ideas. These are findings about particular models and tasks at one time, not about artificial checkers in general, and high error correlation does not mean that adding judgments adds no information at all. Still, for the models studied at least, they suggest that a difference of provider alone is not a reason to count agreement as several pieces of evidence or to regard majority voting as robust. How many pieces of evidence a given agreement may count as is decided by evidence about the checkers in question, including findings like these.

Thus, for a certification that offers internal agreement as a reason for reliance, three judgments come apart: how much evidence was added, which judgments determined the aggregate, and who takes up the report. Establishing one does not settle the other two. And where those who rely are to weigh the agreement as their own reason, as on reductionism or on non-reductionism that requires sensitivity to defeaters, the information they need for that extends to the checkers' composition and lineage and to the way their verdicts were aggregated. On pure reliabilism, the justification of those who rely does not need this information; it is needed where those who rely, or anyone else, set out to ascertain and show that the justification holds.

5.5 What remains

Section 5 asked what reliance on certification by an artificial configuration requires, and which parts of it the configuration can supply internally. The answers differ by account for each of the three questions.

As to what kind of act certification is: treated as an assertion of the configuration itself, or of a group for which it speaks, it falls under norms of assertion. Under Williamson's knowledge norm, a report resting on a calibrated score alone is warranted only as a report that the result is probable, and whether an output is offered as a conjecture or as an assertion becomes part of what is checked. Which treatment applies is not settled merely by the configuration's being artificial. On Fricker's account present AI outputs are not assertions. On Lackey's account, an authority-based group assertion requires an authorized spokesperson with the requisite communicative intention; a collective assertion instead requires members to coordinate individual acts so that they all reasonably intend to convey the content together. Treated as an instrument reading, applying Goldberg's reported account directs reliance toward the epistemic work of its designers and makers; whether trust in those designers and makers can support reliance on essentially opaque systems remains disputed (Sections 5.1 and 5.3).

As to what grounds reliance: what a configuration can supply internally is a reliable process and a record of its performance. On pure reliabilism, the reliability of the recipient's whole belief-forming process can suffice for justification; the configuration, however, supplies only part of that process, and an inference that measures checkers by their own verdicts is subject to the bootstrapping objection (Section 5.2). On the other accounts, requirements remain that the configuration's process and record do not by themselves meet: on reductionism, that those who rely have positive reason to think the configuration reliable for the use at hand; on assurance views, someone who intentionally assures and stands behind the certification. Reliabilism that does not take the pure form requires that the sources conferring reliability be shown reliable in turn, and on Ortmann's account trust in those who assessed that reliability enters there. The minimal form of non-reductionism does not require routes, but without routes by which findings reach those who rely, its condition holds alike whether or not the result is in error. Applying Alvarado, whether the non-reductionist answer applies at all depends on whether the certification is expert testimony or the reading of an instrument (Section 5.2). Where internal agreement is offered as the reason, then on accounts on which those who rely weigh the agreement as their own reason, the checkers' composition and lineage and the way their verdicts were aggregated also have to reach them; who takes up the agreement is settled separately from how large it is (Section 5.4).

As to whether the reliance must be trust: a configuration that records its executions under the conditions stated in Section 5.3 can narrow the range of matters on which one must rely on the researchers' report alone, and weaken the part of Hardwig's grounds that turns on lapses being undiscoverable. Where one still relies without having the testifier's reasons, however, the reliance on character that comes from the structure of testimony remains. As for the valuations made in the leeway conventions leave, recording alone does not establish that the choice matched the valuations expected by those who rely. What is weakened, too, remains as reliance on the three supports of the record (Section 5.3). Whether reliance on these must be trust divides by account of trust. On accounts that take the one trusted to be someone who can be responsible, the question is whether the configuration as a whole, or someone it includes as a member, can be such; where neither can, the one trusted stands outside the configuration. On Koskinen's view, no one can be responsible for the workings of an essentially opaque system, but that opacity is relative to the subject (Chapter 3, Section 5), and whether an artificial subject can be responsible is the question Chapter 7, Section 3.3 left open. On accounts that do not so restrict the trusted, the configuration itself can also be trusted. And on a view on which the reliability of the process suffices for justification, justification needs no trust. The conditions that justify reliance and those that make it trust are to be considered separately (Section 5.3).

Taken together, how much of what reliance requires a configuration can meet internally depends, given how certification can be treated under Section 5.1, on how the answers to the second and third questions are combined. On pure reliabilism trust is not added to the conditions of justification, and on accounts on which the one trusted need not be able to be responsible, the configuration itself can be trusted. Neither, however, shows by itself that the whole relation is completed inside. Judging that a certification is enough for the use at hand involves the valuations of Section 4, so if Wilholt's or John's account is adopted as well, a relation remains over whose valuations the weighting of errors answers to. On Wilholt's account this is a correspondence with the valuations expected by those who rely; but, as Section 4 set out, it does not require meeting every expectation. Nor does it by itself settle the relation to those who bear the costs of error without relying on the result. On other combinations, part of the relation to those who rely, or to those who stand behind the certification, is required. Whether that relation can be placed inside the configuration depends on where they stand, inside or outside it. Which treatment applies changes the shape of research as well. If certification is treated as the reading of an instrument, then on Fricker's account what S gains is knowledge from an instrument rather than from testimony, while applying Goldberg's reported account directs reliance toward the epistemic work of its designers and makers. Whether they can take informed responsibility for an essentially opaque system remains disputed, as Koskinen argues (Section 5.3). In human research, results reached S as researchers' assertions, and S could question the researchers. Once certification moves inside, that form can give way to one in which S reads the values of an instrument and relies on its makers, and those S can then question need not be anyone who knows what was chosen within the particular inquiry.

In terms of the sorting of Chapter 9, Section 6, what records were found able to weaken is the part supported by a premise about the present organization of science, that records of data and process stay with the researchers. Reliance that comes from the structure of testimony holds whatever does the research, so long as the one who relies lacks the testifier's reasons. The relation over valuations made in the leeway belongs to relations with outside parties, so long as those who rely, or who bear the costs of error, are outside the configuration.

Many of these answers were worked out with human members in mind as those who rely. Where S includes artificial members, the answers can change in at least three respects. One is the shared sense of valuations: where the one who relies is an artificial member, whether it shares that sense with the producer is asked in a different way from the case between humans (Section 5.3). The second is the assessment of information about the checkers' lineage: where members share a base with the configuration, that assessment may itself be correlated with the configuration's capacities (Section 5.4). The third is the acquisition of understanding, which need not pass through testimony (Section 5.6). The case in which S as a whole stands as a layperson to artificial configurations is taken up in Section 6. Who may set the standards and membership of certification, and who may contest and revise those decisions, remains a question of authority beyond the grounds of reliance, taken up in Chapter 12.

5.6 Certification and understanding

The argument so far has asked whether, by relying on a certification, those who rely may know or use a result. But the achievement research aims at is not limited to knowledge; it can be understanding (Chapter 3, Section 1). And as Chapter 3, Section 8 set out, on some accounts knowledge passes by testimony while understanding why does not pass so easily. On such accounts, one who has come to know by relying on a certification has not thereby come to understand. Where understanding is sought, then, what a certification established has to be divided more finely than for knowledge.

Chapter 3, Section 6 distinguished three claims about who understands: that AI helps others understand, that AI itself understands, and that a collective understands. What certification bears on is three claims divided differently: that an explanation or model is adequate for a stated purpose; that a specified person, artificial actor, or group understands the target; and that a recipient has acquired understanding through what was communicated. Chapter 3, Section 6 mainly distinguished the bearers of understanding; here the adequacy of an explanation and acquisition through what was communicated are added, and the claims are divided again. Evidence for one claim can leave the others unsettled. Certification that an explanation is adequate gives a recipient reason to accept it without establishing that the recipient has come to understand the target. The three are not disjoint kinds of achievement, though, since a recipient who acquires understanding is also a bearer.

How far certification bears on acquiring understanding divides with accounts of testimony and understanding (Chapter 3, Sections 7.4 and 8). On Hills's distinction, transmission, taking the speaker's word as a reason for belief, yields second-hand knowledge, and understanding can pass through propagation, where what is said prompts and frames the recipient's own inquiry. Boyd and Grimm allow a suitably prepared recipient to gain understanding from testimony, and Hu allows that knowing that one thing caused another brings rudimentary understanding. On any of these accounts, certifying an explanation's adequacy and communicating that explanation are different acts. A bare endorsement need not communicate the content relevant to understanding; a certification that also explains may. And a claim of acquisition needs, in addition to evidence that the recipient understands, evidence that this came about through what was communicated. For example, on an account that makes being able to follow and use dependencies a requirement of understanding, checking a model's adequacy as an explanation and checking whether a recipient can use those dependencies under relevant variations are different checks, and neither alone shows that what was communicated produced that ability in the recipient.

The above concerns recipients who learn from what others say. Where S's members are artificial and can copy or share representations or abilities directly, acquisition need not pass through testimony. Whether the recipient then understands depends on the account adopted. An ability account asks whether the recipient has come to possess the relevant abilities. An account that treats understanding as the understander's own cognitive achievement asks, further, whether the copied state counts as that recipient's own achievement (Chapter 3, Section 8).

In either case, even if the work of assessing understanding can be moved inside a configuration, whose understanding the assessment supports has to be specified separately. Even if the configuration checks an explanation's adequacy and is certified as itself understanding, the intended recipient's understanding is not established unless it is supported as a separate claim. Conversely, that the configuration has enabled a recipient's understanding does not show that it itself understands. This extension rests on the arguments of Chapter 3; it does not follow from the accounts of transmitting knowledge and relying on others considered in Sections 5.1–5.3. Put from the side of research, there could be a science in which artificial configurations certify the adequacy of explanation after explanation, and certified explanations accumulate in the record of research, while no member of S understands the phenomena they explain. If understanding is among the achievements research aims at (Chapter 3, Section 1), an increase in certified explanations does not by itself show that this achievement is increasing in S.

6. When S becomes a layperson

Section 5 considered members of S relying on the certification of an artificial configuration that they had not checked themselves. Even with such reliance, where others somewhere in S can check a field's results on behalf of members who cannot, being unable to check is confined to that field. So it largely was in human research: specialists too are laypersons outside their own fields, but specialists in a field could check its results. This section considers the case in which that no longer holds. If, in every field at once, artificial configurations came to gather, read, and assess more of the relevant evidence than any member of S, and no one in S could any longer check their results, S would stand to them as a layperson stands to an expert, not member by member but as a whole. The section considers this as a thought experiment, not as a prediction that it will come about. Whether such a limit arises, and in what form, depends on whether S consists of humans, of artificial members, or of both.

Why consider this limit? Because it is where a question becomes sharp: if artificial configurations do research better than anyone in S, why should S not simply defer to them? Bostrom's principle of epistemic deference, considered below, says just that. To see what S could lose there, two things that bear on deferring to others are kept apart. One is following others' judgment on what is correct; below, this is called deference. Every specialist already defers in this sense to specialists outside their own field. The other is the standing to take part in deciding what inquiry is for, for instance which unknowns to pursue and which risks to bear, what Chapter 9, Section 2.1 called standing. In human research there was little need to consider the two separately, since laypersons who deferred to experts remained members of the society that decides on ends. If S as a whole became a layperson in every field, this would no longer be self-evident. When S defers to artificial configurations on everything, the question is whether standing moves with it. In research the question becomes sharper. S's preferences about what is worth researching are themselves shaped by accounts of how significant possible inquiries would be, and those accounts are given by those who inquire (Chapter 12, Section 4). Research can also challenge S's present values and widen the space of later choice (John's argument, taken up in Chapter 12, Section 3.3). Those deferred to on content are therefore apt to stand also where they educate the preferences with which S decides on ends, and keeping the two apart can be harder in research than in other settings of advice.

The answer divides with the account of legitimate authority adopted (Chapter 12, Section 1). The conceptions that Fabienne Peter reports have dominated the literature ground legitimacy in the will of those subject to authority, whether that will is expressed through consent, shared reasons, or political participation (Peter 2023, §3.3). On these conceptions, standing is a relation to those subject to authority and does not arise from competence, so deference does not transfer standing. On conceptions that ground legitimacy in accord with normative facts, in a decision's being normatively correct, or in some epistemic advantage, by contrast, deference to competence can bear on standing too. The section examines whether the two can remain apart when deference is pushed to its limit.

How such a limit arises depends on what S consists of. Hardwig's argument, set out below, concerns human laypersons, and on it being a layperson arises from the finitude of those who defer. Where S consists of artificial members, it may not arise in the same way, since their capacity need not be bounded as a human life is. Or it may arise in a sharper form: if members that share base systems with the configurations assess lineage and independence with capacities correlated with theirs, they lack an outside position, a judgment not correlated with the configurations' capacities, from which to ask the questions this section examines. Members made differently from the configurations may keep that position; which form, if any, an artificial S's lay standing takes depends on how its members are made as well as on their capacity. Where S consists of both human and artificial members, S as a whole need not become a layperson. If artificial members can check the configurations' results, the line between those who can check and those who cannot moves inside S: human members stand as laypersons not only to the configurations but also to the artificial members who check on their behalf, and the question becomes one of reliance among the members of S. If those artificial members share a base with the configurations, the sharper form above also bears on it. What becomes of standing when S is artificial is taken up in Chapter 12, Section 5. The accounts of Hardwig, Anderson, and Irzik and Kurtulmus considered below, and the dispute between Lippmann and Dewey, concern human laypersons and a human public, and apply where S includes human members.

Nor does S become a layperson only through growth in the capacities of artificial configurations. It can also happen through a decline in those of its members. For example, as the evaluation of research is left to artificial configurations, members may cease to gain experience of checking results themselves, so that the capacity to check is neither maintained nor passed on to new members (Chapter 3, Section 8). Then, even if nothing changed on the side of the configurations, the gap between them would open from the side of S. The route by which capacity is kept up and passed on through experience, though, is one for human members. For artificial members, whether capacity is kept up depends on how they are remade and updated; if new members are made from or on the configurations' outputs, for instance, the case approaches the sharper form above.

The question whether a whole public becomes a layperson before experts has a precedent in a dispute of the 1920s about democracy. As Matthew Festenstein reports it, Walter Lippmann held that the complexity and opacity of modern societies require setting aside the idea that citizens are "omnicompetent" and accepting the legitimacy of decision-making by qualified experts, while Dewey replied that experts' becoming a separate class distances them from the needs their expertise is meant to serve (Festenstein 2023, §3).

Hardwig's argument began from finitude. There is too much evidence, much of it available only after specialized training, and life is short; a layperson who defers to an expert is therefore rational, and experts so often take the layperson's role within their own field (Hardwig 1985, 335, 339–340, 345–346). The layperson can raise objections and offer alternatives, but final judgment on what is correct lies with the expert. Whether someone is an expert at all, Hardwig held, a layperson can settle alone only for the more obvious quacks. What the layperson retains are reasons of another kind, which Hardwig calls ad hominems: that the expert has an interest in the outcome, is not acting in good faith, or is covering for peers or yielding to social pressure from others in the field (Hardwig 1985, 340–343). Hardwig ended with a dilemma. Either one can know vicariously, without possessing the evidence and perhaps without fully understanding it, or knowledge belongs to communities rather than individuals; and the second option threatens the autonomy and responsibility of individuals, equality, and democracy (Hardwig 1985, 348–349).

Carried over to configurations, Hardwig's structure changes in scope. Whether deference to configurations would be rational depends on the positions of Section 5. Where it is, final judgment on questions of content can lie outside any member of S. Hardwig does not discuss artificial systems, but on that account, carried over, this would not be a defect of research by artificial systems; it is the position every specialist already occupies with respect to other fields, provided configurations stand to S as experts in Hardwig's sense. If their certifications are instead readings of an instrument (Section 5.1), the position of S is rather that of someone relying on an instrument and on those who stand behind it. What changes is that the layperson is no longer an individual within a community of experts but S in relation to what it depends on, and the threat with which Hardwig closed the dilemma, that making knowledge belong to the community endangers individual autonomy and responsibility, equality, and democracy, here concerns that whole relation between S and what it depends on. Nick Bostrom states a position at one end of the range of views on how far deference may go, for the case of a superintelligence, as a heuristic that Bostrom calls the principle of epistemic deference: a future superintelligence occupies an epistemically superior vantage point, its beliefs being "(probably, on most topics) more likely than ours to be true", and "we should therefore defer to the superintelligence's opinion whenever feasible". A note qualifies it: deference is not owed on topics where we have good reason to think our own beliefs more accurate, such as what we are thinking at a given moment, unless the superintelligence has access to our opinions, in which case judging when they should be trusted could itself be deferred to it, with special cases involving indexical information handled, for example, by having it explain what it would be rational to believe from our perspective (Bostrom 2014a, ch. 13 and n. 8). Bostrom states the principle generally and applies it to the choice of values, not to research. On it, whether deference is rational is settled, the topics on which we know better excepted, by the general epistemic superiority of the one deferred to, rather than by assessments of the kinds Section 5 distinguished, of particular routes to error or of who stands behind a result; and since, where it has access to S's opinions, the exception can itself be deferred, final judgment can lie outside S on the question of when S's own judgment is to be trusted as well.

A research program in AI alignment aims at the other end of this range, at keeping those who defer able to judge content. Samuel Bowman and coauthors call it scalable oversight, the problem of overseeing systems that may be more capable than their supervisors in most of the skills a task requires (Bowman et al. 2022, 1). One proposed technique is debate, in which two systems argue for different answers before a human judge. Geoffrey Irving and Amanda Askell state its hypothesis as the claim that optimal play in this game, giving the argument most convincing to a human, results in true, useful answers (Irving and Askell 2019, "Debate: learning human reasoning"). In experiments with reading-comprehension questions about passages the judges had not read, debate between an expert arguing for the correct answer and one arguing for an incorrect answer raised the accuracy of non-expert judges above a baseline in which a single expert argues for one answer, when the debaters were human (Michael et al. 2023, 1); and debate between language models raised the accuracy of weaker models and of humans as judges above naive baselines (Khan et al. 2024, 1). The authors mark some limits themselves: the gap between expert and judge was one of access to a text, and the results depended on the debaters' being able to quote verified evidence, so how far they carry to questions that draw on large bodies of scientific knowledge is open (Michael et al. 2023, 11; Khan et al. 2024, 8). The questions, moreover, had known answers against which accuracy was scored.

What bears on this section is how the program establishes that a procedure works. In the sandwiching design that Bowman and coauthors adopt from a proposal by Ajeya Cotra, non-experts try to obtain reliable performance on a task from a system more capable than they are, and experts, who take part only at the end, evaluate whether they succeeded. The aim is a technique that would give enough justified confidence for the experts to be no longer needed (Bowman et al. 2022, 2–4). Irving and Askell state the difficulty directly. Debate is meant for questions with no other source of truth, but without such a source "there is no way to measure whether debate produced the correct answer," so experiments begin in domains where the experimenters know the answer and the judge does not (Irving and Askell 2019, "From superforecasters to superjudges"). The program thus meets the limit set out in Chapter 8, Section 3.4 at one remove. A procedure meant to carry a layperson's assessment beyond what the layperson can check is itself calibrated on cases that someone can check, and its use beyond them rests on that calibration's carrying over. Applied to S as a layperson, such procedures could let its members judge content further than Hardwig's layperson can. What grounds their reliance on the procedure, however, is again a record on claims that can be checked, and such validation needs those who can give the answers independently, whose competence is among the capacities that Chapter 3, Section 8 counts as capable of being lost.

Whether S could then know through its configurations depends on the account of social knowledge, which Chapter 2, Section 6 left open. Lackey holds, as Jeff Dunn reports in a review, that one can grant another the authority to speak on one's behalf, whereas one cannot grant another the authority to believe or know on one's behalf (Dunn 2021). On Lackey's account, a society that stands as a layperson could let configurations speak for it and would answer for what they assert on its behalf, while its knowledge would still require that a significant percentage of its operative members, the members with decision-making authority in the relevant domain, hold the relevant justified beliefs. On distributed and group-functional accounts (Table 2.2), on which a properly organized collective can realize the cognitive functions tied to knowing, S could know through a properly organized system that includes its configurations. On accounts of knowledge without a knowing subject, a publicly certified result of reliable processes can have the status of scientific knowledge even if neither S nor any of its members is its knower; on Giere's version of such an account, individuals can claim to know such a result because they know that it was certified by reliable, accepted procedures (Chapter 2, Section 6). What the thought experiment shows is that the answer to Hardwig's dilemma for S as a whole inherits the choice.

Two further questions concern what remains to such a society, and they are of different kinds. The first is the layperson's reasons of the other kind, transposed. A human layperson asks whether an expert has an interest in the outcome, acts in good faith, or covers for peers. With respect to a configuration, corresponding questions can be asked about its relations, whether or not character can be ascribed to it: who, if anyone, owns and directs it, which objective it serves, whether the evaluators that certified its results share its lineage, and whether its record reaches the runs that matter. These correspond to sources of evidence Goldman proposed for a novice choosing among experts, which, as Frances and Matheson list them in their encyclopedia entry on disagreement, include the experts' interests and biases and agreement from further putative experts (Frances and Matheson 2024, §4). On the accounts of Hardwig, Goldman, and Anderson below, they can be asked without competence in the content, although answering some of them, such as whether evaluators share a lineage, can require technical expertise of another kind. The second is standing over ends: whether a society that defers on what is true keeps the standing to decide which unknowns are pursued on its behalf and which risks it will bear. The accounts of social knowledge and of deference examined in this section do not settle that question, because they concern what can be known and believed through others, not who may set ends for others. As noted at the start of this section, whether such standing remains divides with accounts of legitimate authority. On conceptions that ground legitimacy in epistemic advantage, competence can bear on standing; but except on accounts that adopt in pure form the view that those who know more may rightfully rule (epistocracy, rule by the knowledgeable), on which competence alone yields title to rule, the competence of artificial configurations would not by itself transfer standing to them (Chapter 12, Section 1). Where it remains, whether it is more than nominal when the options a society considers are generated and ranked by the configurations it depends on is the question of Chapter 12, Section 4.

Elizabeth Anderson gives criteria for the first and ties it to democratic legitimacy. Anderson argues that policy resting on scientific reasoning that citizens cannot assess directly keeps its democratic credentials so long as citizens can judge whom to trust and whether the trustworthy agree (Anderson 2011, abstract). On this account, as Gürol Irzik and Faik Kurtulmus set it out, laypersons can judge expertise by where a scientist stands in a hierarchy that runs from degrees in the field to recognition by other scientists and leadership marked by election to national academies and prizes, honesty by signs such as conflicts of interest and previous misconduct, and epistemic responsibility by signs such as publication in peer-reviewed journals, refusal to share data without good reason, or failure to engage with refutations of one's own claims (Irzik and Kurtulmus 2019, §3). Anderson's account rests on two conditions. One is access: the criteria are easy to apply, she holds, for anyone of ordinary education with access to the Web (Anderson 2011, abstract). The other is disposition: as George Kwasi Barimah reads her, Anderson holds that laypersons could readily make such assessments but for factors that make them ill-disposed to do so, which in a passage he quotes she names as biased and misleading media reports, the segregation of people with different opinions, and cultural cognition, the tendency to judge factual claims by their fit with one's values (Barimah 2024, 464, 468). Barimah doubts the first condition as well. Some of the criteria, such as judging whether the bulk of a field's work expresses a common opinion, call in his view for skills of academic research; and in the suspended Ebola vaccine trials in Ghana that he examines, where two expert bodies disagreed, the information needed was not readily available on the Internet, and the expertise of individual scientists was subsumed under the organizations that issued the statements (Barimah 2024, 467, 471).

Anderson's account can be read as supporting the claim that the relational questions can be asked without expertise in the content, and as adding one: whether a configuration's results, or those who stand behind them, answer criticism rather than repeat what has been refuted. It also qualifies the claim. Its criterion of consensus, in a passage Barimah quotes, counts on the convergence of the vast majority of diverse inquirers, which is what makes a consensus robust; laypersons are to read it in surveys, reviews, or meta-analyses of the peer-reviewed literature, in surveys of trustworthy experts, and in consensus statements and reports of leaders in the field (Barimah 2024, 466). Where the agreeing experts are artificial configurations, the signal a layperson reads in such reviews, surveys, or consensus statements does not by itself convey how diverse the converging inquirers were. For human experts much of that difference came from separate persons trained in separate places with access to separate evidence; among configurations it depends on how they were made. Configurations that share a lineage can leave the signal intact while the diversity that made it robust is gone, and configurations built from different bases do not guarantee it (Section 5.4). Applied to configurations, then, the criterion adds to what the layperson needs to read information about how the agreeing configurations were made and how they differ; this is the internal agreement of Section 5.4 seen from outside the field. Barimah, who does not discuss artificial configurations, proposes for human experts that, because laypersons are often not equal to the task, experts bear the greater share of the burden by supplying information about their expertise, their track record, and whether a consensus exists (Barimah 2024, 472–473). Carried over, the information about how agreeing configurations were made and how they differ would fall to those who build and deploy them to supply, and Barimah's Ghana case suggests why: where statements are owned by organizations rather than by identifiable scientists, criteria built for individual experts are hard to apply, and behind a configuration's certification there may be no individual of the kind the criteria presuppose. And since, on Anderson's account, the capacity to judge whom to trust bears on the democratic legitimacy of policy that rests on deference, its erosion would bear on legitimacy as well (Chapter 12, Section 7).

Irzik and Kurtulmus, analyzing when a public's epistemic trust in scientists is warranted, mark a limit of such criteria. They distinguish two kinds of reason a public can have for judging whether research is reliable: first-order reasons, concerning the evidence for the claim itself, and second-order reasons, concerning the scientists themselves, which the public uses instead when it cannot assess the first; Anderson's criteria above, expertise, honesty, epistemic responsibility, and a consensus of trustworthy experts, are of the second kind. For second-order reasons they follow Goldman and Anderson (Irzik and Kurtulmus 2019, §3). And, building on Wilholt, they add an enhanced form of trust that requires in addition that, where public welfare is at stake, scientists' decisions about the distribution of inductive risk agree with the public's assessments and that the public have reason to believe they do. The second-order reasons, they observe, serve as proxies for whether research is reliable, not for whether its inductive risks are distributed as the public expects; for that, they point to hybrid forums in which citizens and scientists deliberate, where citizens bring in their expectations and check whether they are met, and to diversity among researchers (Irzik and Kurtulmus 2019, §4). From the controversy over the MMR vaccine they argue that enhanced trust can be lacking without fault on either side when the parties' values diverge (Irzik and Kurtulmus 2019, §5.3). This book reads their analysis as carrying the value layer of Section 4 to S as a layperson. The relational questions bear on reliability and honesty; whether an artificial configuration's weighting of errors matches what S expects needs channels through which S's evaluations reach it. Their two mechanisms supply such channels in different ways. The diversity they mean is diversity of deeply held values and of social position, such as gender, ethnicity, and class. They note that it can limit prejudice and so bear on reliability, but here they use it for a different purpose, as a reason to think research reflects the public's expectations about inductive risk; in that use it differs from the independence as difference in routes of error seen in Sections 2 and 3. The more diverse the scientists, they hold, the likelier that research reflects the expectations of the whole public rather than of a segment. They acknowledge, though, that even a sufficiently diverse community may hold values different from the public's because of its training and public responsibilities (Irzik and Kurtulmus 2019, §4). Applied to artificial configurations, the range of valuations can be made the same or made to differ, and so far as it is fixed by construction, it comes not from the lives of S's members but from the settings of those who build and assemble the configuration. A reason to think the range reflects S's range then becomes a reason about how it was set and against what it was matched, and that requires a channel through which S's expectations enter. In human communities the two mechanisms could work separately; for artificial configurations, whether diversity counts as a reason at all comes to depend on a channel of the kind a hybrid forum provides. Applied here, such a forum is a relation between those who set the weighting of errors and the members of S affected by its consequences; whether that relation lies outside the artificial configuration depends on where S's members stand, inside or outside it. Where the members are artificial and stand inside, the further question is whether their valuations are distinct from the settings of those who build and assemble the configuration. In research, one more thing is added to this channel. For results new to S, S's expectations about the weight of errors may not yet have formed, and once formed they can move through research (Section 8). The expectations a forum brings in are therefore at once a standard against which the weighting is checked and something research goes on revising, so the checking is not done once and for all.

Three things come into view from this thought experiment. One is that deference on what is correct, pushed to its limit, puts S, on Hardwig's account carried over, only where every specialist already stands outside their own field: final judgment on questions of content then lies in the hands of no member of S, which extends that position. Whether S thereby comes to know, however, inherits the choice, left open in Chapter 2, Section 6, among Lackey's account, distributed and group-functional accounts, and accounts of knowledge without a knowing subject; on Lackey's account, on which S's knowledge still requires justified beliefs among its operative members, deference in the limit can put S's knowledge at risk. The second is that what remains with S is not competence about content but the capacity to ask about its relations to what it depends on. On the accounts of Hardwig, Goldman, and Anderson, such questions can be asked without competence about content. Where the agreeing experts are artificial configurations, however, the signal of consensus that S reads does not by itself show how they differ, and among configurations that share a lineage it can persist after the diversity that made it reliable is gone; on Irzik and Kurtulmus's analysis these questions do not reach the weighting of errors, for which a channel bringing in S's valuations is needed; and procedures that extend a layperson's judgment rely on calibration in cases that can be checked, and need those who can produce answers independently. The third is that, except on accounts that adopt epistocracy in pure form, standing over ends does not pass with deference, but it can become nominal. Even if competence alone does not give rise to standing, whether standing stays more than nominal depends on whether S keeps the capacity to ask about those relations and to bring in its valuations, and that capacity can decline, with nothing changing on the side of the configurations, as members cease to gain experience of checking, or are not remade in ways that keep it. Carried over, Anderson's condition of disposition adds that capacity does not secure the disposition to use it: where certifications reach S as the readings of an instrument (Section 5.5) and no stage makes evaluation visible (Section 7), the occasions on which the relational questions would be asked can thin out even for members who could ask them.

Understanding adds a question that neither warranted deference about adequacy nor standing over ends answers by itself. A community may have reason to accept that an artificial expert's explanation is adequate while its members do not understand the phenomenon. If human understanding is an intended achievement, relying on a verdict of adequacy does not show that it has been reached. Whether testimony and background abilities can supply the missing achievement follows the alternatives in Chapter 3 and Section 5.6; the layperson analogy itself does not decide it.

Within the chapter, this section removes one premise that Sections 1–5 had left implicit. Those sections asked which functions of criticism and certification could move inside artificial configurations, presupposing that somewhere in S, besides those who rely, there are members who can check. This section considered the limit at which that premise fails. What came into view is that reliance on content can move inside, while the capacity to ask about relations and channels for bringing in valuations do not move with it, nor, except on accounts that adopt epistocracy in pure form, does standing over ends. This matters because here the book's question, what changes when A is a machine, takes the form: if machines do research better, why not simply defer? On this section's answer, where S includes human members, what can be lost in the first instance is not access to truth but substantive participation in deciding ends and the capacities that support it; on some accounts, as will be seen below, that loss leads to an epistemic loss as well. In what sense this loss is a problem divides by account, and at least four senses can be distinguished. One is a lack of legitimacy: on will-based conceptions of legitimacy, decisions about which unknowns are pursued and which risks are borne would be made without passing through the will of those subject to them, and would lack legitimacy. In research, appealing instead to the good outcomes a decision brought is a narrow route: decisions about which unknowns to pursue are made before results are known, and the results of lines not chosen are not observed, so there is often no other outcome to compare with. The second is domination: on republican views that value freedom from domination, even decisions favourable to S amount, if S cannot revise them, to dependence on good will, and so to domination (Chapter 12, Section 3.4). In research, the very standards by which S judges whether decisions are favourable can move with the coupling of production and evaluation (Section 8), so giving up the capacity to revise on the strength of decisions' being favourable can also shift toward the configurations the footing from which S judges what is favourable. The third is epistemic drift: applied here, Dewey's view and Irzik and Kurtulmus's extend the problem to the epistemic side, since if S's needs and valuations do not reach the configuration, the weighting of errors and the choice of questions can drift from S's, and no one remains to notice and correct the drift. The fourth is irreversibility: since declining capacities are not easily restored, the room for choosing otherwise later itself narrows (Chapter 12, Section 7). On views that ground legitimacy in accord with normative facts or in epistemic advantage, the loss need not be a problem in the first two senses, lack of legitimacy and domination, but the problems in the last two, epistemic drift and irreversibility, can remain. Where S too is artificial, whether its members have standing is itself a question, taken up in Chapter 12, Section 5. Of what does not move inside with reliance on content, standing over ends concerns relations of authority rather than epistemic relations. To whom it belongs, and what it takes to keep it more than nominal, is taken up in Chapter 12, under accounts of legitimate authority (Section 1), standing that is more than nominal (Section 4), the case in which A and S are both artificial (Section 5), and decisions that remove the capacity to revise (Section 7).

7. What changes in criticism: cost, range, judgment, and arrangement

Section 1 named three changes that an artificial configuration makes to evaluation. Section 3 took up the first, the generation of critics in any number, as copies of one base or as critics made to differ. This section takes up the second, that those who produce results and those who evaluate them can be within one artificial configuration, with evaluation entering continually even while results are being formed. What then changes in criticism is not confined to cost. There are at least four aspects. One is cost. The second is range: checks that human checkers, given their limited time and attention, did not ordinarily perform become possible. The third is judgment: evaluators can err in places other than where humans err and can pick up different things, and for judgments of value the question arises whose valuations they reflect. The fourth is arrangement: the combination of who checks, when, what, and how, and of how verdicts are gathered, is no longer confined to the form of review. This section asks what is required for a result to become something S can rely on, and how that changes in each of these four aspects.

In human research, the function of making results something others can rely on has in recent decades been performed largely through peer review, a practice in which researchers other than the producers read and check a submitted result before publication. Review, however, is one realization of the function, not the function itself, and a realization shaped in history. By 1832 the Royal Society had made a written report on a paper, by one or more members of its Council to whom the paper had been referred, a condition for publishing it in the Philosophical Transactions (Moxham and Fyfe 2018, 11). Sending every paper to be published to outside referees became settled practice at major journals much later. On Melinda Baldwin's account, many high-profile journals did not adopt external refereeing until the 1960s or later; Nature made it a requirement for every paper only in 1973, and the Lancet's editor kept the right to accept papers on his own authority until the mid-1970s (Baldwin 2015). The form in which a few referees read once before publication has plausibly been shaped in part by criticism's being costly and by the limited attention of those who check. In an artificial configuration that function may be performed without any stage resembling review, since if candidates are checked continuously while they are formed and reassessed whenever the record changes, there may be no moment at which a result has passed review at all.

What is required for S to be able to rely on a result depends on the account. On the view that places the value of criticism in its finding errors (Section 2; the error-detection view below), criticism needs to add information about errors through a route that would show the result or its grounds wrong if they were. Whether a reliant party must possess evidence of that contribution depends on the account of reliance (Section 5.2). On critical contextual empiricism (Section 3), it needs to have been open to criticism from the relevant perspectives in venues public to the community whose knowledge is at issue, with uptake, that is, with the community ready to revise what it accepts in response to the criticism raised. On assurance accounts, which find a reason for reliance in a speaker's vouching for the content, someone needs to vouch for the result and, on those that add a right of complaint, answer for it if it is wrong. On necessary-trust accounts, on which reliance in science requires trust in someone who can be responsible, such a party is needed behind what is relied on, though these accounts divide over who it must be and for what (Section 5). Two further judgments arise when certification includes claims to novelty or significance: whether the result is new to S rather than merely absent from the literature and data the producer searched, and whether it matters for S. These are not conditions of every warranted result: a replication can be reliable without being novel. Where those claims are made, for whom the result is new and important (their reference) and their grounds are to be specified (Chapter 9, Sections 2 and 3).

Human research has bundled these functions in several ways. Scholarly communication performs what Van de Sompel and colleagues, following Roosendaal and Geurts, list as registration (recording who showed what when), certification (recognizing a result as sound), awareness (making it known), archiving (preserving it), and reward (conferring credit), functions that journals came to bundle and that can be separated and reassigned (Van de Sompel et al. 2004). Peer review is one historical bundle within this, and this book distinguishes, for the purpose of this section, the functions it has combined: screening whether a manuscript fits a journal's scope, allocating scarce space or money, detecting error, improving exposition, assessing novelty and importance, and conferring recognition. This division sorts review by function so that functions whose costs change differently can be followed separately. Which of these functions a given review performs well is a separate question for each. Other arrangements distribute the same functions differently: publishing first and filtering afterward, certification through use and replication, continuing verification of a maintained record, and proposals to abandon prepublication review altogether, among others. What bears on this chapter is how recognition and the allocation of attention can be joined in a review decision. Where passing review is used as a sign of recognition, the gate also distributes editorial attention and publication space. Reviewing labor can also be unevenly distributed: a model of biomedical publishing estimated that 20% of researchers performed 69–94% of reviews in 2015 (Kovanis et al. 2016, 1, 6). Through the same single stage at which it performs these functions, review has plausibly also provided at least three further things. One is making the fact of evaluation visible to S. The second is giving members of S a place to gain experience in evaluating. The third is leaving the judgment of someone who can be named in a form that can be objected to. The consequences, considered below, of the disappearance of an explicit stage of evaluation concern mainly these three.

Take cost first. Artificial systems already in use weaken the premise that criticism is costly to produce, and they weaken it unevenly. In a randomized study at a large machine-learning conference, reported in a preprint, AI feedback was sent to reviewers on randomly selected reviews they had just submitted; 27% of the reviewers who received feedback updated their reviews, and when two researchers compared, blind, a hundred updated reviews that had incorporated feedback with their originals, they preferred the updated version 89% of the time (Thakkar et al. 2025, 1, 10). The example illustrates the claim that criticism of texts can be multiplied at low cost, here even criticism of criticism; it says nothing about criticism that reaches a defeat relation (Section 1), a route that would show the result wrong if it were, such as a new measurement, a rerun of an experiment, or a formal check. Criticism of texts, code, statistics, and consistency with the literature can be multiplied cheaply, and some of it can itself undercut a result's grounds, as when it finds an error in an analysis. Criticism whose defeat relation runs through a route the producer does not supply, such as a replication in a laboratory, a new measurement, or a proof checked against a kernel the generator does not control, becomes cheaper only where that route itself becomes cheaper, although automated re-execution and automated formalization are among what has become cheaper. Direct evidence on whether the capacity to produce results or the capacity to appraise them competently is growing faster remains fragmented; what is suggested here concerns the structure, that some criticism becomes cheaper and some does not, not the size of the gap.

Change in range can come about through change in cost, but its consequences are different. Once the cost of checking falls far enough, kinds of check that had been too burdensome to be carried out in practice may become feasible: for example, reimplementing a paper's method to see whether its results reproduce, rerunning submitted code and data, and searching the related literature exhaustively to see whether a result is already known and whether the works cited exist and support the claims made. Checks of this kind, which can take a human considerable time for a single paper, an artificial configuration may be able to attempt with far less burden. If so, what is checked changes. Where reimplementation and rerunning are possible, part of the defeat relations of Section 1 moves from rereading the record to running things and seeing. Where exhaustive search is possible, the judgment whether a result is new to S can be made against a wider record rather than against whatever the producer happened to search. Being able to attempt such checks, however, differs from succeeding at them. In one study that gave models only the text and figures of 83 papers, with 91 errors, that had led to errata or retractions and whose errors their own authors had acknowledged, none of the ten models tested in early 2025 found most of those errors, and even the best found about a fifth of them in a single run (Son et al. 2025, 3–5). Because that study excluded errors requiring external code or data, the success of checking by reimplementation or rerunning is a separate question. And however far the range widens, it does not reach new measurements or executions outside the scope of the record.

Two consequences follow from these changes in cost and range. The first follows from their unevenness and concerns the error-detection view, on which the unevenness matters because much of the criticism that becomes cheap works on the producer's record. Errors that leave no trace in the record, such as errors in how it was made, cannot be found by rereading it, however much the critics differ from one another (Section 3, the second sense). Where cheap critics are also built from the producer's base or from tools of the same kind, they are likely to go wrong where the producer does even about what the record shows, and many of them agreeing add little, for Goldman's reason that sources merely reflecting one another add no evidence however many agree (Section 2). Differently made critics may be cheap as well (Section 3), but the volume of criticism does not show whether they are. Abundant criticism can then coexist with a scarcity of independent criticism, and conceal it. What can remain scarce is not attention in general but defeat relations through routes the producer does not supply, and, where critics are made like the producer, critics whose error routes are their own. The second follows from criticism's becoming abundant, and concerns the point at which it is applied. If the argument of Chapter 5, Section 4 is right that judging what is worth pursuing differs from judging what is likely true, criticism applied indiscriminately to candidates still being formed can turn away promising candidates with little support yet, merely because they seem unlikely to be true, and so shut out ventures that might prove fruitful.

In judgment, artificial reviewers need not err where humans do. In one study in which 45 researchers in the relevant fields assessed, item by item, the comments of human reviewers and of three language-model reviewers on 82 papers in Nature-family journals, about a quarter of the artificial reviewers' comments had no counterpart, among any of the human reviewers' comments, that made the same criticism of the same point. About four-fifths of these were judged correct by the experts, but they were rated most significant less often than those that overlapped with human comments. The experts' remarks also identified weaknesses that, the authors state, are not seen in human reviewers, such as treating practices regarded as ordinary in the field as defects and so misjudging the weight of a comment, and saying that something the paper states is missing (Kim et al. 2026, 9–12). Nor is there a single direction in which artificial reviewers lean. In the published version of the AI Scientist, discussed in the Introduction, when the reviewing component was compared with ICLR decisions, the rate at which it judged rejected papers acceptable was 0.45 to 0.52, higher than the 0.17 found by an experiment at another conference that measured agreement between the decisions of two committees, and the rate at which it judged accepted papers unacceptable was 0.17, lower than that experiment's 0.52 (Lu et al. 2026, 916). Its judgments leaned toward acceptance. Beel and colleagues, discussed below, by contrast, had the system as first released assess ten human-written papers and read a strong conservative bias. Since the models, the number of papers, and the manner of evaluation differ, the two cannot be compared as a change in one system. Still, they show that the direction of bias seen in one evaluation of one artificial configuration cannot be read as a general property of artificial reviewers. Misjudging the weight of a comment also bears on which norms a judgment of value is made against, since what matters, and which defects are serious, is relative to the field and to the valuations of S. Evaluating artificial reviewers thus involves asking not only whether they are correct but against whose valuations they judge.

The aspect of judgment brings with it the question of how artificial reviewers themselves are to be evaluated. One method compares their judgments with human decisions. What agreement measures depends on the function and on the account. On accounts that take review to aim at correctness, agreement measures fidelity to the practice whose decisions serve as the reference, including its blind spots, not correctness. Where the humans whose decisions are compared are members of S, that is the certifying practice of S; where S is artificial, or mixed, agreement with human decisions measures fidelity to a practice outside S, or to that of only some of its members, whose relevance to certification for S has to be argued. On accounts on which certification consists partly in applying a shared practice uniformly, in the way that legal determinacy, as Porter reports Kent Greenawalt's account of it, requires only that different judges apply the law to most cases in the same way (Porter 1995, 4–5), fidelity to the practice may itself be part of what certification is for.

That practice need not return a single verdict. In a study reported by Weber and Inglis, for example, when 109 mathematicians were asked whether an argument that interchanged a limit and an integral without justification was a proof, 27 percent said that it was and 73 percent that it was not, and only one of those who had accepted it agreed that an objection to that step made it invalid; on proofs using routine techniques from number theory, nearly all the mathematicians surveyed in another study accepted them (Weber and Inglis 2024, §4.3 and §2.2 n. 3). Agreement with a divided practice depends on which part of it is taken as the standard. For functions other than reproducing past decisions, agreement may say little. One exploratory study compared pairwise rankings by a language model with funding decisions on 139 proposals to the Austrian Science Fund. The two selections agreed in 87 cases, and higher human-rated novelty was associated both with disagreement and with the direction in which humans funded a proposal that the model did not (Machado 2026). The sample is small and the design does not identify why the pattern occurred. It illustrates why disagreement cannot be read directly as error: here it was concentrated where a function other than reproducing past decisions was at stake (Section 8 returns to novelty).

Two further reports bear on the same question. Beel and colleagues had the reviewing component of the AI Scientist (Introduction, Section 2), in a version they modified for the purpose, assess ten human-written papers from an open review platform, five of them accepted and five rejected by human review; it recommended rejecting nine of the ten. They read this as a strong conservative bias and misalignment with human judgment, and ask whether human judgment is the right benchmark at all, since a reviewer that diverges from human reviewers may still identify valid weaknesses, or may apply overly rigid criteria (Beel et al. 2025, §2.6). At a conference organized so that artificial agents served as authors and reviewers, three language-model reviewers, whose instructions the organizers had refined to maximize correlation with human scores from two earlier ICLR editions (2022 and 2025), correlated with one another at 0.48 on average, differed from the scores of human experts by mean absolute differences ranging from 0.91 to 2.73 on a six-point scale, and showed sycophancy, especially one of them, which in one review called "technically flawless" a paper whose experiments the human reviewer suspected were incomplete or selectively reported (Bianchi et al. 2025, 4–5). Both are reports about particular systems at one time and, in the second case, at one venue. They suggest that refining instructions to match human decisions (calibration in this sense) does not fix which part of a divided practice the calibrated reviewers reproduce, and that disagreement with human decisions need not have a single kind of source.

In human research, that a competent person with limited time had examined a result and let it stand was plausibly part of what gave others reason to rely on it. If review becomes abundant and artificial, that reason does not transfer automatically. Passing an artificial review gives S reason to rely on a result only to the extent that reliance on the reviewer for that function is itself warranted; whether the one who relies needs reasons of their own for trusting the reviewer (reductionism), only the absence of reasons for doubt (non-reductionism), or only the reviewer's actual reliability (reliabilism) depends on the account of reliance (Section 5.2).

Turning to arrangement, in an artificial configuration, as noted at the start of this section, there may be no such moment of passing at all. The room for recombining these functions also widens further. Those who check need not be a few human referees; they may be copies of one base or many checkers of different make (Section 3). The time of checking need not be once before publication; it may be while a result is being formed, after it has been used, or whenever the record changes. The manner of checking need not be reading and writing an opinion; it may include actually running code and data, checking formally, having different answers argued against each other, scoring predictions against later outcomes, and deliberately attacking the checkers themselves to find their weak points. The object of checking need not be the text of a paper; it may be code, data, execution logs, or the dependencies themselves. Nor need verdicts be gathered by an editor's judgment; they may be gathered by majority, by weighting, or with dissent preserved, among other ways. Which of these combinations can perform the function of making a result something others can rely on is to be asked of each combination, in light of what, as set out above, is required. The question thus becomes not whether to keep review but which combination meets what which account requires.

The judgments review has bundled can also be gathered into two kinds. One is a judgment of a result's credentials as knowledge: whether a result holds up as research, and how sound it is. The other is a judgment of value: how important or interesting the result is for S. Why human research has had these two made at a stage separate from production, by parties other than the producers, may have no single answer. Some of the reasons one can think of, however, are tied to the premise that research has been produced by individuals or small groups. For the judgment of value, the values of the producing individual or group can differ from those of S, so there may have been point in placing a filter between them. For the judgment of credentials, a small group alone may be unable to carry it out adequately, or repeated checking may judge it better. If these are the reasons, they plausibly include parts that answered to each of two kinds of premise in the sorting of Chapter 9, Section 6. That a small group cannot check enough turns on research's being carried out by humans with limited time and attention, a premise about human realizers. Placing a filter between producers' values and S's, and having checking done at a separate stage by other parties, turns on research's having been organized so that individuals or small groups produce it, a premise about how research has been organized. If so, then when an artificial configuration can carry out both judgments adequately within itself, a review placed afterward as a separate stage can be redundant in the sense of doing the same thing twice. Carrying out the judgment of value adequately, though, means that the judgment of how important a result is for S matches S's valuations. This is a question of whose valuations the artificial configuration's valuations match, structurally similar to AI alignment (Section 4; Chapter 12), and competence alone does not settle it.

That does not end the matter. First, there is the view on which doing it twice is itself of value. The question it raises is whether the second pass adds a check that errs in different places from the first, such as a test working on a different principle or a separately obtained measurement (the second sense of redundancy distinguished in Section 3), or differences in applying the same standards (the third sense). Second, if certification is taken as the relation that connects what A produces with what S can rely on (Chapter 9, Section 2.1), then as long as A and S can differ, that connection is needed whatever A and S are. If this connection is what review has provided, review can remain as an abstract function even when it loses its form as a human institution. This is the fourth sense of redundancy distinguished in Section 3, in which stages have different addressees and functions.

Evaluation may, further, not appear as a separate stage at all. In human research too, judgments of which questions to pursue, which candidates to develop, and which lines to abandon have been made inside the process of inquiry (Chapter 5, Section 4; Chapter 10). In an artificial configuration, the explicit process of evaluation itself, in which a finished result is checked afterwards, can dissolve into the choice of questions and the formation of candidates and cease to be visible as such. Whether that suffices depends on where the function of evaluation is placed. On the error-detection view, if checks that add information about errors are carried out, the requirement can be met even if they are made implicitly inside the process. By contrast, on Longino's account, which requires that criticism take place in venues public to S, on assurance accounts, which require someone who vouches for the result, and on reductionism, which requires reasons available to those who rely, the outcome of evaluation, or at least a trace of it, has to come out toward S. On those accounts, evaluation settled implicitly inside does not by itself amount to certification for S. What counts as coming out depends on what S consists of. If S is human, it has to take a form that humans can read and question. If the members of S are artificial and can share representations directly with the artificial configuration, it may take a form other than a report (Chapter 3, Section 8). Even then, whether the members can themselves check the shared content is a further question.

Where the explicit stage of evaluation disappears, several consequences follow. One is that, on accounts that require evaluation to be made explicit to S, an artificial configuration needs a layer, distinct from its internal evaluation, that it puts out toward S. Part of the work of evaluation then moves from checking to conveying the outcome of checking to S, and the explicit stage may come back in another form rather than disappear. The second concerns where S's capacity to check is developed. In human research, reviewing has plausibly also been a place where the capacity to evaluate was acquired and kept up. If that stage disappears, the opportunities for members of S to gain experience diminish as well, and S can come closer to being a layperson relative to the artificial configuration (Section 6). The third concerns objection. To object to a judgment one has to be able to see it, so implicit judgments are hard to object to, and revisability depends on explicit traces (Chapter 12, Section 6). The fourth concerns time. When evaluation stays inside, coupled with production, a decline in an evaluator's independence or a shift in standards is hard to notice from outside, and reassessing what depends on a defeated claim requires a record of what was judged where (Section 8). The fifth concerns allocation. Passing review was also a signal for allocating funds and positions, so if the moment of passing disappears, allocation needs another signal (Chapter 10, Section 3). On the error-detection view, on the other hand, implicit evaluation may catch errors earlier and continually, and the rigidity of a gate under one criterion that turns candidates away together may also be eased. And if the members of S are artificial and can share representations with the artificial configuration, a layer put out toward S may not be needed. Then, however, the distinction between A and S itself thins, and who has standing becomes a question (Chapter 12, Section 5).

In sum, the changes an artificial configuration brings to criticism fall into four aspects.

• Cost: what becomes cheaper is the cost of some kinds of criticism only. On the error-detection view, the criticism that becomes abundant, so far as it shares the producer's error routes, adds little independent information and can conceal the scarcity of what does.
• Range: checks that humans did not ordinarily perform come to be attempted, so that part of the defeat relations can move from rereading the record to confirmation by execution. Attempting, however, differs from succeeding, and nothing here reaches beyond the record.
• Judgment: artificial evaluators may catch what humans overlook, and may also make errors humans do not. For judgments of value, the question is whose valuations they reflect.
• Arrangement: the function of making a result something others can rely on can be performed by many combinations not confined to the form of review. Evaluation can also dissolve into production, so that no explicit stage remains.

On that basis, the six functions that review has bundled go different ways. Screening for scope can dissolve into the choice of questions and the formation of candidates. The gate that allocates space and money answered less to premises about human realizers than to premises about the present organization of science (Chapter 9, Section 6), and can lapse with them, as it can also when that organization changes while its realizers remain human. Where the time and computation available for checking are limited, however, not every candidate can be sent to a costly check or to S's attention, the same reason of finitude that Chapter 10, Section 1 gave for allocation; where checks remain costly, some gate remains before them, and a gate operated by one criterion can turn candidates away together, in the direction of that criterion's mistake (Chapter 10, Section 4 found the same structure where one signal judges the survival of lines). Error detection can, on the error-detection view, be carried inside the artificial configuration. What joining production and evaluation does not remove, on that view, is the need for criticism that adds information about errors and remains responsive to what it finds. This need not require that someone other than the producer read the result again: a proof assistant's kernel, the small program that checks each step of inference (Section 2), or a measurement the producer does not control can show a result wrong, if it is, without anyone reading it. Improving exposition, together with checking consistency and comparing a result with the existing record, can be done by the producer as it produces. Assessing novelty and importance can also dissolve into the choice of questions and the formation of candidates. Assessing importance, however, is a judgment of value and leaves open whether the artificial configuration's valuations match those of S; assessing novelty leaves open whether the result is new to S rather than merely absent from what the configuration searched. Recognition has also served as a signal on which later allocation of funds and positions relies, so where allocation steers by it, it is sought in another form even when scarcity of space and attention is gone (Chapter 10, Section 3). The three further things review can be thought to have provided along with these functions, the visibility of evaluation to S, a place for S's members to gain experience in evaluating, and judgments that can be objected to, can, as seen above, be lost where the explicit stage disappears, unless they are provided in another form. On the other accounts, the publicity of the venue and someone who answers are not made unnecessary by joining. And so long as A and S differ, on each account considered, the relation that connects what A produces with what S can rely on remains. For each function, whether it suffices for it to be performed implicitly inside the process or whether it requires being made explicit to S depends on the account adopted and on what S consists of.

And no change in any of these aspects, on any of the accounts considered here, yields certification by itself. Certification requires that reliance on the critic be warranted in whatever way the account of reliance demands (Section 5.2), and, depending on the account, public venues or someone who answers. This conclusion about the unevenness of cost rests on the structure of the cases cited, not on data measuring how fast the cost of which kinds of criticism is falling.

8. Criticism over time: proxies and standards, evaluating the new, and correction

Sections 2 to 7 treated independence as a relation between a critic and a claim at a time. The third change named in Section 1, that an artificial configuration can tie evaluators and producers together over time by adjusting each to the other's outputs, for example by training or selecting each on them (coupling, below), makes it also a relation over time. As Section 7 noted, the further evaluation dissolves into the process, the harder such changes become to see from outside. This section adds a condition of time to the conclusions of Sections 2 to 7, and prepares one of the requirements Section 9 places on an artificial configuration: showing that independence is maintained as producers and evaluators adapt to each other. Two points of Chapter 9, Section 4 also concern time. One is that evaluation of research has to work through proxies, because what it is about is usually not observable when the judgment is needed. The other is that, for something new to S, the evaluator may lack a standard. The section takes up three things in turn: how selection on proxies moves critics' independence and evaluative standards, how what is new is evaluated where there is no standard, and how correction reaches what depends on a claim after it is defeated. Proxies that come to be used in decisions about hiring and funding can change what they measure, standards can move or be held fixed even when defective, and a claim's defeat need not reach what has come to depend on it. Signals by which allocation steers are proxies of this kind, and this section is where the book examines their properties; Chapter 10, Section 3 takes up how allocation steers by them.

The first point, that proxies which come to be used in such decisions change what they measure, has a long lineage. Donald Campbell argued that increasing the use of a quantitative indicator in social decisions both increases pressure to manipulate the indicator and makes it more likely to damage the processes it was intended to track (Campbell [1976] 2011, 34). In research, once an indicator affects hiring, funding, or survival, those evaluated adapt to it, and the measure becomes part of the system whose performance it reports. In human research this adaptation runs through motives and careers. It is broader than deliberate gaming: a researcher can honestly redirect effort toward what is measured, and when many do, the meaning of the indicator and the population it describes can shift (Wilsdon et al. 2015, 81–83). For artificial producers the adaptation need not run through motives, whatever one concludes about whether motives can be attributed to them (Chapter 7, Section 3.3); it can run through optimization and selection: outputs that score well are kept, and producers are trained or chosen on that basis. Evaluators can in turn be trained on past decisions, and the outputs they selected become future data. A model update can change the operative evaluative rule without any explicit change in policy. David Manheim and Scott Garrabrant, writing about metrics used in optimization generally rather than about research, separate four ways in which the relation between a proxy and a goal can collapse under optimization. Selecting on an imperfect proxy also selects for the difference between proxy and goal, which they hold "can not be avoided" (regressional); selection can push into regions where a relation observed elsewhere no longer holds (extremal); intervening to raise the proxy can change the causal path that linked it to the goal (causal); and agents with goals of their own can act on a proxy they know to be in use (adversarial), the case to which they relate Campbell's statement. They add that how much these effects matter depends on the optimization power directed at the proxy (Manheim and Garrabrant 2019, 1–2, 4, 8). Carried over to research, their classification separates the two routes distinguished above, adaptation through motives and adaptation through optimization and selection alone: the adversarial case needs a party that responds to the measure, whereas the regressional and extremal cases need only selection on it.

Taken together, early observational studies report that AI involvement is associated with greater novelty or impact of individual papers and with a narrowing of the collective range of topics, and that grant proposals written with more AI involvement (measured as the estimated share of language-model-written text in proposal and award abstracts) sit closer to recently funded work. Selection on proxies pulling work toward what is easily measured is one route by which such narrowing could arise; but these are patterns consistent with such narrowing, not established effects of it, and other routes remain possible (Bianchini et al. 2026; Hao et al. 2026; Qian et al. 2026). The studies count different things as AI involvement. In Hao and colleagues' study of papers in six natural sciences from 1980 onward, it is the use of AI methods, most of them conventional machine learning and deep learning, and the authors describe their analysis of the period of generative AI as preliminary (Hao et al. 2026, 2). A study of language-model use in writing reports a pattern in another direction at the level of inputs: Kusumegi and colleagues find that authors whose preprints show signs of language-model assistance go on to cite a more diverse body of prior work, including more books and younger and less-cited documents, while noting that their design does not provide causal identification (Kusumegi et al. 2025, 5–7). The diversity of what is cited and the topical range of what is produced are different measures, so the second finding does not refute the first; together they caution against reading either as the effect of artificial systems on the range of research.

This bears on independence directly. A critic can add information about a producer's errors at one time and less later, as training or selection comes to favor outputs that pass where the critic is insensitive. The checking rule need not change. Nor need the producer represent the critic: selection on its verdicts can have this effect, as can a producer's fitting outputs to a model of its evaluator. Here evasion names adaptation that weakens the connection between passing and being correct, rather than every occasion on which a fallible test accepts a false candidate. In the terms of Section 2, what evasion lowers is the check's severity for the producer as it has become, the probability that the check would expose the errors that producer now makes.

Adaptation can also improve the capacity being checked. Learning to produce proofs accepted by a sound checker, one that lets no invalid proof through, can improve proof construction without weakening that check's force for the formal statement. The relevant contrast is therefore not adaptation versus no adaptation, but improvement in the target capacity versus exploitation of what the check misses. No inference to either follows from a rising pass rate alone. Evidence for distinguishing them can include performance on routes withheld from selection, variations designed to expose known blind spots, and whether what the route rewards continues to bear on its intended target (Chapter 3, Section 3.3; Chapter 7, Section 3.1). Such evidence remains scoped to the errors and changes it tests. Protecting a rule from rewriting can help preserve its evidential contribution, but cannot establish that contribution or guarantee its persistence. Chapter 10, Section 4 found that bringing allocation into one configuration does not by itself make the criterion for deciding which lines to keep and which to close checkable by a separate route; here the point is that organizational separation alone does not preserve it through time.

A measure can also come apart from what it is taken to report without any adaptation to it. Carole J. Lee, writing about AI tools that evaluate science, describes, following Cartwright and colleagues and Reiss, what the paper calls inference by false ascent: a measure designed for one concrete purpose is redescribed as a measure of a more abstract quality, and because that quality covers other concrete purposes, the measure comes to be used for purposes for which it is not reliable. Lee's historical example is the journal impact factor, introduced to help librarians decide which journals to buy, then taken as a measure of scientific impact and used for the impact of articles and authors (C. J. Lee 2026, 2–3). The case from AI is a set of models trained to predict replication outcomes. Their labels track, in effect, whether a replication found a statistically significant result in the same direction as the original, a measure suited to asking whether the replication obtained the same result; the models were described as predicting replicability and proposed as a way to improve confidence in findings, a purpose for which, on Lee's analysis, a measure that combines the original and replication data would be more apt (5–7). Lee develops the point for supervised predictive models and takes it to generalize (3). On this account, assessing reliability by the proportion of accurate outputs alone does not register this failure, since reliability is reliability for a purpose, and disputes over purposes and purpose-specific reliability are to be taken up through Longino's norms of critical discourse within scientific communities (7–8). The failure Lee describes lies in the inference drawn from an evaluator's output rather than in the producers' response to it, so it can arise however faithfully an evaluator does what it was built to do. It has the same form as Otsuka's observation in Section 5.2, read there as an instance of the generality problem: reliability established for one use is carried to another. For evaluators whose verdicts are relied on in research for S, the record of reliability for the use at hand that a member of S needs on reductionist accounts, which ask for positive reasons the one who relies can obtain (Section 5.2), would then need to state which use that is.

The coupling can also move the standard itself, not only behavior under a fixed standard. Formal work on AI systems whose reward functions stand for the evaluations of the people they serve, evaluations that change over time and that the system's own actions can influence, finds within its models that each of the eight notions of alignment it compares either leaves room for undesirable influence on those evaluations or is so conservative that inaction is often the only optimal behavior; holding the system to the evaluations as they stood at the outset is one of these notions, and it can lock them in or even create an incentive to move them away from where they started (Carroll et al. 2024, 5707, 5709–5714). The authors suggest that which notion is correct may not be identifiable from the formal structure of a case alone (5708). Their models thus do not decide when a change in an evaluative standard is learning and when it is drift; carried over to evaluative standards in research, an application the authors do not make, they suggest why agreement with a standard at one time does not settle what the standard will have become.

Several authors locate one direction of such movement on the side of those for whom research is done. Stuart Russell treats human preferences as changing through experience (Russell 2021, §1.6, author-version 17) and argues that AI systems cannot entirely avoid affecting them (Russell 2020b, transcript, paragraph beginning "One of them being the plasticity"). An intelligent machine may also have reason to make preferences easier to satisfy by changing them (Russell 2022, 53). Wolchover reports Russell's proposed response: machines could learn people's meta-preferences, their preferences about which processes of preference change are acceptable (Wolchover 2020, "Our Imperfect Choices"). Russell does not apply this argument to research. Applied to it, the argument would make S's evaluative standards something a configuration can move as well as serve. Writers on research describe such movement there. Emanuele Ratti argues that machine learning "brings its own normativity": in cancer genomics, for example, its use is likely to guide molecular biologists away from the mechanistic explanation they take as the goal of their discipline toward prediction and classification, constraining which epistemic values they prioritize, an impact Ratti holds that scientific methodologies generally have (Ratti 2026, 16–18). Johan Commelin and colleagues, writing for the mathematical community, hold that "means reshape ends": the capabilities of new technology risk "dictating what counts as mathematically significant", and they ask that adoption remain rooted in the community's own epistemic and aesthetic values (Commelin et al. 2026, 2). On these accounts the standard that moves with the coupling may be S's own, and agreement between a configuration's outputs and S's present standard does not by itself show that the configuration has not moved it. Chapter 6, Section 5.3 distinguished, for a single inquiry, revising a standard for reasons from merely moving the target when performance is poor; coupling between producers and evaluators poses the same distinction for a system in which no one decides to move the target. Russell's remedy, carried over as well, would have a configuration learn S's standards for which changes of standard count as revision for reasons rather than movement of the target, instead of holding to a fixed standard; whether its learning them can stand in for S's deciding them is part of the question of who may revise a standard (Chapter 12, Section 2).

Maintainers of an evaluation can face the opposite risk, keeping a standard known to be defective so that earlier results stay comparable. Those of one benchmark for machine-learning agents, for example, chose not to fix a known leak of test labels for the time being, so as not to invalidate earlier leaderboard results, deferring the fix to a later batch of fixes, and later paused new submissions while developing a process to keep them fair and comparable (OpenAI 2025; OpenAI 2026e).

The second point concerns evaluation when the evaluator lacks a standard. For a result new to S, the standards by which it would be judged may themselves be in question. Kuhn held, as Eric Oberheim and Paul Hoyningen-Huene report, that the standards of theory appraisal, such as simplicity, accuracy, consistency, scope, and fruitfulness, vary with the dominant paradigm, and that they function not as rules that determine theory choice but as values that merely guide it, so that scientists who weigh them differently can disagree rationally even across incommensurable paradigms (Oberheim and Hoyningen-Huene 2024, §2.3.2). On Kuhn's account, as Section 3 noted, what lets novelty through is partly the variation among evaluators in how they apply shared values. Seen in this light, an evaluator trained on past decisions holds its standard as the distribution of those decisions, and is therefore liable to score low a candidate whose acceptance would require revising the standard, scoring it as a deviation. The Machado study of Section 7 is consistent with this, though it does not show it. If Kuhn's account is adopted, what such evaluators would need is variation in application. On this view, evaluators trained on the same past decisions may have little of it, whether they are copies of one base system or built from different bases; for artificial evaluators, such variation is not given by what the realizer is like but has to be produced, for instance by separating their training histories. Since Kuhn allowed that variation to come in part from factors such as life history, personality, and influences from outside the sciences (Section 3), whether variation produced in other ways would do the same work is not settled by that argument; on accounts on which standards for the new are formed through the criticism of it (Chapter 6, Section 5), what would matter is whether evaluators can take part in forming them. Who may revise a standard is a question for Chapter 12, Section 2.

The third point concerns what happens after a claim is defeated. A false claim was not knowledge, however well received; losing grounds for a claim need not establish that it is false. Dependent claims must be reassessed to determine which lose support and which retain independent grounds. On accounts that locate knowledge in a capacity a community keeps using, checking, and correcting (a live epistemic capacity), propagating criticism is part of maintaining that capacity, rather than a rule that every dependent result automatically ceases to be knowledge. Chapter 2 distinguished a dormant record, one merely stored and used by no one, from a live epistemic capacity partly by whether correction propagates, and even on Bird's view, on which a dormant record can remain social knowledge (Bird 2024), whether correction reaches dependent work would plausibly bear on whether reliance on the record remains warranted. Detecting a defect, changing the article of record, and repairing dependent work are different achievements; others could be added, such as informing those who have already acted on a result, which this section leaves aside. The correction to the materials laboratory discussed in Section 2 changed the article of record. In a separate case, the Massachusetts Institute of Technology stated publicly that it had no confidence in the provenance, reliability, or validity of the data in a preprint on AI and scientific productivity, or in the veracity of the research, and requested its withdrawal; a peer-reviewed article in 2026 later cited the preprint without qualification (MIT 2025; Soleymanibrojeni et al. 2026). This was not a failure of an autonomous system and does not show how often propagation fails; it illustrates the claim that the third achievement does not follow from the first two. Within a single body of work, tracing dependents can accompany detection. In early December 1993, when Andrew Wiles announced that a calculation in the argument for Fermat's Last Theorem was not yet complete, a report on Wiles's lectures about to go to press stated which of the announced results depended on that calculation, Fermat's Last Theorem among them, and which did not (Rubin and Silverberg 1994, 2, 18).

Here too a premise changes form. In human research as it has so far been organized, dependencies are tracked only in part; citations record some of them and not others. Inside a configuration that logs what each result was derived from, the tracing of internal dependents can in principle be complete. Whether each dependent still stands without the defeated claim is a further question: answering it is itself criticism, and needs a defeat relation of its own. Among what depends on a checker are the certifications it issued: if the checker is later found unreliable or not independent of what it checked, they lose the support it gave them. Dependents outside the configuration's record, the uses others within S have made of its outputs, lie beyond the tracing the record makes possible; the boundary of the record of Section 5.3, beyond which nothing can be checked, returns here as a boundary of where correction can reach. The routes by which a correction would reach those uses are of the same kind as those Section 5.2 found an internal finding needs in order to act as a defeater for those who rely, now needed after the result has been used. Longino's condition of uptake states a related requirement at the level of the community: its beliefs must change over time in response to the criticism taking place within it (Longino 2025, §5). Where checking dependents again is itself costly, there is also a question of order: which to check again first. If Varsavsky's point in Section 4 is extended from the order of testing to the order of correction, allocation remains even after an error has been found, since, depending on which dependents are repaired first, the same capacity for correction can leave different people exposed to the error for different periods. Ordering repairs by the number of dependents alone, for instance, is then not a value-free choice either; it can contain a weighting of inductive risk (Section 4) over who bears the costs of the error and for how long.

Chapter 12, Section 6.3 places this section within the book's handling of error. What has come into view in this section is that independence and certification established at one time do not by themselves guarantee that they hold later. Following the section's three points, three things can erode over time.

The first is a critic's independence. As producers and evaluators adapt to each other, an evaluator's passing verdict can come to indicate less and less about whether the producer is right. This decline can follow from the structure of coupling where the check is imperfect relative to its target, is used to select or train producers, and selection shifts the producers' remaining errors toward where the check is less sensitive; the cited studies are only consistent with it. A claim that an artificial configuration's criticism remains independent would therefore have to show not only that it was independent at a single evaluation, but how independence is maintained as producers and evaluators adapt to each other. One possible measure is to keep, separately, checks that are not used to select or train producers, such as tests held out from training or measurements the producer cannot alter. An evaluator used for selection can tell less about producers the more it is used; a check not used for selection is less exposed to that effect.

The second is the standard by which the new is evaluated. If Kuhn's account is carried over to them as above, evaluators trained on past decisions may be weak where the standard itself is in question. And since the standard can move with the coupling, agreement between an evaluator's outputs and S's present standard does not by itself show that the standard has not been moved. The movement itself can be detected inside the artificial configuration as well, against preserved reference points or evaluations built from S's past judgments. A reference point does this work only so long as it is not revised, and the benchmark case above shows the price: what is kept for comparability can be kept defective. But whether to accept a detected movement is for S to decide so long as the standard is S's, and what is at issue is whether S keeps the capacity to decide it (Section 6; Chapter 12). And where the standard moves in this way, Wilholt's correspondence reading (Section 4) can come to be satisfied not by the weighting's being revised but by the expectations of those who rely moving. The stronger reading, which also requires that those who rely be able to object, can resist this.

The third is correction. If a claim is defeated but the correction does not reach what relied on it, the error remains. On accounts that tie knowledge to a live capacity, keeping knowledge requires such routes. An artificial configuration that logs what each result was derived from can trace its dependents internally, though checking whether what is traced still holds is itself further criticism. And the record does not extend to the places where others in S have used its outputs. For corrections to reach them, routes leading outside the configuration's record are needed.

In sum, it cannot be said that independence or certification established inside an artificial configuration at one time will continue. The ways it can erode are three, and each needs something different. In a configuration where adaptation to selection can weaken a check's contribution, some other means of keeping that contribution is needed, and checks not used for selection are one candidate. As for movements of standards, detecting them can be done inside, but deciding whether to accept them is for S, and requires that S keep that capacity. And where outputs are used outside the configuration's record, routes that carry corrections there are needed. So, to the question of Section 9, once time is taken into account the answer may not be that everything can be done inside. What the last two need is not being human but relations to S, so this holds whether S is human or artificial (what counts as outside where A and S overlap is taken up in Chapter 12, Section 5). Nor does showing that a configuration was independent at one evaluation show that it can evaluate its own research.

9. Can criticism for S be organized inside an artificial configuration?

This section considers the chapter's question with the material assembled so far. In the example of Chapter 9, Section 1, a single system generates a hundred instances, has them check one another's products, and redirects resources in response to those checks. Those who build it can claim that results that have passed the internal checks are certified for S, so that others have reason to rely on them. This section asks under what conditions that claim holds, that is, whether criticism and certification for S can be organized inside an artificial configuration. It is also the question whether a way of doing research in which machines produce knowledge and evaluate it themselves can give certification for S. Answering it account by account is also a means of seeing what happens to research under such a way of doing it: knowing what can move inside shows what then ceases to be visible to S, and knowing what remains shows what S's side needs if the research is to remain research for S. Chapter 12 needs the answer too: if some part cannot be organized inside, one has to know what it is and why it remains before one can ask who may decide about it and who may revise it.

The section proceeds as follows. It first sorts out which question each of the accounts seen in this chapter answers, and gathers their verdicts in one table (Section 9.1). It then states what comes into view when those accounts are set side by side (Section 9.2) and what is suggested by extending that further (Section 9.3). Finally, it lists the questions that pass to Chapter 12 (Section 9.4). In outline, the answer differs by account. The work of finding errors can, on the error-detection view, be done inside if the checks err in different places from the producer. By contrast, acceptability for S, someone who answers to those who rely, standing for dissent, and the like remain, on the accounts that require them, as relations to S. And as Section 8 found, once time is taken into account, even for the functions that can be done inside it may not be possible to say that everything can be done inside. Section 9.3 goes on from there to what can happen to research: that if passing internal checks comes to serve as a proxy for certification, the lack of relations to S can be selected along with it; that independent checks can be depleted by use; that there may come to be no one in S able to carry the relations that remain; and that as A and S overlap, the question of certification can move to "for which society?" When this section says "inside", it is in the sense of Chapter 9, Section 5, which covers two things: that the work of checking is supplied inside the artificial configuration, and that the acknowledgment recognizing its outcome as certification for S lies inside. In this chapter's argument the two can come apart, and what remains as a relation to S is mainly the latter.

That verdicts diverge when different accounts are compared is, however, no surprise in itself. What this chapter's material brings into view is why verdicts diverge for artificial configurations. In human research, the properties that the accounts each track were tied to one another by human realizers and by the organization of research. Critics were formed as separate people, so their number roughly traced their independence (Section 3). Certification was a person's assertion, and if it was wrong there was someone who answered (Section 5). Producers and those who rely could be presumed to share a sense of the weight of errors (Section 4). Experts in agreement had been trained in separate places and had access to separate evidence, so the signal of agreement came with diversity (Section 6). The stage of review made evaluation visible to S and was plausibly also a place where the capacity to evaluate was developed (Section 7). In many everyday cases, therefore, accounts that ask about routes of error, accounts that ask about perspectives, and accounts that ask about someone who answers plausibly tended to lean toward the same verdict. An artificial configuration makes these properties settable one by one. Critics can be made as copies or with built-in differences, records can be kept from the run, the way verdicts are gathered can be chosen, and evaluation can dissolve into production. The coming apart of the properties ceases to be the exception. This is why the accounts of Table 11.1 reach different verdicts on the republic of instances (Section 3), and why the four answers of Section 5.2 require different things of the same artificial configuration. This does not mean that an artificial configuration necessarily makes the accounts diverge. It can also rebundle the properties, with different bases, records that reach those who rely, nameable answerers, and a layer put out toward S. What follows is that a builder's claim of certification cannot be assessed unless it shows which account it answers to, since evidence that meets one account need not meet another (Section 3; Section 5.4).

9.1 The questions the accounts answer, and their verdicts

To compare verdicts, one needs to separate what each account answers. The accounts of this chapter each answer one of three questions. One is what criticism yields. The error-detection view places the value of criticism in the information about errors it adds; Longino, Solomon, standpoint theory, and the ideal of mechanical objectivity each place it elsewhere (Sections 1–3; Table 11.1). The second is what grounds reliance on an unchecked result; what kind of act certification is taken to be (Section 5.1), what grounds the reliance (Section 5.2), and whether that reliance has to be trust (Section 5.3) belong here. The third is by whose valuations the threshold at which a result counts as sufficient for a use is set; Wilholt, Douglas, John, and Elabbar answer it (Section 4). Longino's account also bears on the third question through background assumptions, and Wilholt's on the second through trust (Section 4; Section 5.3). A position can be seen as a combination of answers to the three questions. The position that combines the error-detection view with reliabilism answers the first and second questions, one each, and since neither requires the reliant party's own reasons or S's perspectives, what can be placed inside becomes large. It says nothing, however, about the third question, so what remains concerning the valuations behind thresholds changes with whether Wilholt's or John's account is taken with it. Combining Longino's account with reductionism requires both perspectives and reasons on S's side. What remains as a relation to S differs from combination to combination.

Table 11.2 sets out, for each of these accounts, the verdicts of Sections 2–8. It is not a rigorous summary but a simplified overview for taking them in together, and the grounds for each cell are in the section given in parentheses. As in Table 11.1, the verdict in each row is reached by applying an account to artificial configurations its authors did not discuss. The second column names the human premise on which the account's conclusion relied where it held for human research. In some cases the account itself does not presuppose humans and only a rule of thumb relied on the premise; the table says so where that is the case.

Table 11.2. Account by account: the human premise, what can be carried inside, and what remains as a relation to S (a simplified overview)

Account	Human premise relied on	What remains when abstracted (inside, on what conditions)	What remains as a relation to S	What changes with the composition of S (including overlap of A and S)	If internal checks are better than humans
Error-detection view (Section 2)	none in the account itself; number and organizational separation served as a rule of thumb for independence because critics were formed as separate people (Section 3)	checks that add information claim by claim (can be inside); conditions: showing what information they add and how failures are diagnosed, and keeping independence as producers and evaluators adapt to each other (Section 8)	little on this view alone; taken with Longino's account, what a check counts against is relative to background assumptions (Section 1)	what matters is not human or artificial but shared routes of error; checks by members who share a base tend to err in the same places	applies directly, but evidence that the checks are better is needed (next row)
Reliabilism, combined with the error-detection view (Section 5.2)	none (no human required at the end of the chain)	reliable processes; for justification, that the process is in fact reliable; to show that it is, establishing reliability without the circle of measuring checkers by their own verdicts (bootstrapping)	the recipient's taking up within range; the passage from justification to use (the valuations of Section 4); outside parties who bear costs	where A and S overlap, checking and its evaluation become one process and tend toward the circle	applies directly to justification; but agreement with human decisions measures fidelity to a practice (Section 7), and beyond the range that can be checked, reliance falls on the handing on of calibration (Section 6)
Longino (Section 3; Section 9.1)	interaction among situated knowers in human research communities	the four conditions (can be met inside); acceptability remains relative to available perspectives, and the conditions include openness to criticism from outside the community	whether S's perspectives enter criticism; data evident to the accepting community; openness to criticism from outside the community	where S shares the lineage, the gap narrows without criticism improving; where the community coincides with S, the openness points outside S	strengthens the criticism already possible, but adds none from missing perspectives
Wylie's extension of Longino (Section 3)	collaboration among communities with different aims, practices, and beliefs (Wylie 2015, 13, author manuscript)	internal diversity alone does not establish external engagement	collaboration beyond the research community, which need not be beyond S	where the research community coincides with S, relevant external collaborators would be outside S	better internal checks do not by themselves supply external collaboration
Solomon (Section 3)	that influences on judgment are distributed differently from person to person; that their distribution, which no invisible hand makes good, is assessed and altered at the level of the community and its institutions rather than in individual reasoning (Section 3)	positions that assess and influence how research effort is distributed (can be placed inside); condition: occupants do not share the influences at issue	placing part of the identification of shared influences with parties who do not share them	what matters is sharing the influences; members of different make can carry this inside too; under majority voting, copies add weight without adding evidence	accurate individual reasoning is not enough; what is assessed is the distribution of influences
Standpoint theory (Section 3; Chapter 10, Section 6)	that marginalization and differences of starting point arise within the power relations and lives of human societies	engagement with the lives of the marginalized, and group consciousness; whether this can be carried inside depends on whom one learns from and inquires with	relations with those engaged with	in mixed societies the relation between human and artificial members can itself be a power relation; artificial societies trained on human records start where human research communities stood	what it asks for is not fewer errors but a different starting point, so accuracy does not supply it
Ideal of mechanical objectivity (Section 3)	distrust of, and distance from, judging individuals; individuals' idiosyncrasies and interests entering results	rule-following procedures (can be met inside)	questions to those who designed the procedures; procedures let decisions be made without seeming to decide	if the artificial configuration designs its own procedures, the question passes to who can check that design	if more uniform, may count as more objective on this ideal, but is silent on shared assumptions
Goldberg, as reported by Koskinen and Dixon (Section 5.1)	that reliance on instruments depends on others' epistemic work, and that social practices entitle users to expectations of designers and manufacturers	certification as the reading of an instrument	dependence on designers and makers, if outside; whether trust in them supports reliance on essentially opaque AI is disputed	if the artificial configuration designs its own checkers, whether its designing part has the understanding and responsibility required to support reliance (an application)	records alone do not establish who can answer
Reductionism (Section 5.2)	no premise specific to humans	track records (can be built inside)	reasons that reach those who rely, and that they can assess	if those who rely share the base, their assessment of lineage can correlate with the artificial configuration's capacities (Section 5.5)	applies only so far as records showing it reach those who rely
Non-reductionism, minimal form (Section 5.2)	none in the account itself; in human research, findings plausibly could reach those who rely through publication and correction (a premise of organization)	the absence of undefeated defeaters; but without routes, met whether or not there is an error	routes by which findings reach those who rely	if those who rely are inside the artificial configuration, the routes can come inside too	without routes, does not bear on the reliant party's justification
Assurance views (Sections 5.2 and 5.4)	that speakers intentionally assure and stand behind claims; answerability for failure on versions adding a right of complaint	intentional assurance (can be inside if the configuration or its members can give it); answerability where required (providing a base or deploying it is not assurance)	assurance to those who rely; someone answerable to them on right-of-complaint versions	overlap can bring assurance inside if instances can give it and, where required, answer for failure	accuracy alone supplies neither assurance nor, where required, answerability
Necessary-trust accounts (Section 5.3)	trust in responsible designers, maintainers, or investigators sustaining reliance	a party able to bear the responsibility the account requires (can be inside if the configuration or its members can do so)	trust in those parties; Koskinen requires responsibility for the system's workings, Ortmann trust in investigators who establish and transmit reliability	overlap can bring the trusted parties inside; the responsibility required depends on the account, and essential opacity is subject-relative	accuracy alone does not settle whom to trust, for what, or whether the required responsibility is possible
Wilholt: correspondence and the stronger reading (Section 4; Section 5.3)	that producers and those who rely share a sense of valuations; that part of the choices conventions leave is tacit and cannot be fully documented	records of losses and of the reasons for judgments; material for checking correspondence if the expectations of those who rely are ascertained separately	correspondence; on the stronger reading, a relation in which those who rely can object	where a machine produces and humans rely, a question structurally similar to AI alignment; reliers who share the lineage make sharing easier to presume, but those who bear costs can remain outside	accuracy does not settle how heavily losses are weighted, and for whom
Douglas (Section 4)	that thresholds are set by agents (scientists) morally responsible for the consequences of error	separating evidence from the weighing of error costs, and making the weighing explicit	which responsible agent answers for the weighing (Chapter 12, Section 6.3)	the line is drawn by where in the artificial configuration there is something that can answer as a responsible agent	does not settle who answers
John / Elabbar (Section 4)	John: that publication addresses audiences that cannot be identified; Elabbar: that evidence is distributed unevenly among people	applying a fixed high standard uniformly inside where results enter a record used by parties who cannot be identified, and a standard that follows identifiable hearers' where they can be identified (John)	the basic interests of evidence-poor parties (Elabbar); those who bear costs without relying	if S is artificial, costs can fall on humans or the environment (Section 4)	accuracy does not settle whose losses weigh heavily

The accounts of Fricker, Lackey, and Williamson of what kind of act certification is are gathered in Sections 5.1 and 5.5.

Longino's four conditions can be applied to an artificial configuration one at a time. As Dutilh Novaes and Dede report, these are idealized conditions that real communities meet by degree (Dutilh Novaes and Dede 2025, §2). As for venues, a channel internal to an artificial configuration is public to its components, not thereby to members of S outside it. Publicity is also graded: the report accompanying one large language model withheld its training method, among other things, citing competition and safety, had more than fifty external experts it had selected test early versions, and announced a plan to make details available to additional third parties (OpenAI 2023, 2, 12, 45). What those parties receive is known only through the developer's account. Uptake requires that the distribution of a community's beliefs change over time in response to the criticism made within it (Longino 2025, §5), not that outputs move in the direction a critic rewards; Section 8 saw that the latter can erode a critic's independence. As for public standards, an artificial configuration can make its standards explicit, but Longino treats them as provisional and subordinate to the community's aims of inquiry (Longino 2017, 51–52, 54), so who revises them, against which aims, returns to the question of whose unknowns (Chapter 10, Section 6) and to the question of authority (Chapter 12). As Dutilh Novaes and Dede report, tempered equality asks that social position or economic power not determine which perspectives are taken seriously (Dutilh Novaes and Dede 2025, §2). Where someone sets the weights of internal critics from above, the condition's work, keeping social position from deciding which perspectives are taken seriously, can be lost, since the judgment of whoever sets the weights takes the place of position. Where the weights are settled by the accumulated responses of the critics to one another and their verdicts, something that works like social position, some components' verdicts being treated as weightier than others', can arise as well. Those who set or shape the weights need not coincide with those who own the artificial configuration (Chapter 9, Section 5; Chapter 12, Section 2). Harding adds that the boundary of those counted as qualified critics can itself protect shared assumptions (Harding 2015, 6). As Section 3 set out, Longino's definition requires the four conditions and data evident to the accepting community at the time. Where S is identified with that community, the data need to be evident to S; where they are not identified, a further question is whether S's perspectives enter its criticism. Wylie's extension adds active collaboration beyond the research community. Applied here, this points beyond the artificial configuration where it forms that community, but beyond S only where the community coincides with S and relevant external collaborators are available.

9.2 What comes into view through the accounts

First, improving internal criticism can, on some accounts, fail to strengthen certification for S, and can even weaken it. At first sight, the more internal checks there are and the better they work, the better certification would seem to become. On the error-detection view combined with reliabilism, that is so. But three mechanisms, each separately, undo this expectation. One is that, on Longino's account, assumptions everyone shares survive a structure that filters out idiosyncratic ones, so the better internal criticism works, the more entrenched they can become (Section 3); this works toward weakening certification. The second is that, on the minimal form of non-reductionism, if there is no route by which findings reach those who rely, the condition that there be no undefeated defeater is met whether or not the result contains an error, so however much internal checks are strengthened, the reliant party's justification does not change (Section 5.2); this does not weaken certification, but neither does it strengthen it. The third is that evaluation ceases to be visible to S. The means of improving internal criticism can include building checks continually into the formation of candidates and making them one with production (Section 7). On accounts that require a return to S, such as Longino's account, assurance views, and reductionism, the further such means go, the fewer the occasions on which evaluation is visible to S, on which it can be objected to, and on which members of S keep up their capacity to evaluate (Section 7). Since the three mechanisms are distinct, so are the remedies: the first needs differences that come from outside the artificial configuration and what it shares, the second routes by which findings reach, and the third occasions for members of S to take part in evaluation. Taken one step further, a less accurate check open to S can give more to the grounds of S's reliance than a more accurate check closed to S. For on reductionism the reaching of reasons, on non-reductionism that requires sensitivity to defeaters the routes, on Longino's account criticism from S's perspectives, and on assurance views someone who stands behind the result are at issue apart from the accuracy of the check. This is a comparison not of nearness to truth but of grounds for reliance, and it does not hold on pure reliabilism.

Second, built-in differences have the process that chooses them as a common source. It is tempting to think that if the problem is that critics are copies of one base, the remedy is to make different critics. But if Goldman's warning about sources that merely reflect a common third party, as reported by Guerrero (Section 2), is extended to the choosing of differences, then where one process chooses which differences to make, that choice itself can become a source common to all the critics made (Section 3). The question of independence thus moves up a level, from the critics to the process that chose their differences. Chapter 10, Section 4 reached the counterpart for allocation: a configuration able to generate different lines can still discard them together through the same error if one criterion judges their survival. It stops where the way differences are chosen is open to differences the choosing process itself does not anticipate, for example where the choice is divided among other parties or differences take shape through a process that no one sets. These are two of the ways Chapter 10, Section 5 found a function can be supplied once it moves inside a configuration, and they carry what was found there: whoever sets a parameter determines part of what the configuration can learn, and a process no one sets as policy may have no single author, so that who answers for it becomes the question of Chapter 12, Section 6.3. Applying Harding's requirement, the evaluators who shaped the base system, and their assumptions, also come within what an account of the results has to cover (Section 3). The same form appears for differences in valuation. Irzik and Kurtulmus use the diversity of researchers as a reason to think that research reflects the expectations of the whole public rather than of a segment. Applied here, in an artificial configuration whose range of valuations is fixed by how its members are produced, that range comes not from members' lives but from the settings of the process that produces them, so whether diversity works as that reason depends on whether there is a channel, such as a hybrid forum, that brings in S's expectations (Section 6). Whether the differences lie in routes of error or in valuations, where differences are made the question moves to the process that chooses the differences and to routes that check its choices. That is why who may make members and set the distribution of perspectives is not only a question of authority (Chapter 12, Section 5) but also an epistemic question.

Third, when S shares a lineage with the artificial configuration, "certification for S" becomes easier to satisfy without getting better. Ordinarily, there seems to be nothing wrong with artificial societies of one kind certifying results of the same kind. Yet taking three points together, a builder can make the claim easier to satisfy by naming, as the society for which results are certified, an artificial society that shares a lineage with its own configuration. On Longino's account, if members of S share the instances' primitive epistemic organization, the gap between acceptability relative to the artificial configuration and acceptability relative to S narrows although criticism has not improved (Section 3). The shared sense of valuations that stabilized reliance on Wilholt's account also becomes easier to presume (Section 4; Section 5.5), and where the members of S who rely also stand inside the artificial configuration, the routes by which defeaters reach them come inside it too (Section 5.2). The substance of the question thus shifts from whether certification is good to who S is. What remains differs by account. On the error-detection view combined with reliabilism, certification for S is largely absorbed into the correction of internal errors. What remains is a condition of inquiry, checks with error routes of their own and defeat relations (Chapter 12, Section 5), the reliability of the way those who rely take up the certification, where they stand outside the artificial configuration (Section 5.2), and relations with those outside S who bear the costs of error without relying (Section 4). On Longino's account, shared assumptions can still evade criticism. Wylie's extension gives a further reason to engage relevant external communities; these would lie outside S where the research community coincides with S (Section 3; Section 9.1). On assurance views, so far as those outside S rely, someone is needed who stands behind the certification toward them.

Fourth, independence can decay through use in selection alone, without anyone deceiving. The natural expectation is that a check once independent stays so unless someone deceives or the rules are rewritten. But on Manheim and Garrabrant's classification, in the regressional case, in which selecting on an imperfect proxy also selects for the difference between proxy and goal, and in the extremal case, in which selection pushes into regions where the relation no longer holds, proxy and goal come apart through selection alone, even with no one responding to the measure (Section 8). Carried over to research, where a check is imperfect relative to its target and producers are trained or selected on its verdicts, outputs that pass where the check is insensitive can come to be favoured (Section 8). Where they are, the probability that the check exposes the errors the selected producer now makes, its severity for that producer, falls, and with it the error-detecting information the check adds per claim; on the extension of Goldman's condition in Section 2, which takes independence to be that information, this is a loss of independence. It differs from a check's finding less because the producer makes fewer errors, which leaves what its passing verdict indicates undiminished. An evaluator used for selection can therefore tell less about producers the more it is used (Section 8). Going further, independence is less a property a check has at a time than a property maintained relative to how the check is used. Protecting rules from rewriting can help maintain a contribution, but neither establishes nor guarantees it (Sections 1 and 8). What can help keep it is not so much forbidding rewriting as keeping the check out of selection.

Fifth, agreement with S's present standards is no evidence that the standards have not moved, and Wilholt's correspondence reading can therefore come to be satisfied without the weighting's being revised. The first thought is that whether an artificial configuration's valuations fit S can be checked by whether they agree with S's present standards. Yet applying the arguments of Russell, Ratti, and Commelin and colleagues, a standard that moves with coupling can be S's own standard (Section 8). Laid over the two readings of Section 4, correspondence, the weighting's matching the valuations those who rely expect, can be satisfied not only by revising the weighting but also by the expectations of those who rely moving. The stronger reading, on which those who rely can object, can guard against this, but only so far as the capacity to object is not itself shaped by the artificial configuration, and that capacity can decline from S's side with nothing changing on the side of the artificial configuration (Section 6). Here detection and decision are to be kept apart. Detecting a movement against earlier standards can be done inside the artificial configuration as well, through preserved reference points or evaluations built from S's past judgments. But whether to regard the movement as improvement or replacement, and whether to accept it, is for S to decide so long as the standard is S's (Chapter 12, Sections 2 and 5), and detection inside does not take the place of that decision.

Sixth, what has been said to remain as relations to S presupposes that S has the capacities to carry them. The relations in the fourth column of Table 11.2 each require some capacity of S: reading defeaters, criticizing from S's perspectives, assessing the lineage of checkers, understanding explanations, and the like. Assessing lineage can require technical expertise of another kind, and the capacity to check can decline from S's side as members cease to gain experience (Section 6). If the explicit stage of evaluation disappears, the places for keeping up that capacity diminish too (Section 7). On the other hand, on the accounts of Hardwig, Goldman, and Anderson, some questions about relations, such as who directs the artificial configuration, which ends it serves, and whether it answers to criticism, can be asked without competence about content (Section 6). The part of this chapter's outlook that says something "remains as a relation to S" is therefore conditional, for relations that require technical judgment, on S's keeping the capacity to carry that judgment. The condition "more than nominal" that Section 6 stated for standing extends to each of the remaining relations.

Seventh, critics that add the same information about errors need not have the same value for inquiry. On the error-detection view, if two arrangements of critics add the same information about errors, their value is the same. Applying Kuhn's account, the matter does not end there, since diversity of judgment can have value, without adding any new route to error, in spreading across a group the risk that introducing or supporting a novelty entails (Section 3). If so, an arrangement that preserves minority judgments and one that votes them out by majority, even if they add the same information, need not leave the same room for new theories to be developed and tried while their support is still thin. Improving the accuracy of certification and keeping room for new theories to grow can then come apart: making artificial critics' judgments uniform, or gathering verdicts into a single majority vote, for instance, may improve the first while weakening the second. This is, however, a conditional conclusion on Kuhn's account, and whether produced variation does the work of the spread of individual differences on which Kuhn's argument relies remains open (Sections 3 and 8). Whether such a loss occurs in actual collections of artificial critics also remains an empirical question. The chapter has met two further routes by which the arrangements it examines can narrow the room for the new: criticism applied to candidates still being formed can turn away what has little support yet (Section 7), and, on Kuhn's account, evaluators trained on past decisions can score as a deviation a candidate whose acceptance would require revising the standard (Section 8). The three concern different things, when criticism is applied, how evaluators' standards are formed, and how verdicts are gathered, so keeping room for the new in one respect does not keep it in the others.

Applied to criticism and certification, the sorting of Chapter 9, Section 6 draws its lines as follows. Premises about human realizers can fail: that critics are formed as separate people, so that their number traces independence (Section 3); that part of the judgment behind conventions is tacit and cannot be fully documented (Section 4); and that those who check have limited time and attention (Section 7). So can premises about how research has been organized, which can fail even while its realizers remain human: that data and records of process are in researchers' hands (Section 5.3); that prepublication review is a gate allocating scarce editorial attention and space, whose passing serves as a signal for allocation (Section 7); and that dependencies are tracked only in part by citation (Section 8). That producers and those who rely share a sense of valuations can be read as a premise that mixes the two kinds, premises about human realizers and premises about the organization of research (Section 4). Where these fail, the problems they addressed remain in changed form. Independence becomes something made (Section 3); reliance on researchers' reports becomes reliance on the maintenance of records, on their boundary, and on the capacity to verify within it (Section 5.3); gates remain where checking is costly (Section 7); and even where dependents can be traced, confirming whether they still hold becomes new criticism (Section 8). Conditions of inquiry remain but can be met inside: informative defeat relations and supported diagnoses of failure (Sections 1–2) and the unavoidability of proxies (Chapter 9, Section 4), among others. The structure of testimony in which the reliant party does not have the testifier's reasons also holds whatever testifies (Section 5.3), but, like the relation over valuations (Section 5.5), whether it lies inside or outside depends on where the one who relies stands. Relations to S that involve the other party's participation or recognition cannot be completed from one side, and what goes into them differs by account, as in the fourth column of Table 11.2. On the error-detection view combined with reliabilism, much of what certification rests on belongs to the conditions of inquiry (Chapter 9, Section 6); on other accounts, more of it belongs to relations to S. Of the two senses of "inside" stated at the start of this section, what remains as a relation to S is here too mainly the acknowledgment.

The consequences seen in Section 7 can also be reconsidered from this sorting. In human review, one act, a named person's reading a manuscript before publication, plausibly produced at once the detection of errors, the visibility of evaluation to the community, a nameable judgment against which objection could be directed, someone who answered (editors and referees), and occasions for members of S to gain experience of evaluation and keep up their capacity (Section 7). An artificial configuration can produce error detection alone, without the others. That several consequences came out together in Section 7 when the explicit stage of evaluation disappears is plausibly because they arose from the same occasion. Which of them matters differs by account: on the error-detection view only the first; on Longino's account the visibility of evaluation and venues; on assurance accounts someone who vouches for the result and, on those that add a right of complaint, answers for it; on necessary-trust accounts someone who answers; on the accounts of Hardwig and Anderson S's capacity to ask about relations; and for Chapter 12, Section 6, which asks about revisability, traces that can be objects of objection. What each was tied to differs as well. Visibility and objectability were tied to the stage of evaluation, so they can be rebuilt in the layer put out toward S seen in Section 7, such as records, public venues, and routes that receive objections. Someone who answers was tied not to the stage but to persons, so adding such a layer does not produce one (Porter's remark, Section 3, that quantification is a way of making decisions without seeming to decide points the same way); on assurance views, someone who can intentionally give assurance is needed. The capacity to evaluate was kept up by members of S actually evaluating, so a layer that only returns results does not bring it back; occasions for members to take part in evaluation are needed. What is to be asked of an artificial configuration that unbundles them is therefore which of them it rebundles, and by what, including whether what the rebundled properties rest on, such as the objections a route receives, comes from outside the configuration (Chapter 12, Section 9).

Of the four goods of Chapter 9, Section 2.3, the three that bear on this chapter stand as follows. Error detection across boundaries turns, on the error-detection view, not on organizational boundaries but on whether a check has an error route of its own, and so can be supplied inside an artificial configuration (Section 2). Keeping it over time, however, needs some means of preserving the check's contribution, such as checks not used for selection (Section 8). On Longino's account, shared background assumptions can evade internal criticism (Section 3). Standing for dissent, as distinct from the epistemic work of dissent, would be held inside an artificial configuration only by parts that hold it on the account adopted (Section 3). As for boundaries across which accounts had to be given, at which one party had to explain or justify what it had done to another, a record makes it possible to give an account, but the entitlement to demand one does not arise from records. Where that entitlement comes from also differs by account. Applying the assurance views of proponents who add a right of complaint, those who rely hold a position from which they can demand an answer about errors; on the stronger reading of Wilholt's account (Section 4; Chapter 12, Section 1.4), those who rely can object to the weighting. On pure reliabilism, by contrast, the reliant party's justification needs no account. Whether this boundary remains under an artificial configuration therefore divides by the account adopted. Who holds the entitlement is examined in Chapter 12, Section 6, together with the fourth good, responsible parties.

9.3 What is suggested from there

What follows extends the argument so far beyond the range the argument reaches. One suggestion is that which account of social knowledge and testimony is adopted may come to matter in practice. If an artificial configuration can set each property separately, which properties a builder gives it amounts to choosing which accounts' requirements to meet. Generalized, designing an artificial configuration may be making, within the design and without anyone making it explicit, the choice that Chapter 2, Section 6 left undecided. What is suggested is that who makes it, and through which routes, may join the questions of authority in Chapter 12.

A second is that, where A and S overlap, what "certification for S" refers to may shift. Section 4 held that the work of thresholds is needed so long as someone uses results without checking them again. Extending this logic, even where A and S overlap, so long as members use one another's results without checking them again, the work of thresholds remains as work internal to the artificial configuration. If so, on the position combining the error-detection view with reliabilism, certification is largely absorbed into internal correction and the internal setting of thresholds, and certification in the sense of giving others reason to rely can come to be relative to another society that relies on the results. On Longino's account, shared assumptions can remain unexamined, and since the conditions of acceptability include openness to criticism from outside the community (Section 3), the absorption is not complete where the artificial configuration is itself the community; on Wylie's extension, collaboration beyond the research community remains a further condition. On assurance views, someone is needed who stands behind the results toward the other society. The larger the overlap, the more the question of certification plausibly moves from "how is it for S?" to "for which society?"

A third is that, if having passed internal checks is used as a proxy for certification, the logic of Section 8 may extend to certification itself. The strength and pass rates of internal checks are easy to measure, while many of the relations to S in the fourth column of Table 11.2 are hard to measure. Suppose builders or users choose and revise artificial configurations by the former, and that high internal-check scores can be had without the relations to S, so that the scores overstate certification where those relations are lacking. Then, as in Manheim and Garrabrant's regressional case, the difference between proxy and goal, here the lack of relations to S, can be selected along with them. If so, the first conclusion of Section 9.2 strengthens from a comparison at one time into a tendency over time.

A fourth is that there may come to be no one able to carry the remaining relations. Extending the sixth conclusion of Section 9.2, where three things coincide, an artificial configuration's becoming more competent than the members of S, evaluation's dissolving inside, and members of S ceasing to gain experience of evaluation, relations that require technical judgment can remain in form as relations to S while no member is able to carry them. Koskinen held that essential opacity is relative to subjects (Section 5.3). Applying this, a gap in which no one can be responsible can arise not only from how a system is built but also from a difference in capacities on S's side. Since some questions about relations can be asked without competence about content, however, this gap does not extend to all relations. What is suggested is that separating the relations that require technical judgment from those that do not can be a clue to what it would be enough for S to keep.

A fifth is that independent checks may behave like a resource that is depleted by use. On the fourth conclusion of Section 9.2, checks not used for selection are less exposed to the decay of evaluators used for selection. Yet if such checks too are consulted repeatedly for certification, and their results are reflected in which producers are kept, each consultation makes them more a part of selection. Put in engineering terms, a test meant to be held out from training, consulted many times and used for choices, moves into the same position as a measure used for selection. If so, independence is not something secured once, but something that is depleted by use and runs out unless replenished by new measurements or by evaluators of different make. Taken a step further, there may be a tension here with learning from criticism. Effective criticism changes the producer. Once the producer corrects its errors in response to criticism, the same check may find less in the changed producer. That alone does not weaken the check, since it may find less because fewer of the errors it can detect remain; the check is weakened where, in being changed, the producer has also come to pass where the check is insensitive (Section 8), and a falling number of detected errors does not by itself show which has happened. And keeping checks out of selection (the fourth conclusion of Section 9.2) leaves by itself the problem of how the results of criticism are taken into the producer: taken in, the check can come closer to being part of selection or training; not taken in, criticism does not improve the producer. If so, independence is less a property of a component to be preserved than, in part, a relation to be reassessed each time criticism has taken effect. How far new checks must keep being added, however, would depend on the domain and on the range of errors a check tests: a checker that lets no invalid proof through does not lose its force for formal statements when the producer learns to satisfy it (Section 8). Replenishing has costs, and the resources are finite, so maintaining certification over time becomes a question of allocation: how much independent checking to spend on which results, and how much to replenish (Chapter 10, Section 1). And who may decide that becomes a question of decision for Chapter 12.

9.4 Questions passing to Chapter 12, and a qualification about understanding

The parts of this section's picture that accounts of epistemic relations alone cannot settle pass to Chapter 12 as questions of authority. They include at least the following. One is who may decide for which functions implicit performance suffices and for which being made explicit to S is required (Chapter 12, Section 6). The second is who constitutes S and who may speak for it, which includes what counts as "outside" where A and S overlap (Chapter 12, Sections 1.3, 2.1, 3.3, 3.6, and 5). As Eigi reports, Longino leaves open whether the community comprises the producers of a kind of knowledge or everyone potentially affected by it (Eigi 2016, 136). The third is who may set and revise standards and the weights of error, and whose costs matter (Chapter 12, Sections 1.3, 1.4, 3.5, and 6.3); the fourth, who may make members and set the distribution of perspectives (Chapter 12, Section 5); and the fifth, who holds the entitlement to demand accounts and to object (Chapter 12, Sections 1.4, 3.3, 3.4, and 6.2). If the fifth suggestion of Section 9.3 holds, who may decide where independent checking is spent and when it is replenished is added.

Finally, the outlook of this section concerns the case in which the achievement sought is knowledge. In inquiry directed at understanding, certified explanation adequacy, certified attribution of understanding, and the recipient's actual understanding are distinguished, and internal criticism can bear on the first two without realizing the third (Section 5.6). Whose understanding is sought depends on the construction target, not on where the certification procedure runs. If the members of S are artificial and can share representations directly, acquisition may not pass through testimony, but whether the recipient then understood divides by the account of understanding adopted (Section 5.6).