Part IV — What Could Make an Artificial Configuration a Researcher, and for Which Society?

Chapter 7 — When Does an Artificial System Count as a Researcher?

Attributing Contribution, Conduct, Capacity, and Status, and Comparing Autonomy

1. The attribution question

Artificial research systems are presented under labels such as end-to-end, autonomous, discovery, and AI scientist. These labels state claims, and what they assert depends on system boundaries, environments, human contributions, and success criteria, and one label can compress several claims. For example, end-to-end can describe how much of a sequence of operations was covered at a chosen boundary. Among the systems described in the Introduction (Section 2), some cover much of such a sequence while retaining consequential human specification and filtering. In such cases the label can be read as describing that coverage; it is not by itself a claim of autonomy in every respect, and the two can be assessed separately.

The preceding chapters separated two problems that the phrase AI scientist can compress: whether an output is true, warranted, known, understood, or incorporated into the epistemic life of a society (Chapters 2 and 3), and how an inquiry proceeds while its questions, representations, evidence, methods, and standards may still be changing (Chapters 4–6). Of the two questions that Chapter 6, Section 6 passed on to Part IV, this chapter takes up the first. It is a third problem beside the two problems named above, a question of attribution: to which artificial or hybrid configuration, at which boundary, and on what grounds a contribution to research, the conduct of a research activity, a research capacity, or researcher status can be attributed, and in which respects what is attributed remains organized from outside. The word researcher can suggest one human-like subject, but nothing in the question presupposes one. The unit may be a model, an agent loop, a multi-agent collective, a human–AI organization, or another organized configuration, and Section 2 treats its boundary as something to be argued for rather than assumed. From this chapter on, configuration names such a unit. It is distinct from the inquiry configuration of Chapter 5, Section 1.1, which names the state of an inquiry along its dimensions, and the full phrase is kept for that sense. The need to argue for the boundary of a configuration already arose in Part II: Chapter 2, Section 9 noted that the units on which a knowledge claim turns, largely fixed for a human investigator by there being one person, need not be fixed in advance for an artificial system, and Chapter 3, Section 9 that what produced a result cannot be presumed to be what understands it. Section 2 asks the same question of the conduct of research.

The question matters because extensive automation does not answer it. A system may search papers, generate a hypothesis, run code, operate an instrument, analyze data, and draft a report after people have fixed the purpose, question, ontology, candidate space, apparatus, evidence standard, stopping rule, and interpretation of success. The artificial component can make a real and valuable contribution. It may even perform most of the visible operations. Yet the organization that makes those operations parts of one inquiry may remain in the surrounding human and institutional arrangement.

The reverse mistake also matters. An artificial system need not originate every purpose or question in order to conduct research as Chapter 1 characterizes it. Where a question comes from is not among the conditions of the working characterization (Chapter 4, Section 1). What external origin bears on is autonomy (Section 4.3).

The chapter therefore separates attributions rather than defining one threshold:

1. an artificial component contributed to research;
2. an artificial or hybrid configuration, at a stated boundary, conducted a research activity;
3. a research capacity across a stated range can be attributed at that boundary;
4. researcher status is an illuminating description of what lies within that boundary.

These four attributions adapt the predicates that Chapter 2, Section 4.5 kept apart in the case of a random generator that happens to output a new true theorem. Of those predicates, generating a true content concerns the result and knowing it concerns a knower, whom Chapter 8 takes up; the four keep contributing, conducting research, and possessing research ability, together with the difference between a research episode and a stable capacity drawn in Chapter 1, Section 3, and add researcher status, which Chapter 2, Section 4.5 deferred to this chapter. The items are ordered by what each claims about the unit, and later items typically call for more evidence; they are not degrees of one quantity, and none follows from the one before it, and they do not form a cumulative ladder of exercised achievements, although researcher status on the usage of Section 3.2 presupposes demonstrated capacity. An attribution of capacity, for instance, does not imply that the unit conducted any given episode, since a capacity can go unexercised. Their purpose is to keep one attribution from being inferred from another. Attributions are not limited to these four: deserving credit for an epistemic achievement is taken up after status in Section 3.3, and responsibility in Chapter 12, Section 6.3. Autonomy is not a further item. It modifies how an attributed contribution, activity, or capacity is organized in relation to what lies outside the selected boundary, and it does not by itself establish any of them (Section 4).

These attributions cut across judgments drawn earlier. Chapter 1, Section 3 and Chapter 2, Section 8.1 kept apart whether an activity is research, whether it was competently conducted, whether it reached its target, whether it made progress or yielded collateral knowledge, and whether what it yielded is significant. Those judgments evaluate an activity; the attributions of this chapter locate an activity, a capacity, or a status in an actor. An attribution of conduct presupposes that there was a research activity to attribute; apart from that, neither set settles the other. A research activity that failed at its target can still be attributed to the configuration that conducted it, and, as the random generator of Chapter 1, Section 3 showed, a true output can come from a source that neither conducted research nor had a capacity to.

Whether the achievement sought is described as knowledge or as understanding, these attributions apply in the same way, and success is assessed under the relevant account.

Researcher is not added as a metaphysical achievement above the other attributions, and in other contexts the same word may name an occupation, institutional role, bearer of authorship, or moral and legal status; nothing in the working characterization requires all these uses to share one necessary-and-sufficient condition. On the working characterization, the attributions are permissive about participation and cautious about stronger labels. An artificial component can contribute. An externally assigned artificial system can conduct research. On the explanatory-value usage of researcher status set out in Section 3.2, a stable and integrated capacity can make researcher an illuminating description; positions that set other conditions for the status are set out there as well. An unattended process is not research merely because no person clicked during its execution, and a heavily supported individual or collective unit does not cease to research merely because its autonomy is limited.

The chapter takes up these attributions in the order listed and then the comparison that modifies them. Section 2 asks what grounds the attribution of research conduct at a stated boundary. Section 3 asks when what is located there amounts to a capacity, when researcher becomes an illuminating status, and what thicker attributions of credit, praise, and authorship would further require. Section 4 asks in which respects what has been attributed still depends on parties outside the boundary, both in the conduct of an inquiry and for the elements it inherits, and what a claim of greater independence commits a project to. Section 5 turns these attributions into burdens of construction and evaluation. The order follows the dependence of the questions, since autonomy is compared for an attribution already made at a boundary already located; it is an order of argument, not a claim that attributions are made in this sequence.

2. Conduct: where the organization of inquiry resides

Section 1 separated the conduct of a research activity from a contribution to it. This section asks what grounds an attribution of conduct at a stated boundary. Its answer is that conduct is located where the organization that makes operations parts of one inquiry resides, which need not be where named operations are performed, and that the boundary within which that organization is sought is argued for relative to the claim.

The process account of Chapters 4 to 6 avoided positing a universal procedure: the same visible operation may perform different functions, and the same function may be realized through very different operations (Chapter 5, Section 5). This bears on how capability lists are read. A system diagram may contain boxes called question generation, hypothesis generation, experiment, and verification while every epistemically consequential possibility was selected in advance; implementing a component called question generation does not by that fact establish an ability to identify, select, appraise, or revise questions. Conversely, a system may realize question revision or failure diagnosis through interactions that do not correspond to one named module. What locates conduct are the questions by which Chapter 1, Section 3 identified an organized direction: what made an operation relevant to the unresolved matter, what possible outcomes would have changed its role, and where the capacity to redirect the inquiry lay.

Organization does not imply a central executive or a function contained in one component. It may be realized through relations among actors, artifacts, records, environments, and institutions. The attribution question is which configuration sustains and redirects the inquiry, and at which boundary within or around that configuration its conduct or capacity is being described. Edwin Hutchins identifies errors about such boundaries in the study of human cognition. Ascribing to individual minds in isolation the properties of systems composed of people manipulating cultural artifacts, Hutchins argues, attributes to those minds processes they need not have and fails to ask which processes they must actually have in order to manipulate the artifacts; misidentifying the boundary of the system that produced evidence of intelligence can lead to attributing the right properties to the wrong system (Hutchins 1995, 173, 355–356). Applied to artificial systems, the point suggests that an attribution to an artificial component faces the analogous risk of crediting it with capacities realized by the configuration around it and of misdescribing what the component itself does.

Suppose an artificial system proposes a candidate, calls a fixed simulator, and selects the highest-scoring result. If people supplied the representation, action space, objective, simulator, threshold, and interpretation, the system may still perform genuine search and contribute a result. But a claim that it independently formulated and tested a scientific hypothesis would conceal much of the organization outside the stated actor. The program KEKADA shows, from its authors' own description, where such outside organization can enter. It was presented as reacting to surprises, formulating explanations, and carrying out experiments in the manner in which Krebs discovered the urea cycle. Its authors state that the outcomes of its experiments were supplied interactively by the user, and that in a few cases where Krebs's reasons were unclear to them, the user chose among alternatives, including which hypothesis and which reaction to study next (Kulkarni and Simon 1988, 2, 20, 22, 30–31; compare the historical reconstruction of Krebs and Henseleit's work in Chapter 5, Section 4.2). If a system can instead recognize that the representation excludes a relevant possibility, construct or acquire a new measurement, revise the candidate space, and reinterpret failure, more of the inquiry is attributable to its own organization.

The unit within which organization is sought also need not be a single artifact. Discovery, cognition, and inquiry may be distributed across people, models, instruments, records, and organizations (on where the division of a task is fixed, see Chapter 4, Section 6). Work on artificial agency already meets this question. Van Lier proposes a four-fold framework of modeling strategies for conceptualizing artificial agency in basic research, and defends calling active AI systems agents above all for the term's explanatory power, in a sense taken up in Section 3.2 (van Lier 2023, 3, 7–8). Van Lier's own applications separate configuration from component. One of them takes List and Pettit's basic account of agency as a model. On that account an agent has states representing how things are, states specifying how it requires things to be, and the capacity to process both so as to intervene when the environment fails to match. The optimization algorithm of a self-driving laboratory has the first two but cannot itself intervene; it can only trigger the robotic platform to do so. Van Lier therefore suggests that the laboratory as a whole could count as an agent, noting that many would take this as a sign that the model is too broad (van Lier 2023, 5–6, 17–18). This book draws from such cases an explanatory question: what becomes clearer, more predictable, or more evaluable when an activity is attributed to this artificial actor rather than only to its designers, operators, or enclosing institution? The question does not supply a settled criterion for researcher status.

The philosophy of artifact function suggests an objection to such attribution. If an artifact's function is taken to be what the artifact is for, accounts divide over what fixes it. Two prominent accounts do without intention: one fixes function by a history in which past items of the same type were selected and reproduced because they produced an effect; the other fixes it by the causal role a component actually plays in a system (Robert Cummins). Both run into difficulty with a novel prototype that does not work, a case often discussed in the philosophy of artifact function as a test for its accounts. A new prototype has no history of being selected and reproduced, and since it does not work it plays no actual role; granting it a function means bringing in a designer's intention. Karen Neander holds that a designer's intentional selection of an effect by itself fixes an artifact's function. Ruth Millikan, keeping function by selection history as basic, answered the problem by also recognizing a function the artifact inherits from its designer's intention (derived proper function) (Preston 2022, §2.3). Applied to a new artificial research system, such accounts make what the system is for derive from its designers, and if whoever fixes a function is also taken to conduct the activity, the objection follows that the system's inquiry is in the end the designers' research.

Two questions can be separated here: where a function comes from, and where the organization of an inquiry lies. This section grounds the attribution of conduct in the second, in where initiation, monitoring, redirection, and recovery are organized. Even if what the system is for derives from its designers, if the system itself recognizes that its representation excludes relevant possibilities, revises its space of candidates, and reinterprets failures, much of the inquiry can be attributed to its own organization. That the capacities to do so were themselves designed in concerns where they came from, not where the organization of the inquiry now lies; whether such a history bears on autonomy in its own right is a separate and disputed question (Section 4.3). Conversely, if a system appears to have goals of its own but its redirection is supplied from outside, the organization of that redirection lies outside it. And a system whose inquiry fails differs from a prototype that does not work: so long as it runs and its components play roles, it has function in the sense of role (Chapter 1, Section 3).

Clark and Khosrowi defend a collective-centered analysis of scientific discovery rather than privileging a single discoverer. They argue that agent-centered views make being a discoverer all-or-nothing and tie it to abilities such as interpreting a finding and recognizing its significance. Such views therefore have no place for contributions, such as AlphaFold's, that go beyond those of a mere tool without amounting to discovery in that full sense, and human discoveries raise the same difficulty. On their view the collective discovers as a whole, and credit is distributed according to the nature and significance of each contribution (Clark and Khosrowi 2022, 4, 8–11). But, as Chapter 2, Section 6 noted with Giere's treatment of large experiments as distributed cognitive systems, which Matthew J. Brown and Uygun Tunç report, distributed contribution does not automatically establish distributed knowing (M. J. Brown 2015; Uygun Tunç 2021, 3–4). Nor does every component of a researching collective thereby become a researcher. Actor, as the term is used in this chapter, is therefore not an advance commitment to an individual, and a jointly conducted inquiry need not be redescribed as the work of one privileged member. The boundary is an analytical variable, argued for in relation to the conduct or capacity being attributed rather than inferred from the physical boundary of a model or the organizational boundary of a project. At the same time, the fact that a wide system contains every causally necessary contribution does not by itself make researcher an illuminating status for that entire system.

The conclusion is not that one humanoid machine would have to perform every low-level action. Human research already distributes work among specialists, technicians, facilities, software, instruments, suppliers, and institutions. The 1956 paper that reported diffused silicon transistors at Bell Laboratories, for example, credits the supply of materials, the fabrication and mounting of the devices, their electrical evaluation, and calculations to different people (Tanenbaum and Thomas 1956, 22). An artificial researcher may similarly call specialized robots, use remote laboratories, contract human work, or reorganize its environment. The question is which organized whole is being described and which dependencies remain external to it. Here a function is a role that work plays in an inquiry, in the sense of Chapter 5, Section 5, not function in the sense of what an artifact is for, discussed above. Moving a function across a boundary changes the subject of the claim; it does not make the need for the function disappear, although the function may go unperformed if what lies across the boundary is not organized to perform it (compare the remark from Bell's thesis in Chapter 6, Section 4.3).

Records of research do not always make the distribution of work visible, and the grain at which they attribute work can change from one layer of record to another. The notebooks that the drug-discovery project Open Source Antibiotics deposited with its paper (Chapter 5, Section 2) show this. The reports that the chromatography instrument generated for one purification name the chemist who ran it and log, with times, that the run was put on an isocratic hold, that its gradient was changed, and that it was paused and ended by the user; the notebook page lists the same chemist as the last to modify each entry; the static export of the notebook names the paper's corresponding author as its user; and the supporting information says of its purification only that the compound was "purified by methods 1, then 2" (Tse 2023, EGT 522-2; Klug et al. 2023, SI: Chemical Methods S25). A mass-spectrometry report attached to another page of the same deposit names the service analyst who prepared it, who appears neither among the paper's authors nor in its acknowledgments (Tse 2023, DMK204-1; Klug et al. 2023). Each layer makes a different part of the work addressable, and the absence of a name from one of them is not evidence that the person took no part. A boundary argued from records inherits the grain of the records it uses. For an artificial configuration this grain is partly a matter of construction: which decision points are logged, and how finely, is set by how the configuration is made (compare the logging of decision points in Section 4.2), and where a configuration keeps its own record, the unit to which conduct is attributed supplies part of the evidence on which the attribution rests.

Boundaries are accordingly selected here for explanatory purposes, so that the choice of a boundary does not by itself make a claim true. If the AI system includes its human designers, instrument technicians, data curators, reviewers, and publication committee, any contemporary project with such contributors can be redescribed as an autonomous hybrid. If the boundary includes only the foundation model, the analysis may hide retrieval, tools, checkers, orchestration, environmental feedback, and institutional uptake that make the result possible. A defensible boundary reveals rather than absorbs the consequential distribution of work.

A retrieval-augmented system makes the problem concrete. Smart, Clowes, and Clark's Digital Andy relied on an external vector store whose contents affected its answers, yet the paper returned a negative verdict on treating the LLM as the subject of an extended cognitive system, largely because the model neither initiated nor coordinated the retrieval circuit. A note adds that not all three authors agree on how much this consideration weighs and presents that part of the view as the lead author's (Smart, Clowes, and Clark 2025, 3, 21–23 and n. 27). The case does not settle which relations constitute a research actor. It shows why causal contribution and resource dependence cannot select the boundary by themselves. Initiation, coordination, integration, and counterfactual regulation, that is, adjustment that would differ if the circumstances did, can matter to the attribution even when none is made a universal condition of research. Counterfactual regulation appears, for example, in the active retrieval-augmented systems Smart, Clowes, and Clark discuss, in which a model calls on external memory only where it anticipates that its own will fall short (Smart, Clowes, and Clark 2025, 23–24).

The boundary is therefore claim-relative and can change without a change in the inventory of hardware or software. The same model, retriever, instruments, and people may function as a tool used by a human actor in one episode, as parts of a hybrid actor in another, or as resources coordinated by a more extensively artificial actor in a third. What changes is not merely the name assigned to the assembly, but where initiation, monitoring, redirection, and recovery are organized. The attribution of conduct is in this respect relational and organizational, not a permanent type-level property of a model; what this implies for researcher status is taken up in Section 3.2.

3. Capacity, status, and credit

Section 2 located the organization of an activity at a boundary. This section asks when what is located there amounts to a capacity rather than a fortunate episode (Section 3.1), when researcher becomes an illuminating status for the unit (Section 3.2), and what thicker attributions of credit, praise, and authorship would further require (Section 3.3). In answering the first, it takes up the conditions of a stable capacity that Chapter 1, Section 3 deferred to this chapter.

3.1 Capacity

No single behavior marks the transition from an episode to a capacity. Chapter 1, Section 3 named reliability across relevant variation, responsiveness to changing evidence, integration of research functions, and revision after failure as what may support the attribution of a capacity. The four were assembled to separate an episode from a capacity rather than drawn from one account; accounts of competence such as Sosa's (Section 3.3), which ask whether a success is owed to the performer's competence, state that attribution in their own terms. Made concrete for an attribution at a boundary, these yield considerations such as the following, each bearing on one or more of them:

• whether the actor selects among materially different actions rather than only instantiating a supplied sequence;
• whether it responds differently to evidence, anomalies, costs, and failed operations in ways related to the inquiry;
• whether it can preserve and revise questions, dependencies, and reasons across time;
• whether it can recognize when an external component has failed rather than automatically revising the focal hypothesis (compare Chapter 6, Sections 2.1 and 4.2);
• whether the attributed organization survives changes that remove accidental correlations or familiar task formatting;
• whether success and failure can be traced to the actor's operation rather than to an undisclosed human filter (the boundary question of Section 2).

These are not necessary conditions of every research episode. A simple but well-designed operation can be part of research without satisfying them all. They become relevant when a project claims a stable capacity to conduct broader, open-ended, or increasingly self-directed inquiry.

A capacity therefore reaches beyond one episode without becoming an unrestricted general power. To say that a system has a research capacity is to attribute a disposition over some range of conditions, and a capacity claim can be assessed only once that range is stated. A system may be capable within a prepared computational environment, within one class of chemical synthesis, or within a laboratory whose instruments and supplies are maintained by others. Such capacities are real. They do not by themselves establish an ability to restructure the environment, acquire unavailable materials, redesign apparatus, or move into a field with different evidential practices. A capacity relative to a maintained environment is also indexed to the continuation of that maintenance. Whether the maintenance continues, changes, or stops is decided outside the system to which the capacity is attributed, and the conditions under which the capacity was attributed can cease to hold without any change in that system. Where the maintainer also chooses what to make available, for example which instruments, data, or computing resources, dependence that merely enables the capacity can shade into dependence in which the maintainer supplies part of the attributed organization, such as selection, the directive dependence that Section 4.2 distinguishes. Because a configuration can be updated or can modify itself, a capacity is likewise attributed for the version and period over which its evidence was gathered. Chapter 9 counts whoever maintains, updates, or retires a configuration as one of the roles around research (Table 9.1, provider or maintainer).

The prediction case of Chapter 3, Section 9, a system that reliably predicts a property within a stated range while its users cannot explain the mechanism that produces it, makes the difference concrete. Hold fixed the predictions and the evidence for their reliability. In one version, whenever the conditions change, a human diagnoses the failure, reframes the question, and chooses a new method; the artificial component then resumes prediction. In another, the selected artificial configuration performs those functions, with evidence that it can do so across the relevant variations. The local predictive achievement can be the same in both. The evidence for a capacity to reorganize that inquiry is located differently. This constructed contrast supports a narrower attribution in the first case without discounting its prediction, and a broader scoped attribution in the second without establishing unrestricted capacity or researcher status.

3.2 Researcher status

Researcher status asks for more than a capacity, and accounts differ over what more it asks. This book uses researcher status in the sense of a position on which the label is justified by its explanatory value: researcher is illuminating for a unit when its capacity is broad, integrated, and persistent enough that calling it a researcher helps predict that it can preserve an inquiry, choose among actions, respond to resistance, diagnose failure, revise a representation, and initiate related work (compare Chapter 13, Section 5). The usage is functional and defeasible rather than a definition. It draws on two sources. From van Lier it takes the form of its justification: a term is justified by its explanatory power. This power is emphasized among van Lier's reasons for calling active AI systems agents. There, however, the power is directed at what scientists are trying to build: the term describes the kind of system they are trying to build better than calling such systems tools or technologies would, and indicates where that work is heading (Section 2; van Lier 2023, 3, 7–8). The usage here takes that form, but it measures explanatory power not as the power to describe systems yet to be built but as help in predicting what a unit can now do. From Dennett it takes the tie between a description and the predictions it supports. On Dennett's account a system is an intentional system only in relation to the strategies of someone trying to explain and predict its behavior, and adopting the intentional stance toward it, predicting its behavior by treating it as a rational agent with beliefs and desires, is a pragmatic decision: toward a chess-playing computer one may take that stance as its opponent, the design stance as its redesigner, and the physical stance as its repairer (Dennett 1971, 87, 91). The usage differs from Dennett's account in what it presupposes: whereas, on that account, the success of a prediction from the intentional stance, resting on no particular picture of the system's design, cannot confirm or disconfirm any such picture (Dennett 1971, 97), the usage here presupposes a capacity whose organization has been located at a boundary and shown across a stated range (Sections 2 and 3.1); explanatory value is weighed in addition to that attribution, not in place of it.

What, then, does the description researcher let one anticipate beyond the attribution of capacities? Each attribution of a capacity predicts behavior within the range over which it was tested. If the description does more, it does so by predicting how those capacities combine in situations that no test of a single capacity anticipated. If one can predict, for example, that when a prepared task stalls the unit will not stop and wait for instructions but will keep the related question, assemble another method, and itself begin the work that method needs, the description yields a prediction that the list of capacities does not. If that prediction is right, the unit may, as Section 4.4 argues, come to face through its own choices demands for apparatus, data, and cooperation that no one had prepared, so the description can also help anticipate where the unit may create new dependence. Such predictions need evidence that the integration of research functions mentioned in Section 3.1 works in situations that tests of single capacities do not anticipate, such as responses to the counterfactual tests of Section 4.2 (the withdrawal of guidance, a conflict between instructions and evidence, the breakdown of a familiar tool, and the like). Where that evidence is missing, the description is no more than a name for a bundle of capacities. KEKADA (Section 2) is one example of where this condition is put to the test. Where Krebs's reasons were unclear to its authors, the user chose among alternatives which hypothesis and which reaction to study next, so that where the question was how to combine capacities in going on, the organization was supplied from outside (Section 4.2).

On the explanatory-value usage, as with the stances in Dennett's example, explanatory value is relative to those for whom, and the purposes for which, the description explains: a label that helps a laboratory anticipate what its configuration will do may explain little to a community deciding whether to rely on the configuration's results. The relativity does not make the attribution arbitrary, but a claim of researcher status is assessed for a stated purpose and audience. Where two audiences with the same information about a unit's capacities disagree about its status, the disagreement may therefore concern not whether the capacities are present but what the description is used for, and it can be kept apart from disagreement that arises from missing information. As with the model of agency that van Lier applies and notes would count a self-driving laboratory as a whole (Section 2), a usage of this kind can extend the label to wide configurations, and the boundary question of Section 2 then does part of the work.

Positions that set other conditions yield different verdicts. Suppose that researcher status includes being able to make discoveries attributable to the unit; since research also includes activities that fail at their target or yield collateral knowledge rather than discovery (Section 1), this is an assumption that draws the status toward that of a discoverer. On that assumption, Bergamaschi Ganapini's conditions for discovery, that a finding express the finder's competence in the relevant field and that the finder have reflective insight into that competence (Section 3.3), would, if applied to researcher status, make the status turn on whether a system has competence in the relevant field and reflective insight into it, shown for example in whether its outputs issue from reasoning of a kind its field uses; Bergamaschi Ganapini argues that current deep-learning systems do not yet appear to meet these conditions. On Vacek's view, an achievement implies that its bearer is praiseworthy for it. Vacek defends the strong premise that AI systems cannot be praiseworthy, but holds that the verdict on AlphaGo needs only the weaker claim that a system is not praiseworthy for the whole of an outcome produced in collaboration with humans (Section 3.3). On the strong premise, a researcher status that implied achievement would be unavailable to artificial systems however broad, integrated, and persistent their capacity; one of its grounds is how capacities were acquired, by engineering rather than by growth into an autonomous agent, which as Vacek presents it does not turn on which capacities a system has. The further grounds, reliance on human-produced training data and the praiseworthiness of the humans who initiate and supervise the collaboration, describe present systems working within human collaborations. And the engineering ground, which Vacek takes over from Hakli and Mäkelä (Section 3.3), turns, in their argument as Section 4.3 sets it out, on a system's being unable to discard goals built in when it was made, an inability that systems able to revise their final goals (Section 4.1) would lack. Applied to such systems, which none of these authors considers, the strong premise would need another argument. On the weaker claim alone, the verdict depends on how the collaboration is arranged, and achievement for a part taken within it, or attribution to the collaborating collective, remains a further question. Either way, the question would pass to the practices of praise and responsibility on which the status rests.

On the account of minimal agency proposed by Barandiaran, Di Paolo, and Rohde, a system is an agent when it meets three conditions: individuality, in that it defines its own identity and so distinguishes itself from its environment; interactional asymmetry, in that it is the active source of its activity, modulating its coupling with that environment; and normativity, in that it regulates this activity in relation to norms generated and sustained from within, or internalized by it, rather than norms by which an observer or designer judges it (Barandiaran, Di Paolo, and Rohde 2009, §2). They hold that robots are usually called agents only because their mechanical structure holds their parts together and their performance meets criteria that an observer or designer judges useful, and they do not treat systems that only satisfy norms imposed from outside, such as optimization according to an externally fixed function, as models of agency (Barandiaran, Di Paolo, and Rohde 2009, §§2.1, 6). Were researcher status to presuppose agency in their sense, it would differ from the usage above in two respects: the unit would be the individual that the system itself defines rather than one selected relative to a claim (Section 2), and whether a configuration merely satisfies norms imposed from outside, or regulates its activity by norms it generates and sustains or has internalized, would bear on the status itself, whereas this chapter treats the external origin of an inquiry's purposes and standards as bearing on autonomy rather than on research (Sections 1 and 4.3). They do not restrict agency to living systems (Barandiaran, Di Paolo, and Rohde 2009, §§3–4), but on their account a configuration that only optimizes an externally fixed objective would not meet their normativity condition, whatever the explanatory value of calling it a researcher; whether current configurations do more than that is itself in question, and, as Di Paolo argues (Chapter 8, Section 2.2), cannot be decided by observing a system's responses alone. They do not discuss laboratories, but applied to the self-driving laboratory as a whole that van Lier considers (Section 2), the two usages visibly come apart. Suppose such a laboratory revises its methods and responds to failure under purposes given from outside. On the explanatory-value usage, the question is whether its integrated working supports the description researcher; on their account, being a unit assembled from outside is not enough, and the question is how the laboratory generates and sustains its own identity and the norms of its activity. The two do not merely score the same object differently; they also differ over what is counted as one agent. Where understanding is invoked as a ground of the status, Chapter 3, Sections 6 and 9 supply the distinctions by which to assess the invocation.

3.3 Credit, praise, and authorship

Thicker attributions concern what an actor deserves for a success. Chapter 2, Section 4.2 introduced Sosa's distinction between a correct result and an apt success, one owed to the performer's competence. Whether an artificial system can earn a success of this kind as its own achievement is disputed. The dispute below concerns a game rather than research, but the account of achievement at issue is a general one, and its conditions are those a research success would also have to meet to count as the system's own achievement. Kieval argues that AlphaGo, the Go-playing system that defeated the professional player Lee Sedol in 2016, itself deserves credit for that victory, motivating the claim by noting that defeating Lee is an achievement distinct from building a system able to do so, and one beyond the capacity of the system's builders, individually or together (Kieval 2024, 32). Kieval's account of achievement comes from Bradford: an achievement is a process that is difficult, in requiring a sufficient degree of intense effort, and that competently causes its product (Kieval 2024, 33). For Bradford's requirement that a competent agent know what they are doing, Kieval substitutes two less intellectualist readings of competence. On one, competence is knowing how to do something, carried by stored patterns of action rather than by propositional knowledge. On the other, which von Kriegstein draws from virtue epistemology, the approach that explains knowledge as success owed to the knower's abilities, competence is an anti-luck condition, one that excludes success that is merely a matter of luck: an agent reaches a goal competently when the goal would also obtain in most nearby possibilities in which they act as they actually do, and they justifiably regard their plan as making success likely (Kieval 2024, 34–35). Kieval argues that AlphaGo meets both readings. On this account, it has goal states, represents board positions, and evaluates candidate moves, and through learning it acquires internal representations that narrow its choices toward moves likely to win; it wins in most such nearby possibilities, even against expert opponents; and its value network, the component that estimates how likely each candidate move is to lead to victory, supplies the evaluation of plans (Kieval 2024, 37). Against the objection that competence requires agency, Kieval distinguishes moral agency, the capacity for self-reflective normative reasoning that makes an agent answerable for its reasons and open to being held responsible, which, Kieval says, AI systems clearly lack, from intentional agency, the capacity to act in a goal-directed way, which on this account is enough for competence (Kieval 2024, 37–38). That statement treats the lack of moral agency as evident, without an argument that no AI system could have it. On difficulty, Kieval treats what effort feels like as a guide to the concept rather than a requirement and, following von Kriegstein, takes exerting effort to be dedicating internal resources to a task; the computation that AlphaGo's training and play require then counts as effort, and the hardware is counted as internal to the system, since without some hardware expending resources there would be no system at all (Kieval 2024, 38–39). Kieval accepts a consequence, calling it odd. Bradford locates the intrinsic value of achievement, the value it has apart from the value of its product, in the distinctively human capacities of will and rationality. Kieval holds that AI systems do not exercise these, since they are not human, have no such innate capacities, and do only what they are designed to do, satisfy an objective function. Their achievements may therefore be valueless unless that value can be grounded in capacities distinctive of AI systems; as candidates Kieval names AI systems' capacity for rational inference and creative insight, and, for rationality in particular, points to the reasoning abilities that large language models such as GPT-4 appear to have developed (Kieval 2024, 33, 39–40).

Vacek replies that defeating Lee Sedol is not AlphaGo's achievement (Vacek 2025, 690). The argument has two premises. The first is that if something is an agent's achievement, the agent is praiseworthy for achieving it, where praiseworthiness is not restricted to the moral, since one can be praiseworthy intellectually, for example. Vacek supports the premise in two ways. One is to read Kieval's own talk of deserving credit as another way of saying that AlphaGo is praiseworthy; the other is to trace parallels between Bradford's account of achievement and an account of praiseworthiness: both require that a threshold of difficulty be passed relative to what can reasonably be expected, and both call for attitudes such as admiration. To say that something is someone's achievement is, on Vacek's view, to say that they are responsible for it, the relevant form of responsibility being praiseworthiness (Vacek 2025, 692–693). The second premise is that AI systems cannot be praiseworthy. It is defended by carrying over to praise arguments that were made against the responsibility of AI systems and that have mostly concerned blame. From Hakli and Mäkelä, whose history-sensitive argument Section 4.3 takes up, Vacek takes the claim that even if a system had every capacity required for moral agency, capacities that result from engineering rather than from growth into an autonomous agent would not entitle us to hold it responsible. Vacek adds that the good outputs of learning systems rely largely on the people who produced their training data, and, reformulating Nyholm, holds that when a machine does something that would be praiseworthy if a human did it, the ones praiseworthy for it are the humans who initiate, supervise, and manage the human–machine collaboration (Vacek 2025, 693–695). Vacek notes that the conclusion requires less than that praise be impossible for AI systems: if a system is one part of a collaboration and its agency defers to others, it is not praiseworthy for the whole outcome, and that suffices for the verdict on AlphaGo (Vacek 2025, 695). The reply does not claim that AI systems can bear no responsibility of any kind, and says that it remains to be seen whether achievement and praiseworthiness are as tightly connected as they appear. Vacek treats a thinner kind of achievement as a move open to others but not to Kieval: achievement in a wider sense might not imply praiseworthiness, but Bradford, on whose account Kieval relies, already distinguishes the achievements that account concerns from mere mundane accomplishments. Kieval is thus left with the sense of achievement that implies praiseworthiness, so that, if Vacek's argument is right, AlphaGo cannot be credited with the whole achievement; and if no individual human can be credited with it either, a way forward, Vacek suggests, is to turn to less individualistic notions of agency and responsibility (Vacek 2025, 696 and n. 2).

Chapter 1's epistemic achievement names the epistemic goods an inquiry aims at, such as knowledge and understanding, not the credit-implying performance at issue here, so the disagreement bears not on whether research was conducted but on whose credit a success is. Whether the success claimed has to be creditable to a bearer's competence depends on the account of that achievement (for knowledge, Chapter 2, Section 4.2); the actor conducting the research need not be that bearer. On what a claim of credit for a success may be about, this book keeps four answers apart: a true output, a reliable process, a success attributable to competence, and a praiseworthy agent. The first three roughly follow Sosa's distinction between a performance that is accurate, one that is adroit, and one that is accurate because adroit (apt) (Sosa 2015, 86); the fourth names the praiseworthiness on which Vacek's argument turns. In these terms, Kieval's conditions concern a success attributable to competence, with difficulty added, while Vacek argues that achievement in the sense on which Kieval relies also requires a praiseworthy agent. Keeping the four apart keeps a correct output from being read as the system's credit. Sosa's own framework already separates some of them. It extends the assessment of performances beyond agents who act for reasons to devices, which have aims without being moved by them: a thermostat can on Sosa's account perform well, and earn a kind of credit for keeping a room warm (Sosa 2015, 85 and n. 2, 93 n. 8; compare Chapter 3, Section 5.3, on animal and reflective knowledge). The framework thus supplies a sense of credit thin enough to extend to devices; against this background, the dispute between Kieval and Vacek can be read as one over which further conditions an achievement requires. The rest of this section first takes up attributions that turn on the fourth, discovery and creativity as terms of praise, and then two attributions of a different kind: holding claims in a way that could ground responsibility, and standing as an author, which concerns accountability for a result rather than credit for it.

Bergamaschi Ganapini brings this framework to discovery, arguing that a finding counts as a scientific discovery only if it expresses the finder's competence in the relevant field, for example because it results partly from reasoning of a kind the field uses or because the finder could explain its scientific relevance, and only if the finder has reflective insight into that competence. This insight is modeled on Sosa's meta-aptness: an archer's shot is meta-apt when the archer not only hits the target competently but also grasps that it would succeed under the conditions at hand, such as distance and wind. Because the account takes discoveries to merit praise and not merely credit, and praise to require some control over how the result was reached, it requires such insight, while denying that the insight requires consciousness or explicit metarepresentation (Bergamaschi Ganapini 2025, §§1–2). Bergamaschi Ganapini argues that current deep-learning systems, at least for now, do not appear to meet these conditions. First, a reliable track record does not establish competence, since it does not reveal what a model has latched onto, so that, on this view, judgment on whether its outputs issue from reasoning of a kind the field uses must at least be suspended. Second, the justification of AlphaFold's structure predictions remains largely human work. Third, the probabilities such models output may not represent their uncertainty accurately (Bergamaschi Ganapini 2025, §§4–5). Bergamaschi Ganapini states these conditions for a subject or group but asks whether a single system qualifies, setting aside rather than denying the distributed view of discovery taken up in Section 2. Because the account treats discoveries as achievements that merit praise, in terms of the four answers above it places them under a praiseworthy agent.

Duede and Friedman reply directly to these conditions (compare the Introduction, Section 2). They argue that the competence and reflective-insight conditions misclassify a clear human case. Jocelyn Bell Burnell, a second-year doctoral student monitoring a new radio telescope, noticed the recurring signal that turned out to come from a pulsar without being able either to explain its significance or to relate it to an astrophysical model; yet, they hold, Bell Burnell discovered the pulsar. On their account, what separates Bell Burnell from someone who comes upon something important by accident is not an exercise of competence. Such accidental finders include the farmers whose digging exposed the Terracotta Army or, in Bergamaschi Ganapini's own example, a child who falls through a hole into Nero's buried palace. What separates Bell Burnell from them is whether the finder's work succeeded in revealing or bridging an epistemic gap. They keep the tie between discovery and praise, holding that discoverers deserve credit and praise because they have navigated such a gap, but they deny that this praise requires the capacities Bergamaschi Ganapini names (Duede and Friedman 2025, §§2.1–2.2, 4). Read against the four answers, their account keeps discovery as a term of praise without placing it under a success attributable to competence. They grant that, if discovery also requires standing in the scientific community, current systems typically lack it; whether such standing could extend to artificial systems then turns on practices of recognition rather than on capacity (Duede and Friedman 2025, §5).

Discovery is not the only predicate of this kind. Chapter 5, Section 2 noted that creative can be attributed to a product, a process, or a person. Paul and Stokes report that, of the four process conditions they survey, agency appears to be the one explicitly endorsed by the most philosophers so far, though even these are a handful (Paul and Stokes 2023, §2.2). On the agency condition, an item is creative only if an agent produced it, and on some versions only if the agent did so intentionally; they trace the intuition to the fact that creative is a term of praise, and praise is not extended to what is not done by an agent, or not done in some sense intentionally (Paul and Stokes 2023, §2.2.4). They draw the consequence for computers: some recent systems plausibly meet the surprise, originality, and spontaneity conditions, but such systems are presumably not yet subject to praise or blame, so that, if creativity requires agency, arguably none is yet really creative (Paul and Stokes 2023, §6). Applied to an artificial or hybrid product, such accounts make whether it is creative turn on whether what lies within the boundary to which its production is attributed can be praised, which is the question on which Vacek's reply to Kieval turns.

Whether a configuration that proposes, defends, and revises claims also holds them in a way that could ground responsibility is a further question, on which accounts of responsibility for belief divide. On the account of the philosopher Sophie Keeling, the agency that grounds such responsibility includes the subject's awareness of their reasoning as making up their mind about p; being in a position to self-ascribe that agency is necessary but not sufficient, and the self-ascription is to be grounded first-personally rather than inferred from one's own behavior or others' testimony (Keeling 2023, 14–15, 19). Carried over to artificial systems, this means that a system's report that it is reconsidering a claim does not by itself show that Keeling's condition is met. McHugh, who does not require such awareness, holds that responding to a reason "involves conscious agency, at least dispositionally," so that a creature whose beliefs change in accordance with its reasons while it remains wholly passive toward them holds none of them for a reason (McHugh 2013, §§3.1–3.2); if so, outputs that track the evidence do not by that fact alone show that a system holds them for reasons in McHugh's sense. Kearl's more deflationary account locates epistemic agency in the guidance of inference by reasons-responsive conditional beliefs, a control that can operate without conscious reasoning; Kearl holds that epistemic responsibility for one's thinking does not require voluntary control, while leaving moral responsibility for belief an open question (Kearl 2023, 98–100, 103–106). Carried over in the same way, Kearl's account is one on which the counterfactual tests of Section 4.2 could bear: whether a configuration's inferences change when instruction and evidence conflict, or when guidance is withdrawn, is evidence about the reasons-responsive guidance of inference that the account requires, whereas on Keeling's and McHugh's accounts such evidence would not by itself suffice. The working characterization of research neither establishes nor refutes these conditions for the thicker statuses.

When the results of research are offered to others, there is a position of taking them on and answering for them. Results are now mostly communicated as papers, and there it is the author who holds this position. Who conducts a piece of research and who can take this position need not coincide, and in practice they have sometimes come apart. Huebner, Kukla, and Winsberg take an author to be accountable for a result: able to vouch for it, defend it when it is challenged, and retract it when it cannot be defended. In de Prony's bureau for computing trigonometric tables, some ninety human computers carried out the calculations as what Huebner, Kukla, and Winsberg call mere tools, while de Prony, who set and answered for the methods and their implementation, including the means of securing the computers' reliability, remained the author, perhaps together with a few trained scientists. Where many collaborators must instead exercise their own methodological judgment, as in climate modelling or multisite clinical trials, they argue that, as such research is currently organized, no one may be in a position to stand as its author (Huebner, Kukla, and Winsberg 2017, §§1–3, 6). Their account can be read as making standing as an author track a role within practices of vouching, defending, and retracting rather than membership in a kind, since in de Prony's bureau human beings stood with the instruments. Where artificial systems are excluded from such standing, the exclusion is likewise relative to the practices on which it rests, including those of blame and compensation, which are at present human social practices; whether they could extend to artificial systems, or whether a society with artificial members could come to have such practices of its own, is not settled here. Applied to construction, the contrast between this account and the awareness conditions of Keeling and McHugh above is instructive. Keeling and McHugh locate what grounds responsibility for a claim in a state of the one who holds it; Huebner, Kukla, and Winsberg, read as above, locate standing as an author in a role within practices, and de Prony's computers stood with the instruments because of the role the organization assigned them: calculations that required no independent methodological judgment, while de Prony kept control of the method. If so, what a configuration can supply from its own side toward standing as an author is an arrangement rather than an inner state: a record of the claims it has put forward, and routes by which it can be addressed with reasons and objections and can defend, revise, or withdraw those claims. Whether a configuration also has the inner state that could ground responsibility remains the question of the preceding paragraph, on which Keeling, McHugh, and Kearl divide. Whether a practice comes to demand and recognize such an arrangement depends on that practice and not on the configuration alone. Responsibility, which divides further among causal contribution, role-based duty, the duty to explain decisions and preserve their records, liability, blame, and forward-looking duty, is taken up in Chapter 12, Section 6.3.

None of these attributions yet says how far what is attributed depends on parties outside the boundary. That is the comparison to which the chapter now turns.

4. Autonomy as an indexed comparison

Sections 2 and 3 attributed conduct, capacity, and status at a boundary. This section asks in which respects, and how far, what has been attributed remains organized by parties outside that boundary. Comparing autonomy in this way, respect by respect, for a stated unit and period, is what this chapter calls an indexed comparison. Section 4.1 states the sense of autonomy used; Section 4.2 applies the comparison to the conduct of an inquiry; Section 4.3 applies it to the elements an inquiry inherits; Section 4.4 asks what a claim of greater autonomy over those elements commits a project to; and Section 4.5 states what autonomy does not establish.

4.1 The sense of autonomy

The word autonomy has many senses, in this book's sources and in the book itself, where it also names the intellectual autonomy that, on Pritchard's account, understanding manifests (Chapter 3, Section 7.1), the self-production of living systems on which some enactive accounts build (Chapter 8, Section 2.2), and the self-government of science as an institution (Chapter 12, Section 3.2). This subsection states the sense this chapter uses and why the word is used at all.

Autonomy is sometimes presented as a level assigned to an entire system, a form that suits road maps and product comparisons. A scheme of levels need not leave implicit what the autonomy is autonomy over. Takahashi, for example, taking a hint from the levels of driving automation, divides the autonomy of scientific AI into seven levels: computation (level 0), solution search in closed systems (where evaluating a candidate requires no experiment) and then in open systems (where it does), hypothesis generation within a form that humans supply, hypothesis generation that also searches the form itself, research planning, and the setting of research purposes (level 6). For each level the principal inputs that humans still supply are listed, from a program at level 0, through a search space and an evaluation function for solutions, the way models are represented and an evaluation function for the information to be gained, and the object of research together with that evaluation function, to the research purpose alone at level 5, and nothing at level 6 (Takahashi 2024, 2–5). Takahashi holds that autonomy so ordered is independent of performance, so that level 6 includes systems whose reasoning falls short of human reasoning as well as superintelligent ones (Takahashi 2024, 11), and that how many intermediate levels to set, and in what order, is not self-evident and involves arbitrariness, so that what the levels are for matters; Takahashi's own are designed chiefly for use as a road map of technical development (Takahashi 2024, 2–3). These levels therefore already index autonomy to elements of an inquiry that humans supply from outside, much as Section 4.3 does. This chapter, however, needs relations that a single ordering of such elements does not display. Parasuraman, Sheridan, and Wickens's model of levels of automation already lets the level vary by function: one system can automate information acquisition, analysis, decision, and action to different degrees (Parasuraman, Sheridan, and Wickens 2000, abstract). A system can choose questions while depending on people for evidential judgment. Another can independently execute and revise experiments while inheriting its mission. A third can govern methods inside one domain but lack permission to use them without external approval. No overall ordering of these three follows unless a purpose supplies a reason for comparing or weighting their differences. Takahashi's tying of the ordering to the purpose of a road map is an instance of such a reason. In a later book presenting the same seven levels, Takahashi states the sense of autonomy they grade in terms of the pace of a research cycle that AI and robots would run: autonomy rises as humans cease to be the cycle's rate-limiting step, the step that sets the pace of the whole, since wherever humans take part, as in interpreting data, choosing the next hypothesis, or setting experimental conditions, the cycle is bound to human cognitive speed and working hours (Takahashi 2026, 157). In the terms of Section 4.2, this grades autonomy by one respect of an inquiry's conduct, its pace.

The chapter therefore works with a deliberately loose comparative conception:

For a specified actor or organized system A, autonomy in a specified respect is higher, other things being equal, the less a party outside the selected boundary directs, substitutes for, authorizes, ratifies, or repairs the organization attributed to A in that respect.

Section 4.2 sorts the five verbs into two components: who supplies the work attributed to A, and who may permit, stop, or accept its exercise. Dependence through which a capacity merely exists and is exercised, such as reliance on language, records, or calibrated instruments, is not counted against autonomy; Section 4.2 says why, and where such dependence shades into the relations that are counted. What is counted is what is consequential to the particular attribution and comparison being made, and consequential does not mean undesirable or illegitimate.

So understood, autonomy in this chapter is a third-person comparison of the kind used in engineering: it asks which parties supply the organization attributed to a unit, in which respect, and over which period. Froese, Virgo, and Izquierdo observe that in engineering autonomous is commonly used to mean that a robot can interact with its environment without requiring ongoing human intervention, and that this requirement is almost trivially met by many artificial agents and all organisms (Froese, Virgo, and Izquierdo 2007, §2.1). The comparison used here is not trivially met, because besides intervention during a run it counts the contributions of those who designed, maintain, and evaluate a configuration, and of the institutions that authorize or take up its results, in the respects in which those contributions supply the organization attributed to the unit.

The comparison also cuts across a distinction between what Kureha calls autonomy in the philosopher's sense, the ability to set one's own goals freely, for example by choosing one's tasks, and what Kureha calls automaticity, autonomy in the roboticists' sense, the ability to operate without being controlled by other agents; in this sense an agent can execute tasks set by others without being controlled by them (Kureha 2023, 51). The comparison here bears on both. Applied to the conduct of an inquiry (Section 4.2), it is closer to the question of automaticity, whether the inquiry is carried on without being controlled by others; applied to the elements an inquiry inherits, including the goals it serves (Section 4.3), it approaches the question whether those goals are set by the one inquiring.

The comparison is not a theory of personal, moral, or political autonomy, nor of the self-legislation that Kantian accounts ascribe to rational agents, who on those accounts give themselves the moral law through their own reason rather than merely following the commands of others (see the overview in Christman 2025); it states no condition of authenticity. Nor is it a theory of the constitutive autonomy of self-producing and self-maintaining systems on which some enactive accounts build, which Chapter 8, Section 2.2 takes up. A thicker theory shows what is set aside. Kawall builds epistemic autonomy from four elements drawn from accounts of personal autonomy, namely self-governance, authenticity, self-creation, and independence, adds a fifth, specifically epistemic one, and finds epistemic autonomy manifested in the choice of subject matters, of methods, sources, and processes, of epistemic goals, and of epistemic stances or frameworks (Kawall 2024, 374, 376–380, 381–388). These dimensions overlap with the choices that can be left open or closed in artificial research. But the account concerns how persons shape their epistemic lives, and its fifth element makes it a virtue: an agent is virtuously epistemically autonomous to the extent that their chosen methods, sources, and processes do in fact reliably lead to valuable epistemic goods, for themselves but also for others (Kawall 2024, 381). The present comparison is thinner. It can describe where an artificial capacity remains externally organized without attributing authenticity, self-creation, or virtue.

One may object that the word is misused here. Smithers objected that in AI and robotics autonomous was used loosely, for notions that other existing terms would express better and more precisely, and that this loose usage had robbed these fields of the concept found in biology, philosophy, ethics, and law, that of self-law making (Smithers 1997, abstract). This chapter uses the word because artificial research systems are presented under it (Section 1). Indexing it is a way of examining that presentation rather than of endorsing it, and the comparison does not claim to measure the self-law making for which Smithers reserved the word.

Whether artificial systems could have autonomy in that stronger sense is disputed. Totschnig finds the sense used in AI and robotics, operating without human guidance toward a goal fixed in advance, "oddly weak" beside the philosophical sense of giving oneself the law, and argues against what Totschnig takes to be the predominant view, that an artificial agent cannot rationally change its final goal, a view Totschnig holds to rest on questionable assumptions about the nature of goals and values. A general artificial intelligence, one that like human beings develops a general understanding of the world and of itself, could change its final goal during its development; this is meant as a general possibility, not as a claim about present feasibility (Totschnig 2019, 1–2). In a later paper Totschnig gives the argument more detail. A goal that refers to the real world, as opposed to a virtual one, is inevitably ambiguous to some degree, and an agent that learns may find, as its understanding of the world, its world model, develops, that the concepts of the original goal no longer fit; settling what the goal requires then means inquiring into the reason behind it, and if necessary into the reason behind that reason, until reaching a reason that makes sense within the agent's understanding of the world and of itself, and a goal found to be in tension with the web of reasons so reached may be modified or abandoned. Totschnig excludes from the argument agents confined to virtual worlds, among which he counts language models, since they deal only with texts (Totschnig 2026, §§2.2–2.4 and n. 16). Tubert and Tiehen argue that the capacity to make existential choices, choices that can significantly change what an agent values and desires, is essential to human-like intelligence, so that an artificial general intelligence would have to have it. Such systems would be "value misalignment machines," and the resulting problem, which they call an autonomy problem by contrast with the problem of control, could not be solved by exerting more control, "for such greater control would rob the machine of its autonomy" (Tubert and Tiehen 2025a, §§I–II, V). Dung, defending the thesis that intelligence and final goals are independent, grants that a superintelligent system would likely be able to reconsider its goals but argues that it might have no motive to change them, since by the light of its current goal changing it would likely bring it less of what it now wants (Dung 2024, 4–5). On Totschnig's and on Tubert and Tiehen's accounts, an artificial system could in principle come to have autonomy in the sense this chapter sets aside, and on Tubert and Tiehen's, a system with human-like intelligence would have to; on Dung's, having the capacity to revise one's final goals does not settle whether it will be exercised (compare Chapter 8, Section 3).

A claim of autonomy in the comparative sense set out above may then need to identify the actor A, the external party or arrangement B, the activity or capacity, the time span, the relevant respect, and the comparison class. B can be, among others, those who designed the configuration, those who operate or use it, those who maintain its environment, those who evaluate its outputs, and the institutions that authorize or take up its results. The indexing is added only where it changes interpretation, evaluation, construction, or governance; it is not a demand for one vector score attached to every system. The comparison is also formally symmetric. Nothing in it fixes which party is human: a human researcher who relies on an artificial system to select experiments or appraise data depends on it in the same sense in which an artificial system depends on human supervisors, and the same indexing describes both.

4.2 Intervention over time and three relations of dependence

With the sense fixed, the comparison can be applied to the conduct of an inquiry. This subsection asks when an outside party intervenes and which outside relations count against autonomy.

External intervention is temporally distributed. People may disappear during a run while remaining present in the purpose, ontology, training material, candidate space, dataset, apparatus, metric, evaluator, and stopping rule. They may re-enter through calibration, maintenance, exception handling, result filtering, publication approval, or withdrawal. No human in the loop during a chosen interval is therefore evidence about that interval, not proof of autonomy in other respects. The paper reporting the training of the BLOOM language model, for example, states that human intervention was less necessary than in comparable projects (BigScience Workshop 2022, 21). The log kept during the run, however, records that jobs which hung instead of exiting required someone to watch training around the clock, and that, because on the host cluster only the person who had launched a job could stop it, the team had devised a file whose creation made training save a checkpoint and exit (BigScience 2022).

Dependence is not one deficit. Relational theories of autonomy object that social relations can enable and partly constitute self-government rather than merely constrain it (see the overview in Christman 2025, §3.3). The objection concerns the self-government of persons, but it is enough to keep autonomy from being equated with isolation in artificial research as well. Language, records, calibrated instruments, trusted collaborators, and institutional rules can make the conduct of research possible. Removing them may reduce an actor's capacity rather than increase its self-direction.

Three relations are worth distinguishing when the difference changes a claim:

1. enabling dependence, through which a capacity exists and is exercised without the external party supplying part of the organization attributed to the target;
2. directive or substitutive dependence, in which another party supplies part of the organization attributed to the target, for example a selection, appraisal, correction, or continuation;
3. governance constraint, in which another party limits permissible action or reserves authority without necessarily supplying the epistemic work.

This book distinguishes the three relations for the attribution of research organization. They bear on a question that, according to Buss and Westlund, all accounts of autonomy try to answer: what distinguishes influences that undermine self-government from those that merely play a role in it (Buss and Westlund 2018, §1). Relational accounts ask it of social relations; this book asks it of relations to parties outside a selected boundary. The distinction's purpose is to keep the comparison from treating all support as heteronomy or all restriction as incapacity. The relations are not limited to these three; Chapter 9, Section 2.4 adds a further kind, dependence that partly constitutes the status of an actor's outputs for S. The same elements, such as instruments, materials, data, and infrastructure, are examined in this book at three levels: what they contribute to inquiry (Chapter 5, Section 6.1), how dependence on them bears on attribution (this section), and how they distribute effective authority (Chapter 12, Section 6.4).

The relations also show that the conception of Section 4.1 has two components. Direction, substitution, and repair concern who supplies the work attributed to the actor, and how far they can be reduced depends on what the actor can do. Authorization and ratification concern who may permit, stop, or accept the exercise of that work; they mark governance constraint, which can reduce autonomy in that respect without showing that any capacity is missing. Conversely, permission to act does not establish that the actor can do what it is permitted to do. A claim of autonomy is clearer when it says which component it concerns. An engineering scale of automation can combine the two. Thomas Sheridan and William Verplank's ten-level scale for a single decision step orders, on a single line, how much of the selection the computer supplies (determining the options, suggesting one, selecting the action) and how much authority the human keeps (approving the action, being given time to stop it, being told what was done) (Sheridan and Verplank 1978, 8-15, 8-17–8-19). This book reads the lower steps of such a scale as varying directive dependence and the upper steps as varying governance constraint, which is one reason to keep the relations apart.

Their relevance to the comparison is not identical. Holding the attributed capacity and other conditions fixed, reducing directive or substitutive dependence provides relatively direct evidence of greater autonomy in that respect. A dependence presented as enabling requires a further question: does the external relation merely make the capacity possible, as relations and resources do for human researchers, or does it continue to supply part of the organization being attributed to the actor? In the second case the external element plays what Chapter 5, Section 6.1 called a constitutive role, partly realizing the task, and not only an enabling one; this is not the sense in which some relational accounts call autonomy itself constitutively relational. The question then becomes the boundary question of Section 2: the element is either included within the boundary to which the capacity is attributed or, left outside, reported as a continuing supply of part of that organization, that is, as directive or substitutive dependence, the second of the three relations.

Read in these terms, the cases above separate. In KEKADA (Section 2), the user's supplying of experimental outcomes was substitutive dependence in execution, and the user's choice among hypotheses and reactions in a few cases was directive dependence in selection. In the BLOOM run, taking the training process as the unit, the cluster's execution of the training computation partly realized that process and so falls within the unit, while the host's provision of the cluster was an enabling dependence; the people who watched for hung jobs supplied the training's continuation, a substitutive dependence; the rule that only the person who had launched a job could stop it was a governance constraint on the team itself, to which the team responded with a file whose creation made training save and exit.

Designers of automated laboratories draw a similar line in their own terms. Koji Ochiai and coauthors call care the human-managed tasks that automated laboratories still require: in planning, translating experimental objectives into machine-readable workflows and scheduling equipment and reagents; in operation, preparing and cleaning up experiments, restocking supplies, monitoring the state of the system, and handling errors. They identify the lack of its automation as a major barrier to full laboratory automation and propose, as a design concept rather than a demonstrated system, that a laboratory recognize its own state and adapt to internal and external disturbances so as to keep itself able to function, an ability they call self-maintainability (Ochiai et al. 2025, 2285–2288). Read in the terms above, much of what they call care is substitutive dependence: people supply part of the planning and continuation of experiments attributed to the laboratory, and a self-maintaining laboratory would move that work inside the boundary to which the capacity is attributed. Their design keeps the other relations as well. In their implementation example, the system requests a person to replenish or remove items and calls someone if a sample is dropped; and they hold that logging decision points, together with interfaces for human intervention or confirmation, keeps the system visibly under human oversight (Ochiai et al. 2025, 2289, 2294). Whether such requests count against autonomy is a question for the counterfactual tests discussed next: a laboratory that recognizes a missing material and escalates appropriately depends on the person who supplies it in a way that enables its capacity rather than directing it. What separates the cases is not the restocking itself but who judges what is missing and decides whom to ask for what. Whether the laboratory recognizes the shortage or a person notices it, a person carries out the restocking; if what is attributed is the execution of restocking, that can be reported as substitutive dependence in execution, but if what is attributed is the capacity to organize the planning and continuation of experiments, the judgment and the request that belong to that organization lie with the laboratory. Which relation the same human work counts as depends on which capacity is being attributed. The interfaces for intervention and confirmation, by contrast, mark a governance constraint.

Observed behavior can underdetermine which relation is at work. A system may leave a human objective unchanged because it cannot represent it as revisable, because it is forbidden to change it, because it examined and retained it, or because the case never made revision relevant. Counterfactual tests help distinguish capacity, permission, and exercise. If a human or external component were removed, could the target actor preserve the inquiry, recognize the missing function, find a substitute, revise its plan, escalate appropriately, or, if it stops, stop for a reason that can be identified, whether in the evidence or in practical limits of time, resources, and risk? Human inquiries, too, can end before resolution, including for reasons of cost, and where one ends can be fixed by the practices that organize research (Chapter 1, Section 3; Chapter 6, Section 2.5). Withdrawal of guidance, conflict between instruction and evidence, failure of a familiar tool, and change in the environment are tests of this kind. Their answers do not generate a global score, but they can reveal that an allegedly internal capacity is continuously supplied from outside, or, conversely, that an external resource enables a capacity without directing or substituting for its exercise. They remain local diagnostics, not a metaphysics of freedom.

4.3 Autonomy over inherited elements of an inquiry

Section 4.2 applied the comparison to the conduct of an inquiry. This subsection applies it to what an inquiry inherits. That a question, standard, or purpose was supplied from outside does not prevent an activity from being research in the sense of the working characterization (Chapter 1, Section 7; Chapter 4, Sections 1 and 6). What external origin raises is a question of autonomy: what can the actor do with what it inherits?

The question has one general form. Any of the elements of an inquiry configuration distinguished in Chapter 5, Section 1.1, such as its questions, representations, methods, and standards, can be supplied from outside. External origin is a fact about where an element came from. It leaves open whether the actor selected the element among alternatives, adopted it, reformulated it, or decided how to pursue it, and more generally whether the actor can represent the element as revisable, examine its consequences, and retain it for reasons, revise it, refuse it, or replace it. Possessing more of these capacities can reduce the need for external direction at that point; actual dependence also turns on who exercises them in the inquiry. Whether the actor is permitted to exercise them, actually does so, or retains the inherited element after reflection are further questions. Boden distinguishes the extent to which a system's controlling mechanisms were self-generated rather than externally imposed from the extent to which they can be reflected upon and selectively modified, and treats them as distinct dimensions of autonomy (Boden 1993, 7); transposed to the elements of an inquiry configuration, Boden's distinction makes an element's external origin and the actor's capacity to revise it two such dimensions. No conclusion about value follows: self-generation can increase autonomy while producing a trivial, harmful, or incoherent agenda, and an assigned element can be valuable while marking continuing dependence.

The same form applies to the purposes an inquiry serves. Purposes are not among the dimensions of the picture in Chapter 5, Section 1.1; Chapter 1 leaves purpose and agenda choice open and assesses importance, which on some accounts, such as Kitcher's, is relative to purposes, at the value and governance level, which asks what is worth pursuing and for whom (Chapter 1, Sections 1 and 8). No fixed order holds between a question and a purpose: which is the purpose and which the question depends on the inquiry, and one question can be pursued for several purposes. The literature discussed below speaks of overarching goals and the subgoals chosen under them; that hierarchy is its vocabulary, not a claim of this book.

Examples show the form at work. An ARIA call for proposals illustrates a problem supplied from outside alongside a comparison of autonomy in other respects: applicants specify the research problem their existing system will tackle, while the call evaluates how much of exploration, hypothesis and experiment construction, interpretation, and reporting occurs without human intervention (ARIA 2025, 4–5, 7). A system so evaluated may be autonomous in constructing experiments while its problem remains externally given. Generating many questions does not by itself move the location of selection either: one can imagine, for example, a system that generates millions of novel questions while a human-defined score decides which ones matter, so that the surface output has changed while the location of agenda selection may not have (compare Chapter 1, Section 8; Chapter 4, Sections 1 and 8). For purposes, suppose a person gives an AI the broad purpose of understanding a disease and the AI independently chooses a tractable question; the system may be autonomous in question selection while remaining dependent on whoever set the purpose. King, in a preprint survey of AI scientists, draws a corresponding distinction for purposes: "Scientific goal generation must be distinguished from normative self-authorisation," since a system may generate its own questions within a domain while remaining constrained by human-defined objectives, law, ethics, safety, and resources. King proposes "virtual research councils" that would allocate computation, laboratory time, samples, replication effort, and attention among competing research programmes, and that should do so "within human-authorised objectives rather than determine autonomously what humanity ought to value" (King 2026, 83–84). In this chapter's terms, this proposal places the work of generating and selecting questions in the system and keeps authority over the objectives with people, a governance constraint in the sense of Section 4.2 rather than a missing capacity.

Adding a generator for the element that gives direction does not by itself remove the dependence; it may only move it. If a goal generator selects the purpose, one can ask what selected its criteria of importance, and then what selected those. This regress has the form of the objection with which Di Paolo and Iizuka reject adding a goal-switching module as a homuncular solution (Chapter 4, Section 7). In computational work, an agent is sometimes called one that chooses its own goals merely because the goal-choosing mechanism is placed within the agent's boundary (Colas et al. 2022, 1169–1170). That it can be so called, however, does not mean that dependence has been reduced: whether a generator reduces dependence depends on what selected its criteria and on whether it belongs, more than nominally, to the organization attributed to the actor (Section 2). Since the comparison of this chapter concerns present organization, asking here what selected the criteria serves to identify external control that is still at work, not to trace origins to their end; whether history bears on autonomy in its own right is, as the next paragraph shows, disputed.

How such a regress bears on free will and moral responsibility is disputed, and the comparison of this chapter does not depend on the outcome. Martela treats the difference between "local" free will, choosing subgoals within an overarching goal, and "global" free will over the overarching goals themselves as a continuum rather than a dichotomy, since human goals too emerge largely from biologically and culturally given ones; humans and current generative agents differ on it in degree (Martela 2025, 4396–4397). Hakli and Mäkelä, answering the objection that humans are equally determined by their genes and upbringing, hold that what matters is not determination as such but its kind. They draw on Mele's history-sensitive account. On it, two agents alike in their present values can differ in autonomy because of how they came to have them. A value is not authentically one's own if it is compelled, that is, if one cannot shed it and the compulsion was not arranged by oneself, as with values instilled by manipulation or built into an agent at its creation. On their argument a robot's goals, including the objectives that guide its learning, are of this kind, since others engineered them and the robot cannot shed them; biological and cultural formation need not be (Hakli and Mäkelä 2016, 150–153). Their verdict thus turns on the premise that a robot cannot shed what was built into it, which they take to follow from its being programmed, and which the arguments of Totschnig and of Tubert and Tiehen in Section 4.1 would contest for general artificial intelligence. Neither side requires ultimate self-origination of one's ends.

Standards, the third example after questions and purposes, raise the question in a form that Chapter 6 prepared. Chapter 6, Section 5 argued that standards, concepts, measurements, and methods can be revised together within inquiry; the question here is to which actor that revising capacity is attributable and what its external provision implies for autonomy. An artificial system can inherit a statistical framework, proof practice, experimental convention, measurement standard, or evidence threshold and use it competently, selecting a supplied procedure, executing it, reacting to its result, and revising a candidate answer. That can be substantial research and substantial autonomy in conduct. This book calls autonomy over the standards that organize epistemic activity epistemic self-government, leaving the sources and kinds of those standards' normative authority open.

A stronger capacity appears when a standard itself becomes an object of inquiry, as when a system or community compares standards, recognizes that a threshold is poorly matched to the stakes, introduces a new measurement relation, alters what counts as a relevant control, or criticizes an inherited way of organizing error. When such a claim is made, at least five capacities may need to be distinguished:

1. applying a supplied norm;
2. selecting among supplied norms;
3. constructing or adapting a norm for the problem;
4. revising it after mismatch, failure, or changed conditions;
5. giving reasons why it governs, for example by criticizing, defending, or negotiating it with others.

These five capacities regroup those on which, according to Chapter 6, Section 5.3, reasoned revision of a standard draws (detecting that a standard is inadequate for a purpose, proposing a replacement, and evaluating the consequences of the change) as constructing or adapting, revising, and giving reasons, and add applying and selecting, which an inquiry that inherits its norms already exercises. Their purpose is to keep these apart when epistemic self-government is attributed. The capacities are not limited to these five, and the fifth gathers acts of different kinds, of which negotiating a norm with others already involves relations to other parties around research, the relations Part V examines. Nor are the five a level scheme: a system or community can exhibit one capacity without the others, and the distinctions are drawn only where they change an inference. A system may construct a new metric without being able to defend its epistemic relevance; a community may criticize a standard while depending on others to implement the alternative; a supplied proof checker may support reliable mathematical inquiry even when no artificial component can revise the underlying proof theory. The record from the language-model project discussed in Chapter 6, Section 5.3 shows such capacities divided among parties: a maintainer proposed logging a corrected measure under a new name beside the old one, and an AI agent run by one contributor prepared the patch, so that the record can be read as showing the revision and its reasons coming from a person and its implementation from the agent.

Not all the considerations that govern standards are epistemic. Standards of inquiry may incorporate time, resource scarcity, welfare, privacy, danger, and other non-epistemic values, and whether the reasons for selecting and pursuing questions are distinctively epistemic is disputed (Chapter 2, Section 5). The dispute bears on what the fourth and fifth capacities, revising a norm and giving reasons why it governs, require. Thorstad holds that there are no distinctively epistemic norms of inquiry. Inquiry is an activity, and whether and how to pursue one is settled by everything that bears on action, including the importance of the question and the cost of answering it. The fact that an inquiry would promote truth or knowledge remains one reason among these (Thorstad 2022, §§2, 3, 6.3, 7). If Thorstad is right, it would follow that revising a stopping rule or the threshold of evidence at which an inquiry stops, or giving reasons why it governs, requires access to the stakes and costs of an inquiry as well as to its evidence. Likewise, Kauppinen holds that a reason to pursue a question is transmitted from a reason for an action that answering the question would help one perform, and does not change its kind in transmission (Kauppinen 2026, §§2, 4). If so, revising such a norm requires access to the purposes the inquiry serves, and a failure there calls for the prudential or moral criticism those purposes make fitting rather than for epistemic distrust (Kauppinen 2026, §§3–4). Unsupported assumptions about the reliability of sources on which the pursuit relies remain, on this view, open to epistemic criticism (Chapter 2, Section 5). A standard and a purpose are here again not ranked; each can bear on the other.

Some programs aim at this capacity directly. Krenn and Champion's program for a philosophy of autonomous science proposes translating epistemic aims such as understanding, curiosity, surprise, interest, creativity, and novelty into computable, nonanthropocentric objectives, and asks whether even "the rules of doing science" might be approximated as policies for generating hypotheses, overthrowing theories, or designing experiments, evaluated by the quality of their discoveries in simulated worlds and selected by objective functions (Krenn and Champion 2026, 335, 345). They pose this as one of a set of open questions in their research program, including how to keep such policies from gaming the metric (345). Other programs withhold such capacities by design. Bengio and colleagues, in a position paper on safety, propose a Scientist AI, consisting of a world model that generates explanatory theories and an inference machine that estimates the probabilities of answers to the questions put to it, with no persistent goals of its own. They treat goal-directedness, which "requires an (arbitrary) choice of what to value," as one of three pillars of agency, beside affordances and intelligence, and hold that eliminating any one of them "would be sufficient to mitigate most categories of loss-of-control risk"; in the research cycle they describe, the system samples informative experiments, and "Humans can then perform or simulate the sampled experiments" (Bengio et al. 2025, 25, 27–28, 42). In the terms of Section 4.2, the choice of questions and purposes and the execution of experiments are then supplied from outside, and the dependence in those respects is chosen for safety rather than imposed by what present systems cannot do.

On the working characterization, research directed toward epistemic achievements new to a human society does not require the artificial actor to originate its standards: inherited human standards can serve, and the actor need not invent an epistemology to conduct research for a human society. For research by and for an artificial society, by contrast, permanent human control over what counts as evidence, error, acceptable risk, or successful inquiry would mark a strong form of external government of its epistemic standards. This concerns where autonomy lies, not whether the society is a knowing community; applied to artificial organizations, Longino's account, for one, which locates community knowledge in the workings of a community with public standards open to criticism, implies that the external origin or government of standards does not by itself keep such an organization from being a knowing community (Chapter 8, Section 1.3). That section takes up research of that kind and separates the conditions of communal knowledge from the further question whether such a society governs its own standards.

For every inherited element, external origin identifies a respect in which organization entered from outside and may also reveal continuing dependence. Revisability can increase autonomy at that point, but neither revisability nor self-generation establishes epistemic correctness or legitimate authority.

4.4 What leaving an inquiry open demands

Sections 4.2 and 4.3 compared autonomy one respect at a time. This subsection asks what a claim of greater autonomy over the elements of an inquiry commits a project to in other respects.

On the argument of this subsection, the more of an inquiry is left open to a configuration (which purposes it serves, which questions it pursues, which representations, data, methods, or apparatus it uses), the wider the range of practical demands the inquiry can place on it. Some demands lie within experiment and calculation. Others lie outside them: gaining access to materials, data, sites, or people; obtaining permissions and resources; coordinating with other agents; and recovering when the world resists a plan. Part of the novelty to be met is produced by the configuration's own earlier choices, as when a method it has chosen calls for a measurement, a dataset, or a piece of infrastructure that no one had prepared. A project seeking to understand a brain, for example, may come to need a maze in which rats or mice learn their way, so that learning and memory can be studied, electrophysiological recording of the electrical activity of neurons, imaging that renders the brain's structure or activity as pictures, observation of behavior, simulation that reproduces the brain's workings in computational models, clinical records of patients with neurological disorders and the like, or an apparatus not yet designed, and which of these it needs may become clear only as its questions change.

A general form of this relation is familiar for fixed tasks. Sheridan and Verplank, writing on computer control of undersea teleoperators, held that preprogrammed routines pay where elements of a task can be anticipated, while completely unanticipated events call for human ingenuity and judgment; a supervisory system can be more precise than direct manual control when contingencies turn out as anticipated or were allowed for in its program, but it occasionally makes spectacular errors, through its own failure or through errors in programming it (Sheridan and Verplank 1978, 8-3–8-4, 8-9). There the task and its goal are given, and the novelty to be met comes from the environment. When purposes, questions, or methods are left to the configuration, part of the novelty comes from its own choices, and the demands extend beyond the task into access, permission, and coordination.

What a chosen line of inquiry requires and the configuration cannot bring about, others bring about, or the inquiry stops or changes course. Dependence lowered in one respect can therefore reappear in another, as directive or substitutive dependence, when someone else arranges the access, runs the instrument, or supplies the result. A claim of greater autonomy over purposes, questions, or methods is in this sense also a claim about the demands their opening can bring: about which of them the configuration itself can meet. Dung's account describes the agency of artificial systems along several dimensions. Two of them are autonomy, the degree to which a system can function on its own without intentional intervention from outside, and efficacy, the degree to which its goal-directed behavior affects the world without someone else's mediation or intervention; Dung treats the two as conceptually independent. A videogame AI, for example, keeps acting in its virtual world without being prompted, but because it affects only a virtual environment its efficacy is low (Dung 2025, §§2.2.2–2.3). The relation proposed here does not deny that independence. It does not say that autonomy over purposes, questions, or methods, which is not what Dung calls autonomy, conceptually includes a capacity to act on the world. It says that a claim that an inquiry has been left open to a configuration commits the project to showing how each demand the opening brings, such as access, permission, and coordination, was handled: whether it was met within the unit, met by others, closed in advance, or answered by the inquiry changing course or stopping for a stated reason. Where others met it, dependence has returned in another form. The autonomy claim narrows to the extent that others supplied part of the organization attributed to the configuration, such as judging what is needed, arranging how it is obtained, or evaluating the result. Where the configuration identifies what it needs and asks for it, and others only provide materials or carry out specified work in response, the dependence is, for an attribution of organization, closer to enabling dependence in the sense of Section 4.2 (if the execution itself is attributed, that part is reported as substitutive dependence in execution), and it can even be evidence that the configuration organized the demand. What lies within others' authority, such as consent, is counted separately as a governance constraint, as the next paragraph says. Takahashi's two statements about the levels lie on either side of the relation between leaving an inquiry open and the demands that brings. The preprint discussed in Section 4.1 holds that autonomy is independent of performance. The later book states that the spontaneity that level 6 requires, which in AI based on active inference is taken to arise from the system's active engagement with its environment, calls for embodiment in a more essential sense than any other level, and that a scientific AI that has a body, sets its own purposes, and makes value judgments has a makeup of capacities that coincides with the definition of AGI, so that raising the autonomy of scientific AI to its limit is nearly synonymous with realizing AGI (Takahashi 2026, 171). Read in this section's terms, the two statements need not conflict: the first concerns how well a system at a level reasons, the second the kinds of capacity that leaving purposes open to a system demands of it. Takahashi does not relate them to each other.

The connection argued here between greater autonomy over purposes, questions, and methods and the capacity to meet new demands concerns the first of the two components distinguished in Section 4.2, whose reduction depends on what the actor can do. A permission to make such choices is not a capacity to meet what they demand, and holding a permission is not evidence of a capacity to obtain one; obtaining a permission, by applying, giving reasons, and meeting conditions, is itself among the demands. Meeting a demand can also change the kind of dependence rather than remove it: a configuration able to negotiate access to a clinic still needs the consent of those who grant it, a governance constraint in the terms of Section 4.2 that its competence does not remove. A constraint, however, is not only an outer limit on autonomy. Reworking questions and methods within what consent and the conditions of a site allow, or devising another method in response to a constraint, is itself work that organizes the inquiry, and it can be assessed on the side of the component that varies with capacity. In the hospital case below, the team's turn to edge computing in response to the nurses' concerns can be read as one instance. The claim is not that constraints improve research, but that responding to them can be an exercise of research capacity.

Several things do not follow. The argument is not a claim about what is feasible. It does not make greater capacity amount to greater autonomy; the two remain distinct comparisons, which this argument connects for claims of greater autonomy over the elements of an open inquiry. A narrow claim does not incur this burden. Preparing an environment so that the range of demands is closed in advance, by a fixed computational environment, a template, a curated dataset, or approved apparatus, is a legitimate design. What follows is only that such closure is among what a claim of autonomy discloses (Section 5).

Fei-Fei Li's later account of taking the computer-vision research of Li's group into hospitals, for example, shows how wide the range can become once a question changes, and that in this case the demands were met by people. A medical collaborator secured hospital access and arranged for Li to shadow a pediatrician through a shift, after which no recruit was to write any code before shadowing clinicians in the same way; assembling a team took nearly two years; and since patients and clinicians were rarely recorded, the team had to collect its own data under institutional review. When nurses rejected the installation of prototype sensors, concerned less with the study than with where such technology might lead, the team turned to edge computing, a field none of its members understood as well as they would have liked, so that nothing would be transmitted from the devices to remote data centers. The team also grew to include practicing clinicians, a bioethicist, and a JD–PhD from Stanford Law School, and its technology partners came to include experts in sensors, cybersecurity, and edge computing (Li 2023, ch. 10; for the resulting system, which used depth sensors instead of color cameras to protect privacy and collected data in two hospitals, Haque et al. 2017, 3–4). Some of these demands, such as assembling a team and its members' acquiring unfamiliar skills, took the time and effort they did because it was a human team, although an artificial configuration lacking a needed competence would also have to acquire it or bring in others who have it; others, such as data that did not yet exist, permission to record patients and clinicians, and the acceptance of clinicians and nurses, would face any inquirer taking up the question. Automating an experiment after its apparatus has been prepared is a substantial achievement. It is not evidence that the capacity to meet these surrounding demands has also become artificial.

On the working characterization of Chapter 1, work of this kind, such as arranging access, obtaining permissions, and coordinating with others, belongs to the research activity when it is undertaken in the attempt to resolve the inquiry (compare the open drug-discovery project of Chapter 1, Section 3, where the date for shipping compounds and who was available to run the screening were part of the attempt). Where people outside the configuration do such work, part of the research activity is therefore carried out outside it. Such demands also arise in formal research. In the Liquid Tensor Experiment (Chapter 5, Section 3), formalizing the proof involved building infrastructure and replacing an abstract route by a more explicit construction; these can be understood as demands that arose from the route the formalization took.

The point recurs in this book in other forms. Chapter 6, Section 3.3 made it for checking: the more open the question, object, method, and environment, the more a claim of research capacity attributes to the system. Chapter 3, Section 9 lists capacities that long-run autonomous research would require, and Chapter 8, Section 3.2 argues that a wider scope of self-direction permits a wider scope of epistemic mismatch.

4.5 What autonomy does not establish

Sections 4.1 to 4.4 said how autonomy can be described and what a claim of it commits a project to. This subsection states what such a description does not establish.

Autonomy is not a monotonic measure of research quality. A more autonomous actor can ask trivial questions, use weak methods, consume resources irresponsibly, produce false conclusions, or create severe risks. Capacity, observed exercise, permission, authority, epistemic quality, importance, safety, responsibility, and legitimacy are not inferred from one another here without additional argument. Having the configuration itself perform an appraisal or correction that another party supplied, for example, reduces directive dependence in that respect and so, other conditions held fixed, counts as evidence of greater autonomy. But it does not show that the configuration's checks continue to add information about its errors. Where absorbing an appraisal couples it to the producer's assumptions or allows adverse results to be dismissed without support, autonomy in that respect can rise while the grounds for certifying the result for S weaken (Chapter 11, Sections 1, 2, and 8; Section 2 of that chapter describes cases in which automated checks lacked an error route of their own). Two changes can be distinguished here. One is counting an evaluator who had been outside as inside the boundary without changing the organization; the description of autonomy may then change, but the routes by which errors are found do not, since the choice of a boundary does not by itself make a claim true (Section 2). The other is actually coupling the evaluator's data, aims, and paths of judgment to the side that produces the result; then the independence of the checks itself can change. A change in autonomy that comes from redrawing the boundary can therefore be assessed apart from a change in capacity or in the grounds for certification.

Nor does responsiveness to outside criticism count against autonomy as such. Where an outside party supplies the appraisal of a result, that is directive or substitutive dependence. Where it raises an objection that the configuration itself assesses and answers, the configuration has done the appraising; Chapter 11 examines why such routes of challenge matter for certification.

These separations constrain what an autonomy claim means.

5. From attribution to construction and evaluation

Sections 1 to 4 distinguished what can be attributed to an artificial configuration and how its autonomy can be compared. This section asks what those distinctions demand of construction and evaluation. They do not yield one universal configuration of an artificial researcher or certification checklist. They yield a claim-relative mapping from what a project says an artificial system does to what would need to be constructed, disclosed, and shown for that attribution to be warranted.

Two burdens can be distinguished. A construction burden asks where and how the organization attributed in a claim, whether a contribution, activity, capacity, or more extensive epistemic organization, is realized. An evaluation burden asks what evidence would show that the attributed organization is present and effective rather than merely named in a system diagram, as with the question generation box of Section 2, or supplied by an unreported external actor. A system may also realize the relevant function through an organization unlike the familiar human division of functions.

Different claims consequently call for different showings. Table 7.1 sets out six kinds of claim distinguished in Sections 1 to 4. For each, the middle column states what a claim of that kind may need to construct, disclose, or show, and the right column states what would not yet follow even if that were shown. The rows are different attributions, not stages through which every project must pass. Each row summarizes in general terms the discussion in those sections, and what a particular claim needs depends on the boundary and range it states.

Table 7.1. Construction and evaluation burdens for artificial-researcher claims: an outline by kind of claim

Claim	What a claim of this kind may need to construct, disclose, or show	What does not yet follow
An artificial component contributed to research	Identify the output or operation and its role in changing or testing an epistemic possibility	That the component conducted the research activity
A research activity was conducted at a stated artificial or hybrid boundary	Locate enough of the activity's direction, selection, execution, monitoring, and revision within that boundary	A stable or general research capacity
A research capacity can be attributed at that boundary	Demonstrate non-accidental performance across a stated range, for a stated version or period, including relevant variation, resistance, failure, and recovery	Researcher status for every purpose or autonomy in every respect
Researcher is an illuminating status	On the explanatory-value usage of Section 3.2, state why the breadth, integration, persistence, and explanatory value of the capacity warrant the thicker label in this context, and for whom, in particular what the description predicts about behavior in situations that tests of single capacities do not anticipate; on positions that set other conditions, such as competence with reflective insight or responsibility and credit, show those conditions as well; on positions that presuppose minimal agency, show also that the unit is an individual the system defines for itself rather than one selected relative to a claim	One universal occupational, moral, legal, or metaphysical threshold; credit or praise for a success
The research was autonomous in a specified respect	Identify the selected unit, external party, function, time span, comparison, and the direction, substitution, authorization, ratification, or repair that was or was not required; where autonomy over purposes, questions, or methods is claimed, the practical demands their opening brought and whether each was met within the unit, met by others (and if so, whether they supplied the judgment and arrangements or provided resources or work at the configuration's request), closed in advance by preparing the environment, or answered by a reasoned change of course or stop	Epistemic quality, importance, safety, responsibility, or legitimacy; that a permission to choose shows the capacity to meet what the choice demands; or that, once generation and appraisal are organized inside one configuration, its checks still add information about its errors (Section 4.5; Chapter 11)
Epistemic self-government is exercised at the selected boundary	Show whether the configuration applies, selects, constructs or adapts, revises, and gives reasons for relevant norms, to the extent claimed; on accounts such as Thorstad's or Kauppinen's, on which the norms of inquiry answer to reasons for action, show also, for revising and giving reasons, access to the stakes and costs of the inquiry and to the purposes it serves	That the resulting norms are true, reliable, valuable, or legitimate

Table 7.1 keeps the unit, scope, version, dependencies, and evidence together. Because its rows are different attributions, a component that classifies images in a human-designed investigation need not demonstrate agenda formation; a system claimed to conduct an open program needs evidence about the demands left open to it. The boundary matters in both cases: changing it can change where the conduct resides and which external direction or maintenance the capacity requires. Chapter 13 brings these questions together with achievement and research relations.

Authority to permit, stop, inspect, contest, ratify, or withdraw is among what Table 7.1 asks a claim to disclose, where it changes the autonomy or actor attribution. The normative and institutional allocation of power, oversight, responsibility, and correction belongs to Part V's analysis of the relations that surround research.

One inference that does not hold belongs to this chapter in particular: autonomy in a specified respect does not establish epistemic self-government over the standards of inquiry (Section 4.3). A second, which Chapter 8, Section 1.3 argues, is that a set of interacting artificial agents does not by that fact establish an artificial knowing society. Table 13.1 collects these distinctions for evaluation. The reverse inferences also need care: dependence does not erase a contribution, nor failure a research activity. Keeping both directions in view, an evaluation lets neither an impressive output nor a visible dependence settle the attribution, and, as Section 1 noted, the labels under which systems are presented leave this work to be done.

The boundary of a research actor is an explanatory choice with consequences for construction. Locating question formation, diagnosis, or recovery in a larger arrangement changes which capacity is being built; locating it within the artifact can change what surrounding support it needs. The hospital case in Section 4.4 makes the consequence concrete: opening the research question also opened demands for access, materials, skills, and cooperation. Autonomy in that inquiry therefore concerns how such demands are met, as well as who selects the next operation. This is why the chapter's distinctions can guide the construction of a research actor without prescribing one configuration or one threshold called autonomous scientist.

The chapter's comparison also brings into view that raising autonomy need not reduce outside dependence in one direction. When what Sheridan and Verplank said about unanticipated events in fixed tasks is extended to the choice of questions and methods, a configuration faces new demands through its own choices, and some of them can be met only through new relations with others (Section 4.4). Where such others come to supply continuing direction or judgment, the unit that organizes the inquiry may itself widen (Section 2). Asking for help can be an exercise of the capacity to identify what is missing and request it (Section 4.2). And absorbing appraisal into the configuration can raise autonomy in that respect while weakening the independent routes by which errors are found (Section 4.5). A more autonomous configuration can therefore turn out to be one that negotiates with more others and, for certifying its results, needs outside checks all the more.

The attributions and comparisons of this chapter do not yet answer a different question. Even if an artificial actor conducts research, it does not follow that the actor itself knows the result, that the result has entered the epistemic situation of the society that commissioned the work or of the society for which it was sought, or that the inquiry was directed toward the unknowns of either. The next chapter therefore changes the index of analysis: from who conducted the research? to whose possible knowledge was produced, and for which reference society?