Part I — What Is Research Taken to Be?

Chapter 1 — A Working Characterization of Research

What the characterization needs—and what it does not

1. A thin orientation toward epistemic achievement

The characterization needed here has a demanding but limited job. It is meant to range across fields without turning the way one discipline organizes its work into the form of research as such, and to make an artificial-researcher target explicit enough that we can ask what an artifact does, what people and institutions still do, and what an evaluation establishes. It need not identify the timeless essence of research or settle every contested boundary before bounded engineering can begin.

That limitation matters because candidate definitions can answer different questions. A condition for being a research activity is not automatically a condition for good or trustworthy research. For example, producing knowledge is not necessarily the same as producing understanding, making progress, occupying a useful role in a research system, receiving institutional recognition, or producing something important. How it relates to producing understanding depends on whether understanding is taken to be a kind of knowledge (Chapter 3, Section 3), and whether it is the same as making progress depends on the account of progress (Chapter 2, Section 8.3). On the working characterization proposed below, a failed project may still have been research. An institutionally recognized project may be epistemically defective. Data curation may enable major progress without itself resolving the research question it supports.

Candidates for characterizing research include operations and achievements. If research is characterized by an operation such as controlled experimentation, hypothesis testing, or paper production, practices such as proof and interpretation cannot be covered without substantial redescription. If it is characterized by an achievement such as explanation, understanding, or problem solving, a different question arises: whether treating these practices as inquiry directed toward that achievement better captures what they are doing. The cross-domain cases of Chapter 5 give reasons not to identify research with one operation, but they do not decide which achievement research is best described as oriented toward. Achievements such as knowledge and understanding face the same comparison: their adequacy, like that of operational candidates, is judged by what they clarify and what they exclude.

For building an artificial researcher this comparison matters in a particular way. An artificial system can produce novel outputs in great quantity, but whether it has thereby accomplished anything epistemic is a separate question. Producing a novel output and bringing about an epistemic achievement therefore need to be kept apart. This book accordingly characterizes research not by a particular operation but by its orientation toward achievement, leaving the kind of achievement open rather than deciding it in advance, and expresses this orientation in three terms: attempt, epistemic achievement, and novelty relative to a society.

The book's working characterization is as follows:

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

The sentence is not meant to fix a definition of research and then defend it. It is a working characterization, set down as a footing for thinking about what it would be for an artifact to do research, and is itself open to scrutiny. Why each term takes this form is explained below and from Section 2 on.

Here, epistemic achievement is an open term for epistemic success, principally examined in this book through knowledge and understanding. It does not supply an exhaustive classification or assume that the achievements share one structure, are equally valuable, or are mutually exclusive. What their realization involves, including whether it requires possession by a subject, depends on the account. The term does not itself require overcoming a significant obstacle or exercising significant cognitive ability; those stronger senses of achievement are examined in Chapter 3, Section 7.1.

Whether research is described as the production of knowledge or as bringing about understanding changes which aspects of research come into view; this is why the book uses the general term epistemic achievement and examines both. This chapter, however, explains what the working sentence's terms require and do not require mainly by taking knowledge as the example. For knowledge, the questions its success raises (truth, the subject to which it is attributed, and non-accidentality) and its social attribution have been examined in detail in epistemology and social epistemology, which makes it easier to see what each term does. Chapter 2 begins its analysis from propositional knowledge for the same reason. Understanding is touched on wherever a difference arises and examined as a whole in Chapter 3. Theories of understanding ask whether understanding is a species of knowledge or a distinct and perhaps more valuable epistemic achievement, and how its explanatory form (understanding why something holds) and its objectual form (understanding of a phenomenon or subject matter as a whole) are related (Grimm 2021); that question too is taken up in Chapter 3, Section 3.

Here attempt marks an organized direction, not necessarily a conscious intention entertained by every participant. The characterization requires that some feature of the activity—its plan, functional organization, allocation of roles, responsiveness to evidence, or place in a research practice—connect what is done to the prospective epistemic achievement. It is not enough that knowledge can later be extracted by chance from an otherwise arbitrary event. Which kind of organization is required remains open; its immediate purpose is to distinguish a research attempt from a lucky output while preserving distributed and nonhuman cases.

S has two roles: it is the reference against which novelty is measured, and it fixes the society for which the achievement is sought, that is, where the attempt is directed. Neither role attributes the achievement to S. S may be humanity, a specialist community, an organization, a group of humans and artifacts, or a possible artificial community, but not a collection assembled after the fact to make some output look new. Whoever conducts the research need not be S, and naming S does not settle who bears the achievement. Sections 2 and 6 take these points up.

The shorter precursor characterized research through the production of knowledge and then derived questions as a component of that characterization (Takagi 2023). This book retains attempt and novelty relative to a society while widening the prospective achievement and replacing produce with bring about (Section 4). The reason is to connect the general orientation with the alternatives examined in Part II, without deciding in advance whether understanding is a kind of knowledge or treating it only as a means to knowledge.

The sentence specifies the target conception examined in this book. It is not a necessary-and-sufficient demarcation criterion or a complete philosophy of research. Education, journalism, investigation, invention, and ordinary problem solving can fit a broad reading. Conversely, an epistemic orientation can misdescribe practices centered on making, participation, relationship, or transformation. Whether the orientation is necessary for every practice properly called research remains open. These pressures are examined in Sections 4 and 8. Section 4 distinguishes products, roles, and achievements; Section 8 compares both what the orientation captures and the distinctions it imposes with varied research practices.

Hoyningen-Huene's account of science faces a problem like the one this chapter faces: how to say something general about very varied practices. Hoyningen-Huene holds that scientific knowledge differs from other kinds of knowledge, in particular everyday knowledge, primarily by being more systematic, and develops this comparative thesis along nine dimensions (Hoyningen-Huene 2018, 3, 7 n. 12). The thesis compares bodies of knowledge, not research activities. Hoyningen-Huene grants that the abstract term systematicity is vague, and that contrasting terms (not random, accidental, arbitrary, unplanned, or unordered) give it little precision; in the abstract, he holds, apparently no more can be achieved, so the term has to be made concrete in contexts. The concrete notions of systematicity vary with dimension, discipline, and period and are related by family resemblance, overlapping partial similarities without one feature common to all, rather than by necessary and sufficient conditions, so that the unity of the sciences that the thesis claims is, on Hoyningen-Huene's account, an extremely weak one; not every dimension applies to every field (predictions, for example, need not) (Hoyningen-Huene 2020, 89–91). Science, on this account, is not everywhere sharply bounded from nonscience: between scientific research and related activities tuned to practical purposes lie transitional areas in which only a necessarily rough and vague distinction can be drawn (96). In his 2018 replies, Hoyningen-Huene makes explicit that his comparisons establish at best that systematicity is necessary for science, not that it is sufficient, and that the account's demarcation criterion compares how much the overall systematicity of a putative pseudoscience increases over a period with the increase in the established science it competes with; non-quantitative comparative terms suffice for this, a quantitative scale for systematicity being doubtful. He welcomes historians' case studies as a concrete testing ground for the account (Hoyningen-Huene 2018, 1, 3–5). The working characterization is abstract in a similar way: so that it can say something general, and not empty, about varied research practices.

Hoyningen-Huene also meets the first of the pressures above, that a broad reading fits neighboring activities, from the side of knowledge rather than activity. Areas of knowledge production by professionals with a high demand for reliability, such as the Violent Crime Linkage Analysis System (ViCLAS), market research, product development, chess theory, and air-crash investigation, the account argues, can proceed very systematically while their product is unambiguously not science. The account locates the difference in what it calls epistemic connectedness, the connection of knowledge to other pieces of knowledge, which such knowledge lacks to a sufficient degree (Hoyningen-Huene 2018, 2–3). An article in a historical journal and a serious newspaper article on the same events may carry the same message, but the journal article's many footnotes make such connections explicit (Hoyningen-Huene 2020, 96–97). Compared with this chapter's working sentence, the professional cases above are each directed at producing knowledge, and they accordingly suggest that, if research is to exclude them as Hoyningen-Huene excludes them from science, an orientation toward knowledge new to S would not by itself do so. Hoyningen-Huene's own criterion, as just described, concerns the connection of new knowledge to other knowledge. Whether the relation to what S already knows, which a claim of knowledge new to S presupposes, could do comparable work is not pursued in this book, which, as Section 8 records, does not claim that the epistemic orientation separates research from every neighboring activity.

What the sentence organizes—and what requires a different vocabulary

The characterization operates at the level of the activity's broad direction. It is not meant to serve simultaneously as an ontology of research products, a process model, a theory of scientific progress, and a measure of social value. Those tasks can be separated without abandoning the proposed orientation:

1. At the activity level, research is defeasibly illuminated as an organized attempt to bring about epistemic achievements new to S.
2. At the process level, inquiry can change questions, evidence, credences (degrees of confidence), concepts, methods, abilities, recognized ignorance, and social availability before its target is attained.
3. At the achievement level, the book principally examines knowledge and understanding and competing accounts of their realization. Explanation, conceptual reorganization, and increased capacity can be outputs, means, or achievements, depending on the claim; naming them does not establish further achievement kinds or their success conditions.
4. At the value and governance level, importance, utility, beauty, emancipation, legitimacy, risk, and institutional recognition require their own assessment. They are not thereby hidden conditions of every epistemic achievement or of research status.

These are analytical levels, not four independent processes or a universal chronology. One proof or experiment can alter several of them at once. Likewise, an arrangement that sets how long an attempt can be sustained and where its episodes are bounded, for example a grant or a thesis (Section 3), bears on the activity and process levels while being assessed at the fourth, and it is not made a condition of research status either. Their separation is meant to avoid two opposed errors: making the success term so broad that it renames every valuable change, and replacing the common orientation with an unstructured list of field-specific goods. However, on accounts on which values enter warrant, the epistemic support for accepting a claim, or acceptance itself, such as those of Section 8 and those taken up in Chapter 11, Section 4, the fourth level is not separable from the others in this way, and a judgment of value can bear on whether a result is accepted.

The epistemic orientation distinguishes a desirable effect, novel artifact, or accepted statement from epistemic success. It thereby makes the construction target answerable to an account of what has been accomplished, where it is realized, and what distinguishes success from error, accident, or a useful but insufficient contribution. These are questions, not one theory's answers imposed on all achievements.

The working sentence names the intended achievement with an open general term, epistemic achievement. Naming it so has a cost. The general term epistemic achievement by itself specifies little about success: a claim of epistemic achievement does not by itself tell us whether knowledge was obtained or which proposition, if any, is claimed to be known. That becomes clear only once the kind of achievement claimed is specified.

Using the general term epistemic achievement does not, however, loosen the conditions of success. Once knowledge of a proposition is claimed, its truth is required on the view most epistemologists hold (Section 5). Understanding, too, remains answerable to its object under the accounts compared in Chapter 3. Some accounts of understanding allow idealized models, models that contain deliberate falsehoods, to contribute to understanding (Chapter 3, Section 3.4), so the general term might seem to loosen the condition of truth. Idealized representations can contribute to inquiry directed toward knowledge as well, however, so how far idealization may be used is decided not by the general term but by the account of each achievement. The general term settles only what can be said before a kind of achievement is specified; each knowledge or understanding claim has its own conditions to meet.

Nor does keeping epistemic achievement an open general term count everything as an achievement. Naming a concept, method, instrument, or capacity does not establish it as a kind of achievement. Whether something is a kind of achievement and how fully the book analyzes it are separate questions. Practical, aesthetic, and relational goods also remain assessable in their own terms; calling them valuable does not make them epistemic.

The characterization does not establish one global aim

Because the formulation says what research is by saying what it is an attempt to bring about, it can easily be read as a thesis about the aim of research. That reading would make it much stronger than intended. Claims that a practice has an aim can refer to importantly different relations: a practitioner's conscious intention, a collective or corporate goal, an institution's historical or present function, a constitutive feature without which an activity would not be research, a criterion of success, a normative ideal, or an interpretation that makes a practice intelligible, among others. These senses are the grounds that Lusk and Elliott examine and the interpretive route that Hüttemann develops, both discussed below, together with a practitioner's intention. Moving among these senses without argument would make attempt carry more philosophical weight than it can bear.

Some authors do defend such a strong thesis about the aim of research or science. Bird argues that knowledge is the constitutive aim of science, treating science as the social functional analogue of a cognitive system: just as the beliefs that a properly functioning cognitive system produces are knowledge, so what science produces when it functions properly is knowledge (Bird 2019, 171–172, 176, 178–179). There Bird argues from what counts as progress that science aims at knowledge rather than truth. A community that came to believe a true theory by an unreliable method would not thereby have made progress, and giving up that belief would not be a step backward. An activity progresses insofar as it approaches its aim; since the mere accumulation of truth or truthlikeness is not progress, the aim of science is knowledge rather than truth. Bird leaves understanding and problem-solving for discussion elsewhere and also sets aside van Fraassen's proposal that science aims at the empirical adequacy of its theories: it concerns how wide a range of claims the aim covers, a question independent of Bird's own. Even if the aim is restricted to the observable, Bird notes, it remains open whether science aims at theories whose observational consequences are true or at theories whose observational consequences are known to be true. Bird's question thus concerns the kind of success aimed at rather than its scope (Bird 2019, 173, 176; Mohler and Monton 2025, §1.1). Chapters 2 and 3 return to these alternatives as rival achievements. Bird later states a claim in the same direction in terms of function rather than aim: a knowledge-producing function can ground objective scientific norms directly and can help constitute what counts as science (Bird 2025a).

Against this, some hold that research or inquiry need not be assigned one global aim. Jane Friedman argues that we can theorize directly about the structure and norms of inquiry and the purposes of inquirers without positing one aim of inquiry. On Friedman's account, the analogy with games, whose aims are fixed by human convention, does not carry over to inquiry, which is not, as a game is, something humans have made by convention; nor need inquirers all share one aim of attaining a particular mental state such as knowledge, since someone who seeks high credence rather than knowledge is still inquiring (Friedman 2024). At the level of science, Rowbottom recommends dropping talk of the aim, or aims, of science in favor of more specific descriptive, evaluative, normative, and definitional claims, while granting that scientists and research groups have aims. Read literally as a claim about what scientists aim at, Rowbottom argues, such talk turns a philosophical dispute into a question for sociological survey; recast as a claim about what counts as success in science itself, it makes differences in value judgments about which inquiries are worthwhile look like disagreements about facts about science (Rowbottom 2014, 2–5, 12–16). Avoiding the word aim does not by itself settle whether research has one global aim or what could ground one, however, since, as Bird's argument from function shows, the same strong claim can return in other words.

Lusk and Elliott take the question on directly. Their question arises from debates about values in science. Views that let non-epistemic values, that is, values other than those bearing on evidence and knowledge, such as social, political, or personal values, influence scientific reasoning only where the evidence leaves gaps appear to rest on a global aim of science: an aim that holds across all fields and limits which local aims, the aims of particular projects, disciplines, or fields, are acceptable (Lusk and Elliott 2024, 2, 4–6). They therefore ask what could justify such an aim. They first examine five grounds that do not appeal to anyone's intentions and find none of them convincing (7–11). If the aim is a normative ideal, such as truth or explanation, nothing beyond philosophers' conflicting intuitions decides which ideal it is, and reading an ideal off practice would require a principled way of telling practices that fall short of it from practices that never pursued it. If the aim is a function, science is present for many reasons and makes many contributions to society, among them reliable information, prediction, health, success in warfare, and economic development, so no single function stands out. Against Bird's analogy between science and a cognitive system, they object that it favors a single epistemic function only for someone who already accepts a single epistemic aim, since aims such as producing new technologies have no evident counterpart in an organism's cognition. If the aim is an essential characteristic, the absence of any demarcation criterion, that is, a criterion that cleanly separates science from other activities, blocks the identification. A looser cluster of typical characteristics would instead yield many aims, each holding only locally. If the aim is a criterion of success, the analogy with chess, whose rules fix what counts as winning, does not carry over, since science has no comparably clear rules and what counts as its success is the very point in dispute. Drawing on Rowbottom's point above, they add that the dispute turns on value judgments that are unlikely to be settled uncontroversially without appeal to what scientists intend. If the aim is a corporate goal, one that an organization pursues through its structure whatever its members want, science lacks a uniform organizational structure, since its public and private funders have different agendas and incentives. On their account, grounds that do appeal to people's intentions fare no better. Scientists neither share nor rank the same goals, and no one ever designed science as a whole. A standard general enough for everyone to accept would have too little content to limit local aims. Positing several global aims does not help either: they would need a principled ranking that no one has supplied, or else the aims that operate in each context would in effect be local (12–15). They present this as a cursory analysis that shifts the burden of proof onto those who posit global aims, not as a decisive refutation (3, 6–7).

In place of global aims they defend a local aims approach, on which science has only local aims. On their account, the grounds that fail for science as a whole can yield aims once the scope is narrowed. Local aims can be identified, for example, from what a project was designed to achieve or from the mission that a field's founders gave it, and this limited scope is their reason for holding that local aims escape the objections above (15–16). Since a project usually pursues several local aims at once, they combine them through the adequacy-for-purpose view that Parker developed for evaluating models and that they extend to methods and hypotheses: an object of this kind is assessed by whether it is adequate for a specified purpose rather than by whether it is true. The purpose of an inquiry in a given context then specifies how far each of its epistemic aims, such as accuracy and precision, and each of its non-epistemic aims, such as ease of use or completion time, must be achieved (16–17, 20). On this view non-epistemic aims need not be subordinate to epistemic ones, a position they call the equal aims approach. Epistemic constraints nevertheless remain, since both kinds of aim are to be satisfied together as part of achieving the purpose. To the worry that, without a global epistemic aim, nothing would prevent epistemic corruption, they reply that a project's purpose can still be criticized from the local aims of the fields it relies on and from legal, regulatory, and ethical constraints. In particular, honesty requires researchers either to follow the epistemic standards of their field or to make clear that they are departing from them (18–22).

Their reasons against a global aim rest on different premises. Some concern how science as a whole has been organized: its scientists neither share nor rank the same goals, no one designed it, and its funders' differing agendas leave it without a uniform organizational structure. Others do not, among them that philosophers' intuitions about ideals conflict, that what counts as success is the very point in dispute, and that a standard general enough for everyone to accept would have too little content to limit local aims. They do not discuss artificial configurations (artificial systems, or combinations including them, that carry out research; Introduction, Section 6); applied to one whose organization is unified, the first kind of reason need not hold, and an aim could be identified from that organization, much as they identify local aims from what a project was designed to achieve. Whether such an aim would be global or local in their sense turns on whether the artificial configuration is taken to be research as a whole, and the other reasons are not affected by its organization. Their objection to Bird's analogy also changes form for such a configuration. It turned on comparing science as a whole with a cognitive system, although aims of science such as producing new technologies have no evident counterpart in an organism's cognition. An artificial configuration whose organization is unified may itself be a system of the kind the analogy compares science with, and whether a single epistemic function stands out in it then turns on what it was made and selected to do, although what its selection actually rewarded can diverge from its stated objective (Chapter 2, Section 4.1).

Even among views that retain aim talk, whether the aim is knowledge is contested. Beddor argues that the view that inquiry aims at knowledge conflicts with the thesis that a rational agent can know something without being absolutely certain of it. The thesis is closely related to fallibilism, the view that one can know that p even though one's evidence does not rule out that p is false: fallibilism, together with the principle that one should not be completely certain of what one's evidence leaves open, implies it, although the thesis is weaker and can be accepted even by those who reject fallibilism. Beddor's argument starts from a case. A scientist knows a hypothesis on the basis of experiments, yet rationally gives some slight credence to the possibility that all the results are due to confounding factors. A researcher elsewhere then sends the scientist the data of a more comprehensive study. The scientist seems to have a reason to look at the data, and looking is further inquiry into the same question; yet if inquiry aims at knowledge, the scientist has already attained its aim, and there is no point in inquiring further. Beddor backs the case with an argument from epistemic decision theory, which evaluates credences from the epistemic point of view much as decision theory, the theory of rational choice, evaluates actions. A higher credence in a true proposition is more accurate, that is, closer to the truth, and a credence that is not maximally accurate is not maximally epistemically valuable. So if inquiry aims at making one's credences as epistemically valuable as possible, someone who knows without being certain has not yet attained that aim. Finding attempts to reconcile the two wanting, Beddor asks which to give up and judges that requiring certainty for knowledge costs more: people readily say that someone knows something while granting that the person is not completely certain of it, and they distinguish knowing from knowing for certain. Beddor therefore rejects knowledge as the aim and proposes instead that the aim of inquiring into a question is to make one's credence in its answer as epistemically valuable as possible, which requires full confidence, a credence of 1, in the true answer. On this account, part of the appeal of the knowledge aim came from knowledge's being thought epistemically valuable, and the alternative takes that value as the aim directly; it also leaves room for knowledge that falls short of certainty and fits with epistemic decision theory. Beddor does not conclude that inquiry has to go on until certainty is reached: whether it is rational to continue also depends on whether other inquiries promise greater expected epistemic value (Beddor 2024, §§1–3, 5–6). Section 3 returns to this dispute about the aim of inquiry.

A further route gives a different answer to what attributing an aim amounts to: it treats aim attribution as an interpretation that rationalizes paradigmatic scientific practice. On this route an aim is neither posited as a constitutive feature of the practice nor dropped from the account; it is attributed as a reading that makes the practice intelligible. Hüttemann develops this route while also arguing that, if aims are attributed in this way, Bird's thesis that science aims at knowledge may need to be qualified by specifying what kind of knowledge science aims at, without disputing the thesis itself. The qualification may be needed because practice does not pursue just any knowledge. Scientists build accelerators at enormous cost instead of collecting knowledge that is easily had. And on one account of idealization, idealizations misrepresent a system's behavior while helping to identify the structure underlying it. If so, what rationalizes practice is an aim at some kinds of truth and not others (Hüttemann 2025, 1–3, 5–6). On this route, an aim attribution is made answerable to practice, so that what it attributes can turn out more qualified than the strong thesis.

These positions thus disagree not only about which aim science or inquiry has (Bird and Beddor), but about what attributing an aim amounts to (Hüttemann) and what could warrant the attribution (Bird and Lusk and Elliott).

The working characterization therefore makes a defeasible interpretive commitment. In treating the orientation as an interpretation that can be revised in light of research practices, it is close to the route Hüttemann develops. It proposes an orientation toward epistemic achievements new to some S, without identifying that orientation with every participant's conscious intention, an institution's historical function, an exceptionless essence, or, in view of Lusk and Elliott's argument, a global aim that limits which local aims a project may pursue. Lusk and Elliott themselves allow that an aim stated for all of science is not of the kind they reject if it places no such limits on the pursuit of local aims; they make this point about their own adequacy-for-purpose view (Lusk and Elliott 2024, 17 n. 5). In this respect, the orientation proposed here is a claim of that weaker kind, since it does not require that epistemic achievement be placed above a project's other aims. A project can pursue prediction, construction, care, emancipation, or commercial utility while also being organized toward epistemic success. The book does not aim to test this interpretation, but Section 8 compares the characterization with varied research practices to see how far it reaches and where it parts from other ways of taking research.

This limitation matters directly for building artificial researchers. The formulation does not license replacing research with a scalar instruction to “maximize knowledge.” It does not determine which unknowns merit attention, whose purposes should govern inquiry, how epistemic and non-epistemic values should interact, what risks are permissible, or when a result should alter a decision. They do not follow automatically from an epistemic objective, nor are they optional embellishments to it. They are problems of construction and governance that the formulation leaves unsettled, which Part V takes up by asking what each account implies (Chapter 9, Section 1).

2. What the terms are intended to do

Here again is the working sentence from Section 1:

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

Each expression in it has only a limited job and decides nothing beyond it. No term, for example, guarantees success or identifies who bears an achievement; Sections 3–6 treat each term in detail.

• Attempt separates organized activity from successful achievement (Section 3).
• Bring about directs attention to the realization of an epistemic achievement, which is not the same as making an output, need not be a single threshold crossed, and need not create the object known or understood (Section 4).
• Epistemic achievements are successes; claiming one requires making clear which achievement is claimed and under what conditions it obtains, principally through the analyses of knowledge and understanding in Part II (Section 5).
• New to S compares the intended achievement with a society's prior epistemic situation, not merely an investigator's prior state (Section 6).
• Society S opens the reference to a specialist community, humanity, an organization, a mixed collective, or a possible artificial community. It does not itself identify an achievement's bearer (Section 6).

Of the two roles of S noted in Section 1, the second, identifying the society for which the achievement is sought, concerns the direction of the attempt, not a guarantee that the society benefits or that any achievement is realized. Neither role by itself attributes an achievement to S as a collective subject, to its members, or to an arrangement of public resources. Knowledge and understanding require their own accounts of these relations (Chapters 2–3). The reference cannot be freely chosen after the event: the activity's treatment of prior results, its addressees, and the standards to which it answers constrain the choice (Section 6). Where several references are defensible, assessing a claim requires knowing which reference it uses (Chapter 2, Section 6). Neither a single social mind nor the equivalence of storage, accessibility, acceptance, knowledge, and understanding follows from the notation.

A claim about research involves two parties: the actor or organized system A that conducts the attempt, and the society S to which the achievement is to be new and for which it is sought. Since research has so far been done by humans, whoever conducted research and whoever possessed its results have, naturally, both been human; Humphreys notes that the assumption that science is human science has driven, explicitly or implicitly, all previous philosophy of science (Humphreys 2009b, §2). A later essay names an argument that is, Humphreys says, often implicit in discussions of automated science: because science is an activity done by and for the benefit of humans, the ultimate epistemic authority in science should always be humans. Humphreys grants that the attraction of anthropocentrism has good grounds and holds that humans' retaining moral responsibility over the pursuit of science and the use of its products is essential, yet argues that, under an expansive conception of knowledge, we ought to allow scientific knowledge "beyond the current understanding of humans" (Humphreys 2020, 12, 14). On this account, then, that science is done for humans does not by itself make humans its ultimate epistemic authority, nor require that what it comes to know lie within their current understanding. Humphreys had already separated science done by humans from science done for them in the extinction case quoted by Schrape: an activity need not be done either by us or for us to count as scientific, so that both who does science and whose purposes it serves may be nonhuman (Schrape 2015, 162–163; Chapter 8, Section 1.3 describes the case). Westerblad likewise reports Humphreys's rejection of a restriction to human performers and beneficiaries (Westerblad 2026, §2). The two parties distinguished here take up that separation, A on the side of by and S on the side of for, but S is not Humphreys's beneficiary. The for of that separation concerns whose purposes the outcome of an activity serves; S is the society to which an achievement is to be new and toward which the attempt is directed, a direction that, as noted above, does not guarantee that S benefits. Conversely, those who benefit from research need not wait for its achievement to enter S's epistemic situation. Dario Amodei, writing as chief executive of an AI developer about what AI could do for the world if things go well, states the aims of AI-accelerated biology as benefits, such as the prevention and treatment of nearly all natural infectious disease and a doubling of the human lifespan, and notes of the latter that even if it were accomplished within the period Amodei considers, "we may not know it yet" (Amodei 2024, §1 and n. 18).

Nor does either party settle who bears the achievement, and in Humphreys's case of a science carried on after human extinction no human does. Once research may be conducted by an artificial system, whoever conducts it and whoever possesses its results need not coincide. Who actually bears the achievement, in the case of knowledge whoever knows, is therefore asked separately, where it is asked whether the research succeeded. The bearer may be A itself, members or organizations of S, or a third party, and on accounts that allow knowledge without a knowing subject, on which whether a result is knowledge turns on public certification and acceptance rather than on anyone's attitude, there may be no bearer at all. Neither new to S nor for S therefore says that S knows. Both do presuppose that S's epistemic situation, through its members, organization, or records, could come to include an achievement of the kind sought; otherwise neither comparing the achievement with S nor directing the attempt at S would make sense (Section 6). What is presupposed is that S could come to hold such an achievement, not that it could follow the route by which the achievement was reached: on some accounts, a result whose derivation no member can survey can still come to be known in S, as in the case of a machine-generated proof verified by a transparent checker, one known to check what it does and how it does it (Chapter 3, Section 5.1). Because S is presupposed to be able to hold such an achievement, for an artificial S questions of artificial belief and possession arise even where S is only the reference.

What can be asked about an activity and its result accordingly divides into three: whether the activity is research for S; whether an achievement new to S has been realized somewhere; and whether it has entered S's epistemic situation, which the accounts of possession in Chapter 2, Section 6 answer differently. An artificial system may conduct research for human knowledge without itself knowing the output; on some accounts of belief, it may know a result that no member of S yet knows; and a result sought for S may not yet be known by anyone. A human or artificial actor may produce a result that is later taken up by an organization whose components include both humans and artificial systems (what this book calls a hybrid organization). An artificial community could in principle be actor, reference society, and knowledge subject at once if the relevant conditions were met. These separations are analytical; they do not pre-judge which individual, collective, or distributed entities can literally know. Who the bearer is turns on accounts of belief and of social possession (Chapter 2, Sections 2 and 6), and Chapter 8, Section 1 distinguishes the cases by S.

The same separation applies when an artificial actor helps human members understand. The actor, the bearers of understanding, and the reference society are distinct even if they sometimes coincide. A claim of collective understanding requires an account of what the collective's achievement consists in; whether a new explanation available to its members is enough differs among accounts (Chapter 3, Section 6).

Later chapters rely on these separations whenever they state what is required of an artificial researcher. A requirement stated for a system, such as that its conclusions rest on its evidence or that it revise them when its reasons are defeated, applies to the system as stated only where the system itself is claimed to know or understand. Where others are the knowers, whether humans or other artificial systems, the system's outputs serve as their evidence or as part of how they come to believe, and the requirement falls on their beliefs or commitments instead (Chapter 2, Sections 2.1 and 4.3). The same output can therefore be assessed differently depending on whose achievement is claimed.

The characterization does not try to explain research on its own; it shares the work with other theories. Inquiry theory can illuminate the process hidden within attempt. Epistemology examines possible successes and the conditions under which they count as knowledge, understanding, or other epistemic achievements. Philosophy of scientific practice and technology makes material and representational realization visible. Social epistemology asks how production and possession can be distributed. Theories of error control, objectivity, and governance examine quality, criticism, and correction. Theories of scientific progress ask which achievement's increase counts as progress (Chapter 2, Section 8.3). Accounts of agency and autonomy examine what makes an activity an agent's own and self-governed; this book draws on them to ask when a system has a stable research capacity or warrants researcher status. The book does not fold these questions into one definition. Doing so would answer, at the stage of defining research, questions about intention, group knowledge, autonomy, and the like that are still contested in each field. Claims about artificial research are accordingly assessed by asking separately which of these questions they answer.

3. Attempt separates research activity from its success

An experiment can be inconclusive, a proof strategy can collapse, an archive can fail to resolve a historical question, and a prototype can miss its target constraints. If research meant the actual realization of an epistemic achievement, these activities would be reclassified whenever their target was missed. A functional analysis of the treatment–research distinction makes a related point: unsuccessful activity can retain knowledge generation as its function even when it fails to achieve the intended effect (Lewens 2006). That account locates function in planning and intention; it supports the activity–effect distinction without supplying the more general notion of organized direction used here.

If success is not a condition of research activity, what makes an activity research under this book's characterization? It is the activity's being directed toward bringing about an epistemic achievement, which the book calls attempt. Where, then, is that direction located? Here attempt does not require every participant to entertain a sentence such as “I intend to produce knowledge.” The book uses this term, rather than a success term, so that success does not become a condition of research activity, the subject of this section; and it uses the term in this sense so that distributed and nonhuman cases are preserved (Section 1). The relation between activity and epistemic end could be grounded in an individual's intention, a group's plan, an artifact's organized function, a component's role, or participation in a norm-governed practice. The functional organization that this book lists among the grounds of orientation is function in the sense of the role a component plays in a system. An artificial system whose inquiry fails still has organization in that sense if it proposes hypotheses, runs experiments, and revises its hypotheses in response to evidence. Some accounts instead take an artifact's function to be what the artifact is for and trace it, for a new artifact, to its designer's intention. On such accounts the orientation of a new system's inquiry derives from its designers, and the question arises to whom the activity is to be attributed. That is a question of attribution, not of whether the activity is research, and Chapter 7, Section 2 takes it up.

Among the transitional areas that, as Section 1 noted, Hoyningen-Huene places between scientific research and related activities tuned to practical purposes are those between technological research and development and between economic research into markets and the market research that companies perform; in them the scientific and the nonscientific side may not differ in their degree of systematicity in other respects. Work by automobile manufacturers to improve fuel efficiency, for example, is typically tuned to particular engines that are already on the market or soon will be, whereas the analogous work in engineering departments of research universities typically has a wider range of intended applications, or none in sight, aiming at new design principles; the difference, on this account, is a higher degree of epistemic connectedness, at least potentially, due to greater generality (Hoyningen-Huene 2020, 96–97). In an earlier essay on why modern science is technologically exploitable, Hoyningen-Huene makes a related point about single actions. A point-wise testing procedure, the realization of particular values of an experiment's independent variables, has, like every action, an external aspect, its physical parameters, and an internal aspect, which includes its goal. A procedure with positive results may be repeated with a different goal, that of attaining certain values of the dependent variables, so that a test and a technical application coincide in their external aspect and differ only in the goal; a successful clinical test of a new medication is one of his examples (Hoyningen-Huene 1989, 10–11). Hoyningen-Huene also notes that even the experimental test of a hypothesis contains an element of what he calls forced chance: a result that deviates from the prediction may be due to the hypothesis, but may also reveal that some auxiliary hypothesis tacitly taken for granted in the set-up is at fault, so that background assumptions are constantly, though not intentionally, challenged by experimental activity (Hoyningen-Huene 2020, 98–99). The difference in intended applications can be read as a difference in the place an action occupies within an activity, rather than in each participant's state of mind. Applied to attempt as organized direction, the comparison suggests that identifying the direction requires knowing how an action relates to the activity's aims, since, as the coincidence of a test and an application in their external aspect shows, physically identical actions can differ in aim, and forced chance suggests that what an action bears on is not fixed once for all by the plan under which it was undertaken: an experiment set up to test one hypothesis can come, without anyone intending it, to bear on an assumption taken for granted in the set-up. For an organized attempt, this raises questions about what made an action relevant to an unresolved matter, what possible outcomes would have changed its role, and where the capacity to redirect the activity lay. Chapter 7, Section 2 asks these questions where an activity is attributed to an artificial or hybrid configuration.

If an activity's orientation is not identified with participants' states of mind, three things need to be kept apart: the activity's orientation, the norms by which the activity is assessed, and participants' states of mind. The epistemological debate over the aim of inquiry lets one see concretely how the last two come apart. To begin with, as Section 1 noted, views differ on what inquiry aims at. Kelp starts from a widely accepted characterization of the success of inquiry into specific questions as properly settling the question and, as a substantive account that makes it concrete, takes knowledge to be the goal of such inquiry and also takes understanding, which Kelp analyzes in terms of systematic knowledge, to be the goal of inquiry into general phenomena. Beddor instead argues that inquiry aims at making one's credences as epistemically valuable as possible, while Jane Friedman questions whether inquiry has a constitutive aim in the relevant sense and recommends discussing inquiry's structure, norms, and inquirers' purposes directly (Kelp 2021b, S1583–S1585, S1588; Beddor 2024, §§3.1, 6.1; Friedman 2024). Lee points out that claims of this kind admit two readings: as normative claims about the conditions under which an inquiry succeeds, or as descriptive claims about the personal aim that every inquirer has as an inquirer. Lee then holds that the paper's own view, that inquiring partly consists in trying to resolve a doubt one has, holds on either reading. First, it is consistent with any account of success conditions: supposing that knowledge is what makes an inquiry successful, of two people who independently inquire whether it rained last night, one who evaluates the evidence poorly fails to know, so that only the other's inquiry succeeds, although each inquiry was sustained by doubt. Second, it is not threatened even if, as Friedman argues, different inquirers have different aims and none is shared by all: whether one aims at knowing, at certainty, or at a belief justified by one's evidence, one pursues that aim by trying to resolve one's doubt (W. Lee 2026, §3.1). Lee's view is thus a claim not about the aim of inquiry but about the state of mind an individual inquirer is in while inquiring. Following what Lee calls standard in the debate over whether belief is compatible with inquiry, the view takes inquiry to involve an interrogative attitude such as wonder or curiosity, and counts an activity carried out without one as ersatz rather than genuine inquiry, while allowing that it may count as inquiring in more liberal contexts (W. Lee 2026, §1.1). The view that inquiry involves an interrogative attitude, and the distinction of ersatz inquiry, are also discussed by Friedman (Friedman 2013; Friedman 2019).

What this book takes from Lee is that trying and succeeding are distinct. The two inquirers in the rain example differ in success, yet both were inquiring. This supports the distinction between activity and success drawn in this section. Lee's trying, however, is an individual's mental activity sustained by doubt, and applied to an artificial system, which the paper does not discuss, whether its activity counts as genuine inquiry by Lee's standard would turn on whether it has an interrogative attitude (Chapter 4, Section 9), a question of the kind Chapter 2, Section 2 raises about epistemic attitudes. An attempt in this book's sense does not necessarily require one. Because it can also be grounded in a group's plan, an artifact's organized function, or a component's role, it does not require an interrogative attitude of every participant. An activity that would not count as genuine inquiry by Lee's standard can therefore still be an attempt in this book's sense. The two answer different questions: Lee asks what it is for an individual to be inquiring, and this book asks toward what kind of achievement an activity is organized. Nor does this book settle which of Kelp, Beddor, and Friedman is right. Its attempt is not a claim about whether inquiry in general has a constitutive aim but a description of what a particular activity is organized toward. What it takes from the debate is a reason to keep the activity's orientation, the norms by which it is assessed, and a participant's occurrent state of mind apart.

Even so, which view one takes changes what kind of attempt an activity is described as being, because the views describe differently what the activity sought to bring about and what it achieved. An example shows what changes and what does not. Two of the positions in this disagreement specify what inquiry aims at: the view, which Kelp defends, that inquiry aims at knowledge, and Beddor's view that it aims at making credences as epistemically valuable as possible. Either can therefore describe what a single episode aimed at and achieved in terms of what inquiry aims at. The descriptions of the same episode from Kelp's and Beddor's views are therefore given in turn below.

The episode is one in which a mathematical proof was checked in a proof assistant. In 2020 Peter Scholze asked for a formal verification of a theorem proved jointly with Dustin Clausen, in a blog post that launched the Liquid Tensor Experiment. Scholze wrote that for a theorem of such foundational importance, one that others would use as a black box, being 99.9% sure was not enough, and that, thinking nobody else had dared to look at the details of the proof, Scholze still had some small lingering doubts. Half a year later, after the part of the argument Scholze had been unsure about had been checked in the Lean proof assistant, a follow-up post reported that its author had no remaining doubts about the correctness of the main proof (Scholze 2020, §6; Scholze 2021). On Kelp's view that inquiry aims at knowledge, the check is described as inquiry seeking to know that the main proof was correct, and its success is judged by whether knowledge was thereby gained. For whom that knowledge was new is taken up in Section 6. On Beddor's view, on which inquiry aims at making one's credence in the answer as epistemically valuable as possible, and if the 99.9% Scholze wrote is read as the writer's own credence, the same check can be described as inquiry that continued while the credence that the main proof was correct was already high but short of full confidence. The episode has the shape of Beddor's own cases, in which someone who may already know is offered decisive evidence (Beddor 2024, §2). On Beddor's view, even if Scholze already knew before the check that the main proof was correct, a possibility like the one Section 6 leaves open for the theorem, the aim of inquiring whether the main proof was correct had not yet been attained, because a credence of 0.999 in a truth is less accurate than a higher one; the check therefore had a point with respect to that aim. What the check achieved is then described by how far it brought that credence toward the truth: if the main proof is correct, raising the credence toward full confidence made it more accurate. The size of that rise is what confirmation theory calls confirmation as relevance, confirmation as the raising of probability by evidence; on confirmation as firmness, confirmation as sufficiently high probability given the evidence, the claim that the main proof was correct was already confirmed before the check (Chapter 2, Section 5.2). Whether Scholze's writing that no doubts remained marks full confidence, and so the attainment of the aim, or only a credence closer to it, is a further question of interpretation. What does not change between the two descriptions is that the activity was organized to check the part of the main proof about which Scholze had been unsure. This book counts the check as an attempt not because of what epistemic achievement it brought about but because the activity was organized in this way, toward the part of the proof that needed checking. On either view, then, the check is an attempt, but each view interprets it as a different attempt: on Kelp's view, an attempt to bring about new knowledge; on Beddor's, an attempt to make an already high credence as epistemically valuable as possible; and the description of what it achieved changes accordingly.

Separating activity from success also blocks the inference in the other direction. Just as a failed activity can be research, an output that looks like success does not show that research took place; a lucky output does not settle the attribution. If, for example, a random generator emits a novel true mathematical statement, a true inscription may exist before anyone recognizes or supports it. If another actor proves, checks, and incorporates the result, the generator may be causally indispensable to a later production of knowledge. It does not follow that the generator knew the result, conducted the research episode, or possessed a research capacity. Chance can operate within research—through stochastic search, unexpected observation, and serendipitous redirection—when noticing, appraisal, retention, testing, and revision give it an organized role (Copeland 2019).

A research episode is therefore distinguished here from a stable capacity to conduct research. One low-probability or unrepeatable episode may be research, but one fortunate output gives little reason to call its source a researcher. Conditions such as reliability across relevant variation, responsiveness to changing evidence, integration of research functions, and revision after failure may support that stronger attribution. These conditions belong to the analysis of capacity attribution in Chapter 7, Section 3.1 rather than to the minimal work performed by attempt.

Distinguishing episodes from capacities raises the question of what one episode is and where it is bounded. Neither attempt in this book's sense nor talk of research episodes presupposes that research comes in bounded units. Where an inquiry begins and ends, and what counts as one attempt rather than several, can be fixed by practices that organize research—in research as now organized, a paper, a grant, a thesis, a release, the numbering of a notebook's entries (Chapter 5, Section 6.3)—or by the purpose of whoever describes it rather than by the inquiry itself. In one open drug-discovery project, for example, the season in which test larvae could be used and the departure of a researcher, which left one person able to run the screen, set a date by which new compounds had to be shipped; the same researcher's thesis deadline prompted a proposal that a shared manuscript enter that thesis unfinished, while one contributor decided to make further compounds anyway (Open Source Mycetoma 2022). When this book speaks of a research episode, the boundary is drawn for the analytical purpose at hand, and an attribution that depends on how episodes are counted can be assessed only when it says how they were drawn.

Success, for its part, need not be fixed once at a single time. Just as the boundaries of an episode can be fixed by practices or by a describer's purposes, the judgment whether an attempt succeeded can be revised later. Empirical support is a case in point; seen from the side of the achievement, this is taken up in Section 4 as the point that an achievement is not a single threshold. If whether an activity was research depended on its success, that judgment too would be revised whenever support was revised. Characterizing research by attempt keeps the judgment whether an activity was research apart from such revisions of the judgment of success.

Failure, likewise, need not come down to a single verdict. A project may miss its original target yet establish a boundary condition, eliminate a possibility, expose an experimental artifact, or preserve a useful trace. It may instead leave no recoverable epistemic gain. Whether the activity was research, whether it was competently conducted, whether it answered its question, whether it made progress, and whether it produced collateral knowledge are distinct judgments (for the case where the target is knowledge, Chapter 2, Section 8.1 sets out such judgments, and Section 9.1 of that chapter takes up what competent conduct requires of an artificial system that conducts research). This section separated attempt from success so that these judgments are not merged into the single judgment whether the activity was research.

4. Bringing about an achievement is neither making an output nor crossing a single threshold

Bringing about knowledge or understanding need not create the object known or understood. An astronomical relation or past event does not begin to exist when inquiry succeeds. For knowledge claims, the familiar language of production concerns realization of the epistemic relation examined in Chapter 2, with its disputed accounts of acquisition and social possession. For understanding, a new representation can enable a grasp of dependencies, of what a phenomenon depends on and how, without creating those dependencies. In both cases, producing the representation and realizing the achievement are distinct. Changes in evidence, ability, or accessibility can enable success without themselves constituting it.

Research also constructs things that did not previously exist in that form: proofs, formal objects, concepts, interpretations, artifacts, datasets, models, instruments, classifications, samples, performances, and experimentally stabilized phenomena. The immediate product and the epistemic achievement are therefore distinguished here.

At least four relations can otherwise be conflated:

1. an activity produces an immediate output, such as a text, proof, model, dataset, prototype, instrument, interpretation, or performance;
2. the output bears content or makes a claim, pattern, possibility, or distinction available;
3. it occupies epistemic roles, for example as evidence, warrant, explanation, method, representation, memory, or scaffold;
4. an epistemic achievement, such as knowledge or understanding, may thereby be realized. Its conditions and location require an account; an output or enabling role does not establish it.

The four are separated so that the existence of an output and the realization of an achievement can be told apart: a large volume of outputs from an artificial system, for example, does not by itself show that something has been achieved. Where most such outputs are false, even one that passes a check can remain poorly supported, on the Bayesian notions applied in Chapter 2, Section 5.2. The division lets one ask of a claim about artificial research which of the four relations its evidence establishes: a generated proof text, for example, establishes an output and makes a claim available, but does not by itself establish an achievement. The four are not borrowed from a particular author; the book separates them at points where output and achievement can come apart. They therefore do not list every relation that bears on outputs and achievements: relations to a society, such as a result's being recorded, received, and preserved, are taken up separately in Section 6. The four are also relations, not mandatory chronological stages. A proof can establish a theorem, explain why it holds, reveal a dependency, and transmit a strategy. A prototype can demonstrate possibility, expose a failure mode, and reorganize a design space; it can also establish feasibility while its mechanism remains misunderstood. A dataset can support a claim, calibrate a model, preserve observations for unanticipated questions, or fail to support its initial interpretation.

Why it matters to keep the four relations apart can be seen concretely in domain-specific discussions. For data and design, process-oriented accounts of data and research-through-design discuss why artifacts can scaffold inquiry without their bare existence settling whether knowledge has been produced (Leonelli 2023; Zimmerman, Forlizzi, and Evenson 2007; Gaver 2012). Leonelli, for example, treats data, models, and publications as situated scaffolds whose validity and evidential value must be reassessed as they are reused, and takes the achievement they serve to be what Hasok Chang calls active knowledge: an ability to perform and coordinate activities purposefully, which cannot be reduced to storing and retrieving information (Leonelli 2023, 46, 52–53). In terms of the four relations, the existence of an output, a body of data (the first relation), does not fix how it works as evidence (the third), still less that an achievement has been realized (the fourth), since that working is reopened each time the data are reused. In mathematics, Rav distinguishes the conceptual proofs of mathematical practice from formal derivations and holds that the path a proof charts between concepts not only convinces but establishes interconnections that illuminate mathematical ideas (Rav 1989, 67, 73 n. 30). Pelc, who adopts Rav's terminology, reads Rav as treating proofs both as ways of convincing that a theorem is true and as repositories of mathematical knowledge, in which one learns new ideas, concepts, and strategies that can be taken over into one's own research (Pelc 2014, 2–3). In terms of the four relations, what is disputed here is whether mathematical knowledge (the fourth relation) lies in a theorem statement's being made available (the second) or also in the proof, which on Rav's view carries methods and concepts and so does more than work as warrant for the theorem (the third). Thurston emphasizes concepts and ways of thinking as part of mathematical advance (Thurston 1994). On this view, advance is not exhausted by new statements' being made available. In engineering, a method description, a method enacted by skilled practitioners, and a situated sociotechnical method-in-use are not interchangeable (Stacey et al. 2025). A method description may exist and its content be available (the first and second relations) without an ability to enact the method having been realized. In art, Scrivener argues that the art object is not itself a form of knowledge, although it can communicate some: Scrivener defines a knowledge artifact by the intention to communicate knowledge, without implying that knowledge is stored in it (Scrivener 2002, 1, 3, 11). In terms of the four relations, a work's carrying content and playing the role of conveying knowledge (the second and third relations) differ from the work's itself being knowledge, just as an instruction manual is made to convey knowledge although the knowledge is not stored on its pages. A truthful description of an experience may also fail to preserve the ability or sensibility through which the experience mattered. Scrivener's own proposal for art, which centers on what it calls apprehension, is examined in Section 8, which compares the orientation and its distinctions with practices presented as research by their own theorists.

What these discussions bring out is that correct description, warrant, possession, execution, and transfer can come apart. Evidence that an artificial system has output a correct theorem statement or method description therefore does not by itself establish warrant by proof, that the system or anyone who reads the statement possesses the knowledge, or an ability to enact the method. Keeping the four relations apart lets one ask which relation such a claim has established without settling these disputes. These literatures do not settle whether proofs or methods are themselves knowledge, although Rav's account, as Pelc reads it, treats proofs as repositories of mathematical knowledge, and Leonelli's account locates the achievement an output serves not in knowing a proposition but in an ability to perform and coordinate activities. In this book's terms, such an ability is a candidate for an epistemic achievement other than propositional knowledge. As the achievement level of Section 1 noted, however, abilities can be outputs, means, or achievements depending on the claim, so whether a given ability counts as an achievement has to be specified claim by claim. Whether knowledge includes abilities is among the questions Section 5 registers, and taking propositional knowledge as a starting point, as the opening of Chapter 2 notes, does not settle it.

Finally, bringing about an achievement in this sense does not imply one irreversible threshold at which a paper is stamped as knowledge. Section 3 noted that the judgment whether an attempt succeeded can be revised later. The achievement itself need not be fixed at a single time either. Partial proofs, calibration, replication, criticism, theoretical integration, and later correction continuously alter support, confidence, usability, and the space of live alternatives. For knowledge, since on many accounts whether knowledge obtains depends on the state of support and on the presence or absence of defeaters, contrary information that removes support (Chapter 2), a change in support can change whether knowledge obtains. On Dewey's account, each stage of reflective inquiry presents what may be taken for granted as the outcome of previous thinking and as determining further inquiry, and the intellectual content so reached is tentative, corrected as it is acted on (Dewey 1916, 164, 241). Understanding, moreover, is sometimes treated as coming in degrees. On Kelp's account, for example, maximal understanding of a phenomenon is maximally comprehensive and maximally well-connected knowledge of it, and understanding short of that is a matter of degree (Kelp 2021b, S1588; see also Chapter 3, Section 7). On such accounts, bringing about understanding can be a matter of understanding's deepening rather than of crossing a single threshold. Bringing about an achievement therefore does not imply that every project crosses one discrete threshold.

5. Epistemic achievement requires making clear what was achieved

As Section 1 noted, the term epistemic achievement in the working sentence is an open general term that includes knowledge and understanding. Because it does not decide in advance which kind of achievement is sought, the characterization applies both to research directed at knowledge and to research directed at understanding. Leaving the kind open does not, however, lighten the burden of any particular claim. To claim that an activity has achieved, or is trying to achieve, something is to need to make clear which kind of achievement it is and under what conditions that achievement obtains. This section looks concretely at what that need demands in the case of knowledge, because the success conditions of knowledge have been discussed in detail in epistemology and many accounts treat them as demanding (compare Section 1). Chapter 3 supplies the corresponding inquiry into understanding. What is listed here are conditions of knowledge, not conditions of every epistemic achievement.

Where an activity is claimed to be an attempt to bring about new knowledge, that is, where the kind of achievement sought is specified as knowledge, the word knowledge cannot remain an honorific for praising a result. Using it takes on the responsibility of answering philosophical questions such as these. What distinguishes a true result from knowledge? Must a subject believe it? What kind of evidential, reliable, modal (concerning whether the belief would still have been right had things been slightly different), agential, or social relation is required? Does knowledge include abilities and understanding? Can a collective know what no single member currently entertains?

The present chapter does not answer these questions in advance; Part II examines them by comparing rival accounts. The chapter lists three issues that anyone claiming knowledge will face, whichever account they adopt. First, knowledge is treated as world-directed and factive: as Ichikawa and Steup report, most epistemologists agree that what is false cannot be known (Ichikawa and Steup 2026, §1.1); on this view, a false claim does not become knowledge even if it is widely accepted in a society and that acceptance is stable. Second, there is the question whose knowledge it is. The standard analysis in epistemology attributes knowledge, in the form "X knows that p," to a subject rather than to an unowned proposition; the subject may be an individual or, as many accounts in social epistemology allow, a group. On accounts derived from Popper's epistemology without a knowing subject (Chapter 2, Section 6, Table 2.2), by contrast, whether a result is knowledge turns on certification and public acceptance rather than on anyone's attitude. Third, a lucky true belief or output appears to lack the relevant non-accidental epistemic standing: a system that hits on a correct answer by chance does not thereby know it. This book does not aim to settle, on these issues, what truth is, who can be a knowing subject, or what relation excludes luck.

Not every account of science, moreover, uses knowledge in a sense that requires truth. Hoyningen-Huene, for example, uses scientific knowledge as scientists do, for established belief that the relevant community widely holds and does not treat as tentative or refuted, and leaves open whether it is true (Hoyningen-Huene 2018, 13). Hoyningen-Huene's thesis introduced in Section 1, that scientific knowledge differs from everyday knowledge chiefly by being more systematic, concerns knowledge in this sense. That thesis therefore does not supply a success condition of the kind examined in this section.

Contemporary epistemology offers competing accounts of how to answer the three issues listed above. They draw on different resources: belief, evidence, accessible reasons, reliable process, safety, competence, ability, freedom from defeaters, social criticism, and knowledge-first explanation, among others. One source of the competition is the Gettier cases. Gettier presented cases of justified true belief that hardly seem to be knowledge, challenging the analysis of knowledge as justified true belief, but no single replacement analysis has been established. Knowledge-first approaches reject the assumption that knowledge must be reduced to independently specified components at all (Ichikawa and Steup 2026). Williamson, for example, as Ichikawa and Steup present his view, treats knowing as itself a mental state, the most general factive one, and holds that it is a mistake to analyze knowledge in terms of more fundamental epistemic notions (Ichikawa and Steup 2026, §11).

Which account one adopts places different burdens on those who build artificial research systems and on those who evaluate them. On an account that stresses reliable process, for example, one has to show that the process by which the belief in question was formed is reliable; when a human comes to believe a result through a system, that process includes the system's production of the output. On an account that stresses evidence, one has to show that whoever is claimed to know has evidence that supports the result. Part II neither decides which rival account is correct nor combines them into one checklist; it examines what comes into view about artificial research on each. The book's own working characterization of research, unlike the rival accounts it leaves standing, is itself open to revision. If treating central research practices as attempts at epistemic achievement requires distorting their own distinctions or calling every desirable change epistemic, the content or scope of the working characterization needs revision.

6. New to S makes the unstated reference explicit

Novelty concerns the reference society's prior epistemic situation. For knowledge, what one person knows another may not, and what a specialist community knows another may not. For understanding, an individual can newly grasp something already understood in the community. Neither change is socially new merely because it first occurs in one investigator. A claim of novelty specifies which achievement is new, in what respect, and to which society.

The motivating observation is modest. In ordinary academic practice, learning an established theorem does not by itself become a new research contribution. Independently rediscovering what a field already knows may demonstrate ability or provide an independent check, but novelty cannot be assessed only from what the investigator previously believed; it has to be compared with what the field already knew. In the Liquid Tensor Experiment, for example, a complex that Johan Commelin constructed while formalizing a proof turned out, several months later, to be a construction of MacLane's, while a lemma Commelin proved about it was, as far as its prover knew, new (Commelin 2022). Conversely, a first result unknown to the relevant community can be a contribution even though only a small subgroup initially understands it. Talk of new knowledge therefore presupposes, without stating it, a social reference: the society to which the knowledge is new. Being new to that society does not, however, require every member to update their beliefs at once.

The working sentence's variable S makes that reference explicit and generalizes it. It can denote humanity, a specialist community, an organization, a mixed group, or a candidate artificial society. It is minimally a collectivity for which the relevant comparison of epistemic situations is meaningful. The comparison is meaningful only if S's epistemic situation, through its members, organization, or records, could come to include an achievement of the kind compared: if it never could, such an achievement would not be new to S but simply beyond its reach, and an attempt could not be directed at S (Section 2). The notation by itself does not require a corporate mind, institutional acceptance, universal accessibility, or one theory of group knowledge. Competing accounts, grouped into families in Table 2.2, may yield different verdicts (Chapter 2, Sections 6 and 7).

Making an unstated reference explicit in this way has a precedent in the debate over scientific realism. Van Fraassen, for whom what is observable is what is accessible to the unaided human senses, holds that observable, like portable, is nonetheless indexical (its reference shifts with the context of use, as that of here or we does): it means observable by us, the epistemic community, and the range of beings counted as us is "not fixed, either necessarily or even historically." Epistemology, van Fraassen argues, must therefore consider futures in which we and humans do not coincide, as when dolphins or extraterrestrials are admitted to the community (van Fraassen 2005, 1–4, 7). This us fixes what is observable rather than what is new, and van Fraassen does not discuss research. What this treatment of observable shares with S is that a reference left unstated in ordinary use is made explicit and allowed to range beyond humans, so that who belongs to it becomes a question of its own; Chapter 8, Section 1.2 takes up what admitting members does to a mixed S.

This openness has limits. An arbitrary set assembled after the event cannot automatically become S merely to make an output novel. Members, specialist subgroups, records, or practices are required to be connected enough for the social comparison under consideration to make sense. What kind and degree of connection is required belongs to social epistemology and to the practical role the achievement claim is meant to play.

Nor does the notation imply that knowledge appears socially at one moment. A result can be produced, recorded, appraised, preserved, made retrievable, taken up, used, corrected, and withdrawn through different events (compare the relations distinguished in Chapter 2, Section 6). An unread result may make knowledge possible without yet being knowledge possessed by S, depending on the account. The fuller social-epistemological analysis belongs to the achievement part of the book.

Whether something is new to S is settled by comparison with what S already knew. Results obtained by deduction can also be new by this comparison. Because mathematical theorems follow logically from axioms, on an idealization taken up at the end of this paragraph a community that knows the axioms would already know the theorems that follow from them. But a result can follow from premises already known within S without anyone there yet knowing the result. That a conclusion follows logically from premises is a relation among propositions; it is not the same as someone's actually carrying out a derivation and moving from not knowing to knowing. A principle much discussed in epistemology, called closure, holds that one who competently deduces a conclusion from what one knows comes to know the conclusion. Not every epistemologist accepts it in this general form, but even granting it, what it says is that knowledge extends when the deduction is performed; it does not credit a subject in advance with every consequence not yet deduced. Nor does the existence of a derivation, however long or inaccessible, by itself put S in a position to know its conclusion. To be in a position to know is not yet to know but to be able to come to know by means within reach. Chapter 2 (Sections 4.4 and 7) distinguishes actually knowing from being in a position to know in this sense. This book applies that distinction to the actual epistemic situation, to who has done what and what is available, rather than inferring it from what can be derived. A proof can therefore produce knowledge new to a mathematical community even though its conclusion follows from accepted axioms. Whether a particular proof does so still depends on its warrant and on what the community already knew. An evaluation that tracks this distinction separates such a new derivation from a system's re-deriving a result already known within S but withheld only from the system being evaluated, that is, from a rediscovery that is new only to the system (taken up at the end of this section). An idealization such as logical omniscience, on which a subject already knows every logical consequence of what it knows, would erase this distinction.

Returning to Scholze's check from Section 3, what a check of a proof made new depends on at least two things. One is whether the relevant community already knew the theorem, which the accounts of social knowledge in Chapter 2, Section 6 answer differently. What was new may also have been knowledge that a particular argument was correct rather than knowledge of the theorem. The other is which account one takes of whether one can know something while some doubt remains. On Williamson's knowledge-first account, one can know something without being ready to stake much on it, and in mathematics one knows, to a first approximation, what one has a proof of (Williamson 2000, 86, 209, 251–252, 263–264). Williamson does not discuss this case, but applying the account, the 99.9% Scholze wrote can be read as describing the writer's own credence, leaving open that Scholze already knew the theorem before the check. If so, the check need not have given Scholze new knowledge of the theorem. Both points concern not whether the check was an attempt but what was new, and to whom.

Novelty to S can be asked about understanding in the same way, but what is compared depends on how understanding is conceived. On accounts that treat understanding as a kind of knowledge (such as Khalifa's, taken up in Chapter 3, Section 3, and Kelp's, Section 4 above), novelty in understanding is itself a question of novelty in knowledge: whether the relevant knowledge, for example of why a phenomenon holds, of its explanation, or of how the facts about it connect, was already present within S. Even on Kelp's account, however, understanding a phenomenon requires knowledge that is comprehensive and well connected, so whether pieces of such knowledge were present within S does not by itself settle whether an understanding is new (Kelp 2015, 16). On accounts that treat understanding as an achievement distinct from knowledge, the question is whether the relevant grasp, for example of dependencies or explanations, was already present within S. On either kind of account, that a result such as a theorem was already known does not by itself settle novelty in understanding, as when a new proof of a known theorem discloses dependencies that were not previously understood; a proof that is merely written differently does not show this. Whether newly available explanations or resources enabling understanding already amount to understanding credited to a collective or to science, or must be separated from understanding actually realized in its members, also differs among accounts (Chapter 3, Section 6). On accounts that separate them, just as the existence of a derivation was separated above from actually knowing its conclusion, a society is not treated as already understanding everything it could come to understand.

These comparisons make a concrete difference where artificial systems are involved. Suppose an artificial system comes to understand a proof already understood in the human mathematical community (whether artificial systems can understand is taken up in Chapter 3, Section 6). A new bearer of understanding does not alone make the mathematical understanding new to that community. Building that system may nevertheless yield new knowledge about artificial understanding, for example about the mechanism by which a machine comes to understand the proof. What is new then is not the mathematical understanding but knowledge about artificial understanding, and the research aims at a different achievement from that of mathematical research. By contrast, where the system grasps a dependency the human community did not know, there can be novelty for the community, which turns on that dependency and on what was already known or understood there. This does not yet mean that the community itself has acquired the understanding: if humans cannot access the system's internal state, the community's grasp may be unchanged. Whether an explanation has become available to the community, whether members have taken it up, and whether the collective understands each add different conditions, depending on the account adopted. What was grasped, whether it is new to S, and whether S itself has acquired it are therefore distinct claims. If S is instead an artificial society of which the system is a member, the system's understanding of a proof that human mathematicians already understood can itself be new to S when no other member understood the proof, and whether S has then acquired it turns on whether other members can access or share it, which accounts again settle differently (Chapter 3, Section 6). If a novelty claim is mistaken, what fails is the claimed success of the contribution, not the fact that research was attempted; this is an instance of the distinction between attempt and success drawn in Section 3.

A structure like that of these comparisons appears in the philosophy of creativity. Paul and Stokes, setting out how the requirement that what is creative be new has been refined, note that Newton's calculus was creative even if Leibniz reached it first, and that a beginning student's idea that freedom is compatible with determinism can be creative although philosophers have defended such views for millennia; in Margaret Boden's terms, such ideas are psychologically creative, new to their creator, without being historically creative, new in all of history, and psychological novelty is the more fundamental, since whatever is new in all of history is new to its creator but not conversely (Paul and Stokes 2023, §2). Boden's distinction concerns ideas new to the individuals who have them, not novelty for a reference society. Applied to novelty for S, what carries over is that assessing novelty relative to a restricted domain is coherent in its own right, with novelty in all of history a special case of it rather than its source; that the domain is S rather than an individual rests instead on S's being the collectivity for which the comparison of epistemic situations is meaningful. Read this way, Boden's distinction gives a reason to think that assessing what is new relative to a chosen S, rather than to whatever is known anywhere, is a coherent choice of scope rather than a derivative of novelty to the world. In the cases above, an achievement can then be new to a human S although an artificial system outside S already holds it, much as Newton's calculus would have been new to Newton even had Leibniz reached it first. Where the machines are themselves members of S, as in a mixed S, whether they already hold an achievement bears instead on whether it is new to S, as with the artificial S discussed above.

Judging whether something is new to S also becomes less straightforward when S has lost something it once knew, since the question arises whether recovering lost knowledge is new to S or not. Chang argues that specialist science, which must take much for granted and cannot pursue every question, can lose knowledge, and that rehabilitating discarded or forgotten knowledge is a form of knowledge creation (Chang 2021, 108, 110). One of Chang's examples is Pictet's experiment on the apparent radiation and reflection of cold. Chang treats confirming that such a reported phenomenon can be reproduced as recovering a piece of forgotten scientific knowledge, and notes that a 1985 replication of Pictet's experiment itself seems to have been largely neglected (Chang 2017a, section "Recovery"). In this example, the result of Pictet's experiment was once scientific knowledge and was later forgotten. Does unearthing the experiment now and confirming its result again bring knowledge new to the community that forgot it? The same question arises when an artificial system unearths forgotten literature or experiments. The answer depends on how far a society is taken to retain knowledge, that is, on which account of retention one adopts (Chapter 2, Section 7). On Bird's view, a record still available through normal channels remains the community's knowledge, so as long as it is so available, unearthing it adds nothing new to the community's knowledge. On Lackey's view, in the form Kallestrup develops from it, such a record leaves the community only in a position to know, so unearthing it restores knowledge to the community and may in that sense count as new. Where S is artificial or mixed, the specialization Chang describes can remain while its records persist within reach of its members; knowledge set aside then need not have been forgotten, so that on Bird's view returning to it adds nothing, while on Lackey's it can still restore knowledge that no member held. On either view, this recovery differs from a system's rediscovering a result withheld only from the system under evaluation: there the result has been known within S all along, and it is new only to the system.

A short story by Ted Chiang poses a further case, in which a record has entered S's channels while its content has not. It is written as an editorial in a journal of a future in which "metahumans", humans whose brains were altered by gene therapy before birth, dominate research and make their findings available only through a digital neural transfer that unaltered humans cannot use. Journals for human readers publish second-hand translations that leave "even the most brilliant humans" puzzled, and human scientists, expecting to make no further original contribution, turn to interpreting metahuman publications, reverse engineering metahuman artifacts, and remotely sensing metahuman facilities; the editorial counts this hermeneutics as a legitimate method that increases human knowledge "just as original research did" (Chiang 2000, 517). The piece is fiction, and its metahumans are not machines. Read as a thought experiment, it separates three things: the results are realized in bearers outside the human S, records of them stand in S's journals, and no member can yet follow those records. Applied to this case, the views of retention above place the novelty of a successful decipherment differently. If a record that no member can follow still counts as available through normal channels, Bird's view, extended from the retention of a result once read and accepted to such a record (Chapter 2, Section 7), would make its content the community's knowledge already; on Lackey's view, the record would leave the community at most in a position to know. And on the sense of being in a position to know used in this section, being able to come to know by means within reach, a record that no member can follow by such means does not put S even in that position. On the latter readings, a decipherment can be new to S although its result is not new to those who hold it.

7. What follows comparatively directly from the working characterization: being directed at something unsettled

No universal method or procedure that research must follow can be derived from the working characterization. By comparison, one implication follows comparatively directly: an attempt at an epistemic achievement new to S addresses something not yet epistemically settled in the relevant respect, for example, what is the case, why it is so, or how a target is to be grasped. The working sentence's "new to S" specifies the epistemic achievement sought as one that S's epistemic situation does not yet include, however that situation is construed (Section 6), and "attempt" requires that the activity be directed at that achievement; together they direct the attempt at something not yet settled for S. For knowledge, the attempt is directed at something not yet known in S; for understanding, at something not yet understood there. Since S's epistemic situation could come to include an achievement of that kind (Section 2), what is not yet settled for S is something that could become settled there, even if the achievement, once realized, is at first held only by A or by some other bearer. That something is not yet known or understood also means that a question can arise about it. This book therefore calls such directedness at something not yet settled the minimal question structure. It does not require a verbal question formulated at the outset, nor ignorance of every fact concerning the target. Understanding can remain unresolved even when some facts about the target are known. Nor does it require that the point at which the attempt is directed be in fact unsettled in S: what follows here is where the attempt is directed, not what S's actual state is. If the achievement aimed at had already obtained in S, then, as Section 6 noted, what fails is the novelty claim, not the fact that research was attempted. Which activities belong to the attempt depends on their organized connection to that prospective achievement.

This does not mean that every project begins with one explicit interrogative, that one actor entertains it, or that the final question existed unchanged at the outset. How questions take shape and are recast within inquiry is examined in Chapter 4, Section 4, in light of research records and accounts of inquiry. Here it is first made clear in what sense the minimal question structure is a question.

The minimal claim concerns directed epistemic unresolvedness, not a fixed question represented in one mind. The word question is itself used for an interrogative sentence, for its content, and for the act of asking (Ciardelli 2022, 16 n. 3; Habgood-Coote 2026, 2). Accounts of the act differ in what they require of the asker. Watson takes the third use and identifies a question, as an act, by its characteristic function of eliciting information, so that a question need not take interrogative form, and one asked with some other explicit goal, such as embarrassing the hearer, can still be a genuine question (Watson 2015, 86–102). Habgood-Coote calls default the view that to ask a question is to express an interrogative attitude, to aim at acquiring its answer, and to direct the hearer to supply one, and argues that this view does not fit three kinds of asking: a teacher asking a class a question to which the teacher knows the answer; a speaker asking a question in order to answer it, as Frederick Douglass did in an 1852 address; and shared asking, as when David Hilbert, at the International Congress of Mathematicians in 1900, set out problems in the hope that the mathematical community would solve them (Habgood-Coote 2026, 2, 4–6). What identifying an act of asking requires and does not require on the two accounts, and what view Habgood-Coote proposes from there, are taken up in Chapter 4, Sections 1 and 6.

The question structure of this section need not yet be a question in any of the senses of interrogative, content, or act of asking. Watson's account is close to it in separating a question from interrogative form and from the asker's conscious motive; but what Watson identifies is an act, one with the function of eliciting information. The structure of this section is not an act; it is a point at which an attempt is directed as something not yet settled in S, which interrogatives may later express, whose content may become determinate only during inquiry (Chapter 4, Sections 3 and 4), and which acts of asking by one or several participants may pursue. The minimal claim therefore does not depend on locating such an act in A or in any single member of S.

Nor does the minimal claim fix where a question comes from. A question may arise within the inquiry of A, or be inherited from those who supervise or fund the work, from communities, instruments, or a discipline's history (Chapter 4, Section 1); whether the actor originated it bears not on whether research is attempted but on autonomy (Chapter 7, Sections 1 and 4.3). This structure alone therefore does not settle in what sense asking can be attributed to a machine, whether the content of a question is fixed from the outset, who holds it, or for whom it is unknown. The three chapters of Part III view the process of inquiry from three sides, the relation between an inquiry and the unsettled points that organize it, the relations among candidate answers, representations, and means, and the relation between a result and what it bears on; of these, Chapter 4 asks what it means for a machine to ask, along four points: in what sense it is a question, whether its content is fixed from the outset, who holds it, and for whom it is unknown.

8. What does not follow

The working characterization does not establish that every research activity:

• begins with a fully formulated hypothesis;
• represents a determinate answer space;
• separates discovery from justification (treating the process of hitting on a hypothesis and the process of giving grounds for accepting it as separate stages);
• performs one procedure called verification;
• produces explanatory understanding;
• publishes a paper or makes its result immediately available;
• is governed by one actor;
• is valuable, important, or progressive;
• or contains a researcher with maximal autonomy.

Nor does it prove that knowledge, or epistemic achievement generally, is the only aim of research, that epistemic purposes always override practical, moral, political, aesthetic, or transformative purposes, or that epistemic orientation uniquely separates research from every neighboring activity. The formulation identifies the target conception whose construction consequences are examined here.

The characterization also leaves substantially open the choice of purposes, what research is pursued for, and the choice of agenda, which questions are taken up and in what order. A society is ignorant of indefinitely many facts, explanations, constructions, and possibilities. Novelty does not decide which deserve attention. Question selection can reflect, for example, explanatory interest, expected epistemic gain, practical need, moral concern, feasibility, institutional priorities, opportunity cost, and distributions of risk and benefit, among other things. On Kitcher's account, what makes a question significant depends on characteristics of people, their aims and values, and their social organization, and answering questions that express human curiosity is one way of conceiving significance (Kitcher 2013, 216–217). Other positions, set out in Chapter 9, Section 3, hold that the value of knowledge is independent of S; on them, that value is not, or not only, relative to S's purposes and interests. They come in two versions, whose consequences for choosing among S's unknowns differ. On one, which adopts a principle that Grimm attributes to Kvanvig and Lynch, every true belief has prima facie value; this by itself neither makes their values equal nor orders S's unknowns, so it leaves the choice among them unsettled. On the other, which Kitcher calls the traditional approach to epistemic significance, the structure of the world fixes which truths are significant, and choosing among S's unknowns would reduce to finding what is significant. On neither version does novelty to S decide the choice: the first supplies no ordering, and the second orders the unknowns by the structure of the world, a ground other than novelty. This book locates importance at the fourth level distinguished in Section 1, that of value and governance, rather than treating it as a hidden condition of knowledge or research status.

Comparing the orientation and its distinctions with varied research practices

As this section noted, the working sentence identifies the target conception whose construction consequences are examined. Section 5 also stated when it would need revision: if treating central research practices as attempts at epistemic achievement required distorting their own distinctions or calling every desirable change epistemic. The book does not aim to test the working characterization. Even so, comparing the characterization with practices presented as research that do not fit this conception easily, to see how far it reaches and where it parts from other ways of taking research, shows concretely where the revision condition could become an issue. It also shows concretely, not only by definition, what would and would not enter the target if an artificial system were built and evaluated by this conception.

This section takes up three such practices: artistic research, participatory action research, and research within the Indigenous paradigm that Shawn Wilson describes. Each is presented as research by its own theorists, and their accounts put the book's commitments under pressure. These two conditions are used so that the characterization is compared with practices its own terms have not already screened out; if practices were chosen because they fit those terms, where it parts from other ways of taking research would be harder to see.

The three practices bear on different commitments of the characterization. The comparison asks three questions. First, on its theorists' own account, does the practice include an organized orientation toward an epistemic achievement new to a relevant society, as the working sentence's "attempt" and "new to S" require? Second, what would constitute that achievement, as distinct from a mere output or enabling condition (Section 4)? Third, can achievement be analytically distinguished from value and relationship in the way proposed here? These are this book's questions, not neutral terms binding the practices examined.

The artistic case bears mainly on the first two questions. It lets us separate what follows from the working sentence's epistemic achievements including understanding as well as knowledge, namely that the understanding that institutional definitions of artistic research pair with knowledge is not excluded from the start, from questions that remain: whether orientation toward an achievement captures the openness to what is not yet known that Borgdorff describes (the first question), and whether the object of what Scrivener calls apprehension is, in the terms of Section 4, a product (the second). Action research bears on the third question, asking whether epistemic achievement can be separated from worthwhile action. Wilson's relational account bears both on the third, the separation of achievement from relationships, and on attribution. The cases do not exhaust such pressures or form a consensus, but together they bear on all three questions above and on the location of attribution, which the relational account places differently. Accounts of understanding are examined more fully in Chapter 3, and the roles of outputs in Section 4 and Chapter 5; here the issue is the characterization of the practice.

The institutional definitions to which the debate over artistic research has responded pair knowledge with understanding (Borgdorff 2011, 54–55, 63 n. 20; Scrivener 2002, 2). Borgdorff holds that in artistic research the creative process is the pathway, or part of it, through which new insights, understandings, and products come into being (Borgdorff 2011, 46), and sets out criteria on which it counts as academic research, among them that it seeks to contribute to what we know and understand and to add new knowledge or understandings to the existing corpus (Borgdorff 2011, 54–55). Borgdorff's chapter, as announced at its outset, also characterizes artistic research as the articulation of pre-reflective, non-conceptual content, content prior to reflection and to concepts, enclosed in aesthetic experiences, enacted in creative practices, and embodied in artistic products, and concludes that it does not really involve theory building or knowledge production as those terms are usually understood but is "more directed at a not-knowing, or a not-yet-knowing" (Borgdorff 2011, 44, 47, 59, 61). Scrivener argues that the art object can communicate some knowledge but is not itself a form of knowledge, and that what art offers are possibilities and ways of seeing rather than conclusions (Scrivener 2002, 1, 7, 11–12). Scrivener accordingly proposes that research in art be defined as "original creation undertaken in order to generate novel apprehension," where the novelty is cultural, not merely novelty to the maker or to individual viewers, and treats what the art community chooses to call research as its own decision (Scrivener 2002, 2, 12–13).

Borgdorff reports that many in the debate view knowledge not primarily as justified true belief or warranted assertibility but as world disclosure or world constitution (Borgdorff 2011, 56; compare 61). That the working sentence's epistemic achievements include understanding does not thereby establish a match with Borgdorff's account. Among the views it reports in the debate is one that treats understanding as a form of knowledge in which theoretical knowledge, practical knowledge, and acquaintance (knowledge by being directly acquainted with something) may intersect. For artistic research, Borgdorff adds insight and comprehension to emphasize a perceptive, receptive, and verstehende engagement, one that seeks to grasp meaning from within, that is often more important to the research than getting an explanatory grip (Borgdorff 2011, 55). Whether the accounts compared in Chapter 3 capture that understanding remains a substantive question. Nor does that inclusion establish that every insight, world disclosure, or articulation Borgdorff describes meets an account of epistemic success. Borgdorff's account still leaves the question whether orientation toward an achievement adequately captures openness to what is not yet known. Section 7 drew from the working sentence a directedness at something not yet settled for S, but as part of an attempt to bring about an achievement there, whereas Borgdorff sets being directed at a not-knowing, or a not-yet-knowing, against knowledge production in the usual sense. Set against Section 7, what is at issue is therefore whether that openness is directed at settling what is not yet known. Among Borgdorff's conditions of academic research are that a study justify its own importance to the academic forum and that its significance extend beyond the artist's own oeuvre (Borgdorff 2011, 54, 57), considerations that this book assigns to significance rather than to research status.

Scrivener explicitly contrasts making art with investigation undertaken to gain knowledge and understanding. Apprehensions in Scrivener's sense are objects to be grasped by the senses and intellect, offering ways of seeing and being rather than knowledge of how things were or are (Scrivener 2002, 11–12). That the working sentence includes understanding therefore does not by itself answer Scrivener's proposal. Such an object is made as a work, so in Section 4's terms it is an immediate output of the activity (the first relation); whether grasping it brings about an epistemic achievement such as knowledge or understanding, and under what conditions, requires a further argument. Cultural novelty is relevant to the proposal, but does not settle that issue. The book's terms carry no authority to decide what the art community may call research (Introduction, Section 4). It leaves the fit between Scrivener's proposal and its own characterization open. Scrivener's arguments include premises about knowing that this book does not adopt, such as that knowing is experienced with felt certainty and that a claim to new knowledge must be justified by argument (Scrivener 2002, 4–5, 10, 12). The conclusion that the art object is not itself a form of knowledge is therefore a claim about knowledge in Scrivener's sense, not about knowledge in the book's sense. What the book takes from the case is not an answer to the second question, what would constitute the achievement as distinct from a mere output, but the point that showing that grasping such an object brings about an epistemic achievement such as knowledge or understanding needs an argument of its own.

Reason and Bradbury name as one primary purpose of action research the production of practical knowledge useful in the everyday conduct of people's lives, and write that it leads to "new practical knowledge" and to new abilities to create knowledge (Reason and Bradbury 2001, 2). Measured against the working sentence, their own account therefore contains its orientation, at least for the part of practical knowledge that can be stated as propositions, alongside purposes that Section 1 allows to coexist with it, such as liberation and healing (Reason and Bradbury 2001, 2, 10). Following Heron and Reason, they set out four interdependent forms of knowing: experiential knowing, through direct encounter with persons, places, or things; presentational knowing, which first expresses that encounter in story, picture, or movement; propositional knowing, through concepts and ideas; and practical knowing, knowing how. The one they treat as primary is practical knowing, which, following Heron, consummates the other forms of knowing in action and consummates experiential knowing "in worthwhile action"; in action research, they write, knowledge is "a verb rather than a noun" (Reason and Bradbury 2001, 1–2, 9, 11). Practical knowing in their sense can be separated into three layers: claims that can be stated as propositions, which can be assessed as knowledge in the sense of Section 5; skill, which this book does not decide to be or not to be a kind of knowledge and assesses as an ability (Chapter 2, Section 1); and the worth of the action, which this book assesses at the level of value. They also give an epistemic reason for participation: outcomes grounded in the perspective and interests of those immediately concerned, rather than filtered through an outside researcher's preconceptions, make for "better" research, a reason that the book, which does not adopt one account of warrant, can accommodate on accounts that give weight to the perspectives brought to bear in criticism (Chapter 2, Section 6). But they add that the characteristics of action research are not simply questions of methodology and that, "far more than that," in joining knower with known in participative relationship they undercut the foundations of the empirical-positivist worldview (Reason and Bradbury 2001, 4). Among their questions of validity and quality they include the significance of an inquiry and the relationships through which it was conducted (Reason and Bradbury 2001, 10, 12). This book keeps these at the level of value and governance as its own analytical placement; Reason and Bradbury do not accept that separation.

They also describe action research as creating new forms of understanding (Reason and Bradbury 2001, 2); whether that understanding is understanding in a sense examined in Chapter 3 remains open. Because the working sentence's epistemic achievements include understanding, assessing the epistemic orientation need not depend only on the propositional component identified above. But the book has not thereby supplied an account of practical knowing that identifies skill with knowledge or understanding, or separates its success from worthwhile action. The general term epistemic achievement leaves this substantive issue open; it does not settle the disagreement about participation and value.

In Shawn Wilson's account of an Indigenous research paradigm, relationships are reality, and relational accountability grounds its axiology and methodology (Wilson 2004, vi, 147–148). Wilson lists principles that must guide such research and what results from it, among them that the research question be founded in the reality of Indigenous experience, that the methods be process-oriented, and that the researcher take responsibility for the transformations and outcomes the research brings into a community (Wilson 2007, 194–195). Drawing on Stan Wilson, Shawn Wilson holds that knowledge is relational and shared with all of creation rather than owned by an individual researcher, and writes that the knowledge being presented belongs neither to its presenter nor to the group who took part in the research (Wilson 2003, 172; Wilson 2004, 223). The account speaks of the knowledge gained from relationships, but holds that knowledge cannot be owned or discovered: what research does is to create a new set of relationships, and bringing a new idea into being carries a responsibility to the relationships around it (Wilson 2004, 152, 209, 232). Under relational accountability, Wilson writes, "Right or wrong; validity; statistically significant; worthy or unworthy: value judgements loose [sic] their meaning," and research that meets criteria of validity and reliability, or accurately describes a "fact," will not be accepted within an Indigenous paradigm if it is separated from its relationships (Wilson 2004, 147–148, 188). The account adds that such research has its own ways of ensuring that what it says is "true" without externally imposed tests, that continuous feedback lets participants check the accuracy of the analysis, and that accuracy matters more in describing the set of relationships that make up a phenomenon than in describing the phenomenon (Wilson 2004, 188, 222–223). Research, on Wilson's account, is a ceremony and must be respected as such: "a life-changing ceremony" in which the relatedness of all things is respectfully reinforced (Wilson 2003, 175).

Wilson declines to articulate the paradigm by comparing it with other research paradigms, since any comparison would create a binary that leads to judgment and subjugation (Wilson 2007, 194). Whether research in this paradigm produces knowledge in the book's sense is not judged here; a verdict reached in its terms would in any case carry no authority for the community whose practice it classified (Introduction, Section 4). It records only three places where its terms and Wilson's part. The first is whether relations can bear knowledge: the book starts from the standard analysis, which attributes knowledge to a subject, but also notes accounts of knowledge without a knowing subject and leaves open which entities can be subjects of knowledge (Sections 2 and 5), and Wilson's denial that knowledge can be owned is not by itself a denial that knowledge is attributed to someone. The second is where considerations of relationship stand: the book does not make them conditions of research status, whereas Wilson makes them conditions of acceptance as Indigenous research (Wilson 2004, 147–148, 188). The third is how Wilson's talk of truth, accuracy, and new ideas relates to the factive knowledge of Section 5.

Read against the three questions, the cases come out as follows. On the first, whether the practice includes an organized orientation toward an epistemic achievement new to a relevant society, Reason and Bradbury's own account includes the working sentence's orientation, at least for the part of practical knowledge that can be stated as propositions; Borgdorff's criteria include it in asking that artistic research contribute to what we know and understand, though whether it captures the openness to what is not yet known that Borgdorff also describes remains open; the fit of Scrivener's proposal is left open; and the book does not judge research in Wilson's paradigm. On the second, what would constitute the achievement, the object of Scrivener's apprehension is, in Section 4's terms, a product, and whether grasping it brings about an epistemic achievement such as knowledge or understanding needs a further argument. On the third, whether achievement can be analytically distinguished from value and relationship, the book keeps the separation of achievement from value, governance, and relationships as its own analytical proposal; Reason and Bradbury contest it, Wilson's account does not separate achievement from relationships, and the book does not settle the difference. The working characterization, then, neither takes in the three practices wholesale as research in its sense nor shuts them out. In parts of its theorists' accounts the working sentence's orientation can be found, and in other parts whether it can be found is not settled.

The line falls where it does because the working sentence names the achievement with the term epistemic achievement. Since the term includes understanding, the understanding that Borgdorff's criteria pair with knowledge is not placed outside the sentence's scope from the start. The term is also a success term, as Section 1 noted: a claim to knowledge or understanding has to meet that achievement's own conditions, and a claim to any further achievement needs its own argument. The orientation is therefore found where theorists say they aim to contribute to knowledge or understanding, and it remains undecided where it has not been shown that what the practice aims at would, if realized, meet those conditions or be supported by such an argument, as with openness to what is not yet known, Scrivener's apprehension, the part of practical knowing that is not propositional, and relational insight. Restating them in the book's terms would not by itself bring them within its scope. To claim that such a restatement had brought them in would be to distort the practices' own distinctions, and Section 5's revision condition would then apply. As long as what remains undecided is left undecided, the condition does not apply. The comparison with the three practices therefore comes out as follows: the characterization does not, for now, need revision, since Section 1 left open whether the orientation is necessary for every practice properly called research; but what it covers stops there, and beyond that point it parts from other ways of taking research.

For building and evaluating an artificial researcher, this examination has two consequences. First, if an artificial system is built and evaluated by this conception, what remained undecided above does not thereby enter its target, and a claim that such a system does research has to say in which sense. What remained undecided also rests on different premises. Borgdorff's content enclosed in aesthetic experience and the experiential knowing that Reason and Bradbury set out after Heron presuppose a knower who has experiences; applied to an artificial system, which none of these theorists discusses, whether these have counterparts turns on what such a system encounters and how. Reason and Bradbury's reason for participation and Wilson's relational accountability concern who takes part in an inquiry and in which relationships; applied to a system, they ask whose perspectives and relationships enter its research, and, where S is human, how its human members take part. If institutions of evaluation used the conception, what it leaves out of the target could be overlooked within them; the consequence of institutional use described in the Introduction, Section 4 applies here too. Second, even within the conception, describing research as an attempt to bring about epistemic achievements does not by itself fix the targets of an artificial system. Those targets and their conditions still have to be specified. Unless they are, an evaluation that speaks of epistemic achievement can still leave understanding out, or count an enabling artifact as the intended achievement.

9. The debts the characterization leaves open

The book uses the working sentence mainly for the questions it makes explicit. The sentence raises many questions that it does not itself answer. As Section 5 noted, for example, a claim to knowledge brings the burden of explaining what distinguishes knowledge from a true result. This book calls such questions debts, since once the characterization is adopted, the later chapters are responsible for answering them.

The book groups the debts into four: the achievement, process, attribution, and research-order debts, each taken up by one of Parts II to V (Introduction, Section 6). They are kept apart because an answer to one does not settle another. Showing that an artificial system produced a correct result, for example, does not by itself pay the attribution debt, since whether the conduct of the research or a research capacity can be attributed to that system has still to be shown. The grouping thus lets one ask which debts a claim about artificial research pays, and how far.

The four debts organize the questions that the book's parts take up; they are not derived term by term from the working sentence. Nor do they correspond one to one with the terms of the sentence or with the four levels of Section 1. Each debt bears on more than one term, unlike the correspondence Section 2 drew between the terms and Sections 3–6. As for the levels, the achievement and process debts concern the levels of the same names, and the attribution debt, which asks where an activity, capacity, or status is located, cuts across them. The relations on which the research-order debt bears, described below, are assessed chiefly at the level of value and governance; questions that Section 1 placed at that level have no debt of their own, and Part V takes them up where they bear on these relations (Chapter 9, Section 1).

First is the achievement debt. What distinguishes the intended epistemic success from a novel output, warranted acceptance, useful artifact, or enabling resource? Chapters 2 and 3 examine knowledge and understanding: their conditions, possible bearers, and relations to explanation and ability. Production, acquisition, possession, availability, use, and preservation need not coincide. No general word settles these differences, and a further achievement claim incurs a further burden of explanation.

Second is the process debt. How does inquiry proceed when an answer is not available to the relevant society, a project may fail, and its question, concepts, evidence, methods, and standards may themselves change? Candidate formation, experiment, proof, modeling, interpretation, construction, criticism, replication, anomaly diagnosis, and representation are possible roles, not a fixed sequence.

Third is the attribution debt. To what, and at which stated boundary, can research be attributed? Producing one correct result does not establish a stable research capacity. Answering the question requires asking whether a contribution, the conduct of an activity, a capacity, or researcher status is being attributed to an artificial actor; which parts of inquiry it selects, executes, monitors, revises, or governs; which remain externally supplied; and in which respects, and how far, it is claimed to be autonomous. Attributing the conduct of research also leaves open where its intended achievement would be realized: the actor need not be its bearer, nor belong to the society relative to which it would be new.

Fourth is the research-order debt. By the research order this book means the social order surrounding research, made up of the division of labor, the allocation of inquiry, criticism and certification, memory and correction, credit, and authority. Research already distributes cognitive labor across collaborators, tools, archives, review practices, organizations, and publics, and divides responsibility among the people and institutions involved. Artificial researchers may change agenda setting, attention, allocation, validation, memory, correction, credit, and authority. This metascientific problem has an origin independent of the working characterization, but the two meet when an artificial actor enters a research world, because three of the characterization's terms, new to S, epistemic achievements, and S itself, mark where the activity is tied to S, and, taken together with conditions outside the characterization, give rise to relations such an actor may change (Chapter 9, Section 2.1). Because S's unknowns are countless while the means of inquiry are finite, and many researchers pursue them, the allocation of inquiry among S's unknowns arises. Because members of S cannot each gather the evidence for many results and so rely on results through others, the criticism and certification of results arise. And because S consists of many members without a single voice, and the consequences of inquiry can reach beyond S, authority arises: who may decide on, and who may revise, allocation and certification. The characterization makes none of these relations a condition of research status. On accounts on which social organization enters warrant or possession, however, certification bears on what counts as S's knowledge, and on accounts that make group knowledge depend on the members with decision-making authority in the relevant domain, so does a form of authority (Chapter 2, Section 6). Accounts of public or collective understanding can also make social organization bear on the achievement claimed (Chapter 3, Section 6). Where a relation does not enter an achievement's conditions, its assessment remains at the value and governance level.

These debts constrain one another. An achievement account can misdescribe a practice; a practice description does not establish warrant; an attribution boundary can hide distributed work. Chapters 2 and 3 assess knowledge and understanding claims where made, while Chapters 4–6 examine the formation and recasting of questions, how candidates and representations become available, and the localization of failure, Chapters 7 and 8 locate conduct and capacity and ask whose achievement would result, and Part V examines the allocation of inquiry, the criticism and certification of results, and authority over both. A different target conception could change their answers without erasing their questions. Readers who reject the opening orientation can still identify which conclusions depend on it and assess the remaining arguments.