Part III — How Does Unfinished Inquiry Move?

Chapter 1's working characterization identifies research as an attempt to bring about epistemic achievements new to a reference society S. The two chapters of Part II examined knowledge and understanding as possible targets and compared their conditions. But stating the conditions under which an achievement obtains is not describing the process that reaches it. What Part III asks is what is done, and what can change, in an inquiry whose answer has not yet been obtained and which may fail. Inquiry here refers broadly to the process that moves forward within an attempt at research. The accounts seen in Chapter 1, Section 3 of what it is for an individual to be inquiring use the same word in a narrower sense.

Chapter 2, Section 8 called question formation, evidence acquisition, credal change, acknowledged ignorance, and the like candidate epistemic transformations, since whether a particular change constitutes progress depends on its relation to a worthwhile target, a correct answer, error exposure, an improved capacity for inquiry, and so on. Ross argued that formulating a worthwhile question or acquiring relevant evidence can itself be progress, and Harris replied that such changes only promote progress (Chapter 2, Section 8.2). For what bears on later advances, such as datasets, methods, and classifications, constituting an advance, promoting one, and merely being rationally expected to promote one were also distinguished (Chapter 2, Section 8.3). A research process can alter questions, candidates, representations, evidential relations, methods, abilities, recognized ignorance, and more before it yields knowledge. How such changes arise within inquiry, and how they can be related to one another, can be asked separately from whether they constitute progress or only promote later progress. Part III takes up the former question.

Chapter 1, Section 7 drew out, as an implication that follows comparatively directly from the working characterization, that an attempt is directed at something not yet settled in S, and called this the minimal question structure, but it left undecided in what sense that something is a question. Moreover, where the answer is not yet known in S, it may be unclear not only which proposed answer is correct, but what could count as an answer, what the problem is, which objects to compare, which representation makes the problem tractable, what would count as relevant evidence, which operation could produce it, and what a result will alter. Rheinberger, who studied the history of experimental systems, makes a related point about how the unknown is approached (Chapter 4, Section 5). Still, there is also research in which, even where the answer is new, the procedure for reaching it is substantially established in advance (Chapter 5, Section 1).

How unclarity about what could count as an answer is handled in ongoing work can be seen, for example, in the following short record. In October 2015, the health department of the Brazilian state of Pernambuco was notified of 26 newborns with microcephaly by several public and private hospitals that serve as centers of maternal and child care. The department's technical note dated the 30th of that month also cited 22 records of microcephaly in the state's birth registry for January to September 2015 and, comparing these with a median of 9 a year for 2005–2014 (range 4 to 16), stated that the information, though partial, already indicated a change in the pattern of occurrence of this congenital anomaly. Some of the mothers had reported a rash during pregnancy, but the note stated that on current information the anomaly could not be related to any disease, and that its cause could be a congenital infection or some other, non-infectious factor. It then fixed a case definition, by head circumference according to gestational age, to be used for notification from that point on, and stated that all ninety patients reported from various sources would be assessed against it. It recommended that tests for infectious diseases be carried out only in cases whose tomography showed findings suggestive of infection, announced that investigations to formulate hypotheses about possibly associated factors were under way, and stated that the protocol would be updated as clinical and epidemiological evidence developed (Secretaria de Saúde do Estado de Pernambuco 2015, 1–3).

Within this document of a few pages, open points, provisional answers, and means of checking them appear nested within one another. To the point of whether the pattern of occurrence had changed, a provisional answer based on partial information was given by means of the registry records and a standard of comparison, the median of the preceding ten years. What was counted there, however, were cases entered in the birth registry under a code of the International Classification of Diseases. What was to count as a case in subsequent notifications was fixed separately, as the definition by head circumference. This definition made it possible to ask which of the ninety reported patients met it. Unlike a vague "aren't there rather many?", this is a question in which what could count as an answer is fixed. On the other hand, what the anomaly was due to remained open. The note only listed broad candidate answers, a congenital infection or a non-infectious factor, and left the formulation of hypotheses about the factors to the investigations. Even so, the candidates bore on how means were chosen. The recommendation to restrict tests for infectious diseases to cases with findings suggestive of infection fixes, in the light of infection as a candidate, which cases are to be examined for what. Further, the results of assessing the ninety could bear not only on the provisional answer that the pattern of occurrence had changed but also on whether the case definition itself was appropriate. The announcement that the protocol would be updated can be read as leaving room for the results of checking to revise the definition used in checking.

Part III uses a picture to describe and compare research processes. The picture describes an inquiry configuration at a given point through dimensions such as what is asked and what remains unresolved, what contents are taken as candidate answers and what attitudes are directed toward them, representations, evidence, methods and instruments, standards, practical capacities, actors, records, and social arrangements, together with the relations among them. Research can proceed as a change in one dimension alters which changes are possible in other dimensions, and the configuration so reconfigured is reconfigured further. The picture is not a model fixing an order of stages but a vocabulary for comparing processes (Chapter 5, Section 1.1). The cases it draws on were selected, to suit the points at issue, from records such as papers, historical studies, laboratory notebooks, training logs, repository histories, and oral histories (Chapter 5, Section 1.2), and what they show, the cases of Chapter 4 included, is that certain relations can occur, not how often they do.

For many of the systems reported today, the question whether a machine does research becomes a question of where the work that organizes a research process lies, in the machine or outside it. Labels for stages such as hypothesis generation and verification leave unexplained how a question becomes available, how a candidate is made thinkable, how a mismatch is diagnosed, and the like (Chapter 5, Section 1). The three chapters of Part III therefore describe that work, so that Chapter 7, Section 2 can ask which of it is realized within an artificial actor and which remains distributed across other actors and arrangements.

The health department's note describes one inquiry from three sides at once. One is a description of what is open: the cause of the anomaly, and which of the ninety meet the definition. Another is a description of what was fixed for the time being: the standard of comparison, the case definition, the broad candidate causes, and the range of cases to be tested. A third is a description of what results yet to be obtained can bear on: the results of assessing the ninety, and the announcement that the protocol would be updated. Moreover, the same element appears on more than one side. The case definition by head circumference is something fixed for the time being, and at the same time what made it possible to ask, of the ninety, a question in which what could count as an answer is fixed, and something that the results of the assessment could revise. These three sides correspond, not only in this note, to three relations that are written apart when an ongoing inquiry is described: 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. Below, these are called the three relations of inquiry. The elements of the picture's dimensions are not each assigned to one of the relations; the same element can stand in different relations. The three relations do not exhaust the relations within an inquiry, but the three chapters of Part III each look at the same inquiry from the side of one of them. Chapter 4 treats the relation to unsettled points as the question of what it means for a machine to ask; Chapter 5 asks how candidates, representations, and means become available; and Chapter 6 asks what results bear on and what they revise.

Chapter 4 — What Does It Mean for a Machine to Ask?

The 1950 paper of the maize geneticist McClintock opens by stating that, in the course of an experiment designed to reveal the genic composition of the short arm of chromosome 9, a phenomenon of rare occurrence (or recognition) in maize began to appear with remarkably high frequencies in the cultures: mutable loci. They first appeared in the progeny of the self-pollination of about 450 plants whose chromosome 9 short arm had undergone drastic structural modifications during the cycle in which chromosomes repeatedly break, fuse, and form bridges early in their development, and about 40 different mutable loci were recognized in them. Because the stocks from which they arose had been under investigation for some years without showing such a large number of unstable loci, McClintock concluded that some part of the mechanism of this cycle, or some of the structural modifications it produced, was responsible for the conditions that produced the burst (McClintock 1950, 344–346). The 1944 annual report likewise reports a study that used this cycle to produce a large number of newly broken chromosomes 9 and to obtain mutations in their short arm (McClintock 1944, 127–132).

Let us separate what changed. What organized the original experiment was what genes the short arm of chromosome 9 carries. What came to organize the work after the unexpected mutations appeared, and became the title of the 1950 paper, was the question how mutable loci arise and how they behave. The conclusion that the mechanism of the cycle or its products created the conditions is a candidate answer to this later question, not the question itself. An unexpected observation changed what was asked. A memorandum of January 1949 formulates, for c-m1, the first mutable locus isolated in the cultures having Ds, the conclusion that it arose when a locus called Ds left its original position, was inserted at a locus called C, and inhibited the action of C (McClintock 1949a, 4, 9; McClintock 1950, 349–350). This, too, is a candidate answer to the same question. Moreover, even after the question of how mutable loci arise came to organize the work, narrower questions arising under it remained open. The same memorandum leaves undecided whether large colorless sectors appearing on an ear were due to the loss of a factor called Ac or to the stabilization of the unstable mutation called c-m2 itself. An addendum later attached to a manuscript of April 1949 likewise leaves undecided, of a difference in Ac between a tiller and the main stalk, whether the tiller had gained Ac or the main stalk had lost it (McClintock 1949a, 49; McClintock 1949b, 33). The interpretation that Ac was lost and the interpretation that c-m2 became stable are candidate answers to a narrower question that arose under it, what the colorless sectors on the ear were due to; they are neither the question of how mutable loci arise itself nor answers to it. Two things can be read from this example. One is that what is asked can change in the course of the work, and that the change can be ascertained from the records. The other is that a shift in the question and the persistence, under the new question, of undecided candidate answers are different things.

Suppose a similar shift were seen in the record of a machine's work. If one then says that "the system recast its question in response to an unexpected result," what is being claimed? That the system output a new interrogative sentence, that it performed an act of asking something, that it directed an attitude such as curiosity toward the question, or that its subsequent work came to be organized toward a new unsettled point? When human researchers are said to have "recast the question," these are often spoken of without being distinguished. When A is a machine, they can hold separately or be missing separately. And what prompted the recasting is something to be ascertained separately, from the records, as the published account does in McClintock's case.

The last of these, that the subsequent work came to be organized toward an unsettled point, corresponds to the minimal question structure that Chapter 1, Section 7, drew from the working characterization. According to that section, the structure does not require that a project begin with an explicit interrogative, that one actor entertain it, or that the final question have existed unchanged from the outset. And what was derived there is where the attempt is directed, not S's actual state. From this structure it therefore follows that an attempt is directed at an unsettled point, but not in what sense that point is a question (including whether the output of an interrogative, an act of asking, or an attitude must be added to it), whether its content was determinate from the outset, who holds it, or for whom it is unknown. What this chapter asks, on that basis, is what it means for a machine to ask.

Sections 1 to 7 of the chapter are divided along the four points that the structure leaves undecided: in what sense it is a question, whether its content is determinate from the outset, who holds the question, and for whom it is unknown. In what sense it is a question is treated in Sections 1 to 3; that its content need not be determinate from the outset is treated as the formation and recasting of questions (Section 4) and as a question's ceasing to stand without being answered, or being settled (Section 5). Who asks is taken up in Section 6, and for whom something is unknown in Section 7, which separates what is unknown to A from what is unknown to S. Section 8 reads empirical studies' claims that machines generated questions or research ideas in light of these distinctions, and Section 9 looks at how descriptions of the same activity of a machine diverge between this book's way of understanding research and the views of Ravetz in his later account, Simon, Nersessian, Funtowicz and Ravetz, Laudan, Bromberger, Kelp, and Lee, which address problems, questions, and inquiry, and then looks back over the chapter as a whole.

1. Dividing the claim that "a machine asks"

When it is said that a machine asks, what is being claimed? A machine's outputs remain in records and can be examined, but what acts or attitudes they amount to is just what is disputed, and when S is a human society, whether they were taken up as questions pursued by members of S is a further question.

Whether machines can ask is also discussed in debates about machines doing research. It has been said, for example, that current training produces obedient students who answer known questions rather than the skill to ask the right questions (Wolf n.d.), and that asking the right question is the hardest part of science, so that for some time conjectures and hypotheses will come from human experts (Hassabis and Manyika 2026). A physicist who worked with a machine reports that it can supply candidate calculable problems in large numbers but that its judgment of whether something is interesting cannot yet be trusted, so that the human supplies that judgment (Schwartz 2026). Most such remarks concern the systems of a particular time and are not claims that machines cannot do this in principle; Hassabis adds that intuition could in the limit be modeled by an AI system. Separately, declarations by mathematicians hold that solving problems is only a tool and proxy for conceptual understanding (a declaration by Fields medalists, Avila et al. 2026), and that credit and responsibility continue to belong to humans (Alper et al. 2026). Whether a machine can supply candidate questions, whether it can judge which questions are worth pursuing, and who sets the aims of research and answers for them are different claims. The distinctions that follow also help to tell which sense of question such remarks concern and what they claim about it.

As noted in Chapter 1, Section 7, the word question is used for an interrogative sentence, for its content, and for the act of asking it. According to Cross and Roelofsen, who survey the philosophy and semantics of questions in the Stanford Encyclopedia of Philosophy, distinguishing these three is standard in the philosophy of language, and analyses of the act of asking divide into, among others, Searle's account, on which it is a kind of request, and that of Åqvist and Hintikka, on which it is an imperative addressed to the hearer to bring it about that the speaker knows the answer (Cross and Roelofsen 2026, §§1.1, 4.1). Ciardelli, who defends inquisitive semantics, which captures the meaning of an interrogative through the information that resolves it, distinguishes the content of an interrogative from the act of asking for the same reasons for which the content of a declarative is distinguished from assertion, and calls this content an issue, in the sense of a matter awaiting resolution. On Ciardelli's account, the same issue can be the content of different acts, such as asking, wondering aloud, and taking it off the agenda, as in "let's set aside who did it," and what is investigated when something is investigated is not a speech act but a content grasped in thought (Ciardelli 2022, 24–31). What Watson's account of the act of asking (Watson 2015, 86–102) and Habgood-Coote's (Habgood-Coote 2026, 1, 8, 11), seen in Chapter 1, Section 7, require of the asker, and what they do not, is taken up in Section 6.

Besides content and act, there are other things spoken of with the word question. One is a mental attitude directed at a question. Jane Friedman, writing on inquiry and interrogative attitudes, extends the standard idea that questions are the semantic contents of interrogative sentences to the contents of minds, and argues that interrogative attitudes such as wondering and curiosity are attitudes that have questions as their contents (Friedman 2013). A second is an unsettled point that organizes an inquiry, and the minimal question structure of Chapter 1, Section 7 is of this kind. It is a point at which an attempt is directed as something not yet settled for S; it need not in fact be open in S (the opening of this chapter), and it may not yet have become an interrogative, an act, or one mind's attitude. A third is an agenda, that is, the list of questions that a conversation or group has taken up as what it pursues. These senses are not exhaustive; they are listed to the extent the distinctions below require.

Once these are separated, the presuppositions that a claim that "a machine asked" needs differ according to which sense the claim concerns. If it is only that the machine output an interrogative sentence, that is a claim about form. Watson argues that being interrogative is neither necessary nor sufficient for being a question (Watson 2015, 95, 102). Watson does not discuss artificial systems, but applying the account, an interrogative output by a machine may fail to be an act of eliciting information, and conversely an operation containing no interrogative, such as sending a search term or querying a database, may be a question. A stage label such as question generation (Chapter 5, Section 1) therefore does not by itself determine which of the senses above it refers to. To move from output to act, applying Watson's account, requires an account of the activity within which the machine's operation functions to elicit information. Whether that function is fixed by the purposes of designers or users, or by the organization of an activity that includes the machine itself and by the goals and norms of the practice it joins, depends on a further account of function (Chapter 1, Section 3; Section 6). On either route, the point at issue moves from whether the machine is curious to within which activity, and from which standpoint, the function of eliciting information is attributed. If Ciardelli's distinction is brought in, the form of an output alone does not determine whether it is an act of asking, something closer to wondering aloud, or the arranging of an agenda, and attributing an act of asking requires, separately, an account of, for example, what commitment the output is treated as within a conversation with S. The same gap can arise for attitudes. On Jane Friedman's account, a higher-order representation of one's own ignorance, that is, representing that one does not know something, is neither necessary nor sufficient for an interrogative attitude (Friedman 2013). Applying this, a machine's outputting "X is not yet known" is not by itself evidence of an interrogative attitude, and conversely, the fact alone that a machine does not represent its own state of knowledge does not exclude it from what can bear such attitudes.

In the sense of an unsettled point that organizes an inquiry, matters differ again. As Chapter 1, Section 3 noted, the book's attempt can be grounded not only in an individual's intention but also in a group's plan, an artifact's organized function, a component's role, and so on. That a machine's activity is organized toward an unsettled point therefore depends neither on the machine's having performed an act of asking nor on its having an interrogative attitude. Conversely, even if a machine outputs interrogatives in great numbers, there is no question in this sense if they organize no work. Further, the one who produces a question and the one who receives and pursues it need not coincide. A human team may pursue an interrogative that a machine output, and a machine may pursue a question that a human gave it. For a question that a machine output to organize S's inquiry, it must be taken up by members of S and placed on the agenda. If S includes machine members, whether exchanges among machines can be regarded as proposing questions or as functioning to elicit information depends on the account of the act of asking (Section 6).

A machine's output is sometimes evaluated as having "formulated a good question" (Section 8), but different things can be said about the "goodness" of a question, too. The distinctions drawn by the authors below are not a list of necessary conditions for a question's being good; they serve to separate which point a given evaluation concerns. Collingwood, who took metaphysics to be the historical study of the absolute presuppositions that certain people's thought has made, separated a question's arising, that is, its presupposition's being actually made, from that presupposition's being true (Collingwood 1940, 26, 28). In Collingwood's sense, a question with a false presupposition can arise (Section 5). Ravetz, in a 2016 interview, connects research priorities with the ignorance that results when some kinds of investigation are not pursued (Ravetz 2016, second answer). Watson separates the value of questioning from good questioning, and argues that goodness has a thin sense, succeeding in any way at all, and a thick sense, competently eliciting valuable information, and that competence is both narrower and broader than success: a vague question that happens to hit the mark is not competent, and one can question competently even when no answer is obtained (Watson 2015, 135–140). What Watson evaluates is questioning as a performance and the questioner's skill that supports it; the choice of what to ask is also evaluated as part of that skill (Watson 2015, 144), but how the content of a question is semantically determined is not an object of the evaluation. How a question's answer conditions are semantically determined (Section 2), whether it is practically tractable (Section 3), and whether there is value in pursuing a candidate or reason to direct resources at it (Chapter 5, Section 4.2; Chapter 10, Section 6) are also evaluations distinct from these.

When A is a machine, these distinctions offer a handle for asking what an evaluation that "the system formulated a good question" has checked. A tractable question need not be one at which S has reason to direct resources. Even if the presupposition of a question a machine produced is actually made in S, that presupposition may be false. Applying Watson's distinction between competence and success, a machine's question's happening to lead to valuable information is not a sufficient reason for saying that the machine questioned competently. This is the form, on the side of questions, of the point seen in Chapter 1, Section 3 that one fortunate output gives little reason to call its source a researcher. Attributing competent questioning, like attributing capacities, involves conditions such as reliability across varied conditions and responsiveness to changing evidence (Chapter 7, Section 3.1). And who may determine research priorities is distinct from whether a machine can estimate the considerations relevant to them; Chapter 10, Section 6 and Chapter 12 take this up. The remarks at the opening of this section can be read in the same way, and they concern different things. The physicist's report separates supplying candidate problems from judging which of them are interesting; the remarks about the skill to ask the right questions concern the choice of what to ask, which Watson evaluates as part of the skill of competent questioning; and the declaration that credit and responsibility continue to belong to humans concerns who answers for the aims, which is settled apart from either.

The distinctions so far have concerned what a question is and how it is evaluated. At a different level are claims about the working characterization itself. As Chapter 1, Section 7 noted, the minimal question structure does not fix where a question comes from. A question may occur in the middle of an experiment or of reading, in a conversation with colleagues or at a meal. One may also ask by going back to the presuppositions one is relying on: while working through a procedure, one may ask why one is doing it this way in the first place, or, after a series of improvements, whether the approach serves its aim at all and what one was trying to do in the first place. Such questions turn what had been working as a presupposition into an object of questioning; this overlaps with Collingwood's relative presuppositions, taken up in Section 5, which work as presuppositions while in use and pass to the side of answers to questions on inspection. One may also ask anew about what one had taken for granted, as when, looking at the sky, one asks why clouds are white in the first place. A question may also be inherited from supervisors, collaborators, funders, communities, instruments, a discipline's history, and so on. That a researcher did not originate a question does not by itself, therefore, deny that research is being conducted in the sense of this book's working characterization. An individual, team, or artificial system can conduct research on a question supplied by another actor, or on part of a question held by a community. Bromberger, who wrote on the logic of questions and of ignorance, also holds that on a view that evaluates scientific contributions as the resolution of items of ignorance, a contribution need not be a result sought, in response to that ignorance, by someone aware of it (Bromberger 1992, 116–117).

Not fixing where questions come from does not mean that where they come from does not matter. If one holds standpoint theory, on which the social position from which research starts bears on the knowledge it yields, which community's records a question is inherited from bears on how one-sided the knowledge that research yields becomes. Chapter 10, Section 6, drawing on standpoint theory and on Varsavsky's argument that research in one society can take its criteria of importance from elsewhere, examines how an artificial configuration that inherits the questions its records already pose, that is, an artificial unit to which research is attributed, can start from where those who produced those records stood. Also, the capacities to notice an unresolvedness, formulate or select a question, and revise its framing can expand what can be attributed to an actor and increase its autonomy (Chapter 7, Section 4.3), but they are not included in the working characterization as necessary conditions for every researcher or participant in research. How questions are actually formed and recast within inquiry is a matter at a level distinct from this claim that the origin of questions is not fixed, and Section 4 takes it up as a description of the processes that research records and accounts of inquiry depict.

2. What is asked, and what counts as an answer

What Section 1 called the content of a question cannot be separated from what counts as an answer. This section deals with questions in this sense, that is, with cases in which an unsettled point has taken the form of a question with conditions on what counts as an answer. The questions machines output, and the questions used in evaluations, usually have this form. The philosophy of language since Frege has emphasized propositions and declarative sentences, and much work on the semantics of questions has stayed within that framework, identifying a question with a set or function involving the propositions that are its answers; there are also attempts at a logic in which neither questions nor propositions are more basic (Cross and Roelofsen 2026, §§1, 2.1.1). What is taken as an answer also changes what a sentence embedding a question says and what follows from it: whether knowing what John reads amounts to knowing what John does not read holds on Groenendijk and Stokhof's account, on which an embedded question denotes its true exhaustive answer, but not on Karttunen's, on which it denotes the set of true answers (Cross and Roelofsen 2026, §4.2). For artificial systems, on the other hand, claims that "the system answered the question" are routinely counted in benchmarks and evaluations. What is counted as an answer then?

Cross and Roelofsen hold that there is no univocal pre-theoretical notion of an answer; it is a technical notion that can differ from theory to theory, and some theories leave it implicit. Notions of answer range from the view, shared by many theorists, that counts as direct answers, which may be false, only what answers a question exactly and completely, to accounts that count any information about the subject matter as an answer, and theorists also differ on whether answers must be propositions. A response that denies a presupposition is called a corrective answer, and many authors do not count it as a direct answer. The same interrogative can also have a reading on which giving one answer suffices and a reading on which all must be given (Cross and Roelofsen 2026, §§1.1, 4.3, 4.4). The same word can also refer to different notions in different authors. Wiśniewski, who developed inferential erotetic logic, assigns to each question as its direct answers the sentences that answer it exactly in their expression (Wiśniewski 2022, 108 n. 4, 109), and treats as distinct notions a partial answer, which says that the true direct answer lies within a certain proper subset of the direct answers, and an eliminative answer, which rules out at least one of the direct answers; the two come apart when the direct answers need not be mutually exclusive and none of them need be true (Wiśniewski 2013, 43–45). Habgood-Coote, by contrast, drawing on the Stalnaker–Roberts model that regards conversation as cooperative inquiry, follows Roberts in treating a question as a partition of the portion of logical space that satisfies its presupposition into mutually exclusive possible answers, and calls an answer that rules out at least one possible answer a partial answer; completely answering a subquestion gives a partial answer to the original question (Habgood-Coote 2022, 1107–1111 and n. 28). In Wiśniewski's terms, this is roughly an eliminative answer. Ciardelli calls the relation by which, given background information, a collection of subquestions logically determines the original question a dependency, and takes it to be a fact about meaning, not about discourse; where it holds, any answers to the subquestions, together with the background, answer the original question (Ciardelli 2022, 23–24, 29, 41–42).

Brought to the judging of benchmarks, these distinctions make the truth of the claim "the output answered the question" vary with the notion of answer adopted, and counting something as "answered" implicitly adopts one theory of answers. Even if a corrective answer that rejects a presupposition is not counted as a direct answer, whether it can end an inquiry, or make a question cease to stand, depends on each author's account (Section 5). Here the judgment that counts whether something "answered" and what was settled in research part ways. According to Wiśniewski, moreover, a formal question only represents a natural-language question under one interpretation (Wiśniewski 2013, 15). The same output may therefore be a direct answer under one formal representation and yet not be an answer to the question S is asking. Unless which reading is being asked is fixed on S's side, judging whether an output amounts to an answer cannot begin. Even if S includes machine members, using the same formal language does not by itself align the members' interpretations. Further, Habgood-Coote defines as a partial answer in a weaker sense (a casual partial answer) what, together with the available information, rules out at least one possible answer (Habgood-Coote 2022, 1109–1110 and n. 34). Applying this definition, the same output may be a partial answer for a system with access to a certain database and yet not be one for members of S without access to that information. Whether something is an answer becomes relative to what is available to whom.

Bringing Ciardelli's dependence relations to a system that divides a task into subtasks and collects the answers, whether the collected answers answer the original question can be asked at the stage of division, independently of who carried out the subtasks and how. If the division does not realize a dependence relation, there is no guarantee that the original question is determined even if every subtask is answered successfully. Further, a component in charge of a subtask's rejecting a candidate is not the same as narrowing down the answers to the original question in Wiśniewski's sense. Elimination becomes narrowing only when one adds the presupposition that the true answer lies among the listed options, and that presupposition is just what can be in question when the set of candidates may be incomplete (Section 3). Dependence relations can be asked about at the stage of division, however, only so long as the space of possibilities is fixed. In the course of research that space itself may be revised (Section 4), and a question may cease to stand in connection with its presuppositions (Section 5).

3. Understanding a question and making a question tractable

The minimal question structure of Chapter 1, Section 7 says that an attempt is directed at something not yet settled in S, but it requires neither that this point be expressed by an interrogative nor that what counts as an answer to it already be fixed. The accounts seen in Section 2 mostly dealt with questions for which what counts as an answer is already fixed. This section takes up cases that the structure allows but Section 2 set aside: an unsettled point whose answer conditions are not yet fixed, and the work, on the side of the question, of recasting it into a question whose answer conditions are fixed. Moreover, that a point is unsettled does not mean that the set of possible answers is already at hand. What, then, is it to understand a question when its possible answers are not at hand? For machines this question and the work of recasting carry weight for two reasons. One is that outputting interrogatives can be cheap for a machine; the other is that a machine can generate many candidate answers. What is attributed to a machine therefore divides according to whether being able to produce candidates amounts to understanding a question, and whether there is work of fixing answer conditions distinct from generating candidates.

Authors divide on how understanding a question relates to its possible answers. In the logic of questions there is an idea, deriving from Hamblin, that to know what counts as an answer is to know the question, and Wiśniewski retains it by dividing it into two steps: to know a natural-language question is to know a formula that represents it, and to know a formula is to know its possible answers. Wiśniewski anticipates the objection that for certain why-questions and how-questions it makes no sense to speak of a set of answers determined in advance and, without disputing that there are such natural-language questions, responds by distinguishing the possible answers to a natural-language question from the possible answers to a formula that represents it (Wiśniewski 2013, 16). Ciardelli states, as a slogan, that to understand a question is to know what information is needed to resolve it, and identifies the content of a question with the set of information states that resolve it in the model at hand (Ciardelli 2022, 36). Bromberger, by contrast, held that we understand many questions whose possible answers we are in no position to state, and rejected the view of many logicians of questions that identifies a question with an exhaustive set of possible answers. According to Bromberger, understanding a question cannot be knowing the truth conditions of each of its possible answers, since one must also know that each answer stands to the question in the relation of being an answer to it. And to know that one is in a situation in which one can think only of answers known to be false yet has reason to think the question sound, one must be able to judge that the set of candidates is incomplete, even if one cannot extend it (Bromberger 1992, 4, 130–131, 133).

Some authors stand between these positions. Nickles follows the "spirit" of Hamblin's insight while defining a problem not by its solutions themselves but by the constraints that fix the range of solutions (Nickles 1981, 109–111). The constraints include theoretical background, methodological conditions, and implicit conditions, and they are discovered as inquiry proceeds (Nickles 1981, 111–112). The more constraints one knows, and the more sharply, the better the problem is understood, and the constraints not only make it possible to recognize a solution on seeing it but also indicate where to look (Nickles 1981, 88–89, 113). Nickles is close to Ciardelli in identifying a question by its answer conditions, but differs in taking those conditions to include implicit ones that are discovered as inquiry proceeds.

These positions do not fit a simple opposition between having a question and having its possible answers. One way of reconciling Ciardelli and Bromberger is to read Ciardelli's "know" not as a capacity an inquirer possesses but as the specification of a question's semantic content. On this reading, a semantic model can specify resolution conditions without requiring an inquirer to enumerate the answers or possess a method for finding one. If so, failure to construct a candidate does not by itself show that a question lacks semantic content, and what Bromberger's cases make visible is a practical gap: understanding what is being asked may leave the concepts and procedures needed to answer it unavailable. Bromberger's own point, however, is a theoretical claim about understanding a question, so this reconciliation holds only on condition that Ciardelli's "know" is read in that way. Representing a question's content and explaining how an inquirer makes progress on it are different tasks, and Ciardelli's slogan can be read as placing the understanding of a question on the side of the former, while Bromberger places it on the side of the latter, as an inquirer's grasp that an answer is an answer to the question.

Bromberger discussed a further gap concerning why-questions. On Bromberger's analysis, for what and where questions, understanding the question suffices to tell what must hold for there to be an answer, whereas the form of a why-question shows only that what it asks about must be the case. For a why-question to have a true answer, the fact must be construable under suitable predicates, attributes, and the like, but the form of the question gives no guide to these, and such concepts may be conceived long after the why-questions that call for them have been raised. One can thus be placed in a predicament in which one knows that one does not know the answer, yet does not know what one would have to find out to learn whether there is anything to know at all (Bromberger 1992, 157–168). This predicament is explicitly distinguished from the p-predicament discussed in the next paragraph, and is not peculiar to why-questions: "what is" questions can produce the same predicament in some circumstances (Bromberger 1992, 168–169 n. 46). As a rational way of dealing with this predicament, Bromberger mentions forgetting about the why-question for a while and turning to other questions with reliable credentials, and conjectures that answers to why-questions usually emerge from the study of such questions. Bromberger adds that this conjecture does not belittle answers to why-questions, and that the value of why-questions is beyond doubt (Bromberger 1992, 169). Read through the distinctions of this section, what is missing in this predicament is less candidate answers than clues to what to find out, and the response of turning to other questions whose answer conditions are known is work on the side of the question, distinct from producing candidate answers.

Distinct from this predicament is the p-predicament discussed in Chapter 2, Section 8.2. It is what Wilholt identified with deep ignorance: a subject is in a p-predicament with regard to a question if, on the subject's views, the question admits of a right answer, but the subject can think of no answer to which, on those same views, there are no decisive objections (Wilholt 2020, 9–10). What is missing in a p-predicament is a candidate answer that escapes decisive objection; one may be unable to think of any candidate at all, or every candidate one thinks of may have been rejected (Bromberger 1992, 4, 29). If one hardly understands a question, one can produce candidates without end and so is not in a p-predicament (Bromberger 1992, 31). Jane Friedman, too, argues that one can grasp a question and have an interrogative attitude even if one cannot represent a single possible answer (Friedman 2013). Grasping a question and having candidates can hold separately. How candidates compatible with a question's constraints become available is treated in Chapter 5.

One typical form of the work of making a question tractable is specifying a contrast for a why-question. Van Fraassen analyzed a why-question as a triple of topic, contrast class, and relevance relation. As Cross and Roelofsen present the analysis, "Why P?" has not been identified as a particular question until one has fixed a set of propositions to be compared alongside P (the contrast class), and the relation fixing what is explanatorily relevant (the relevance relation) is contextually determined (Cross and Roelofsen 2026, §3.2). In Woodward and Ross's presentation, context enters into specifying both, and the relevance relation depends on the context and on the information the questioner is interested in obtaining (Woodward and Ross 2025, §6.2). On Lipton's account, to explain why something happened rather than a specified alternative, the foil, requires citing a causal difference between the two. One need not cite a cause sufficient for what happened, but in return the answer must now say something about the foil as well (Lipton 1993, 43–44, 52). Lipton takes up the case of syphilis and paresis. Since most people with syphilis do not go on to contract paresis, philosophers have disputed whether a person's syphilis explains that person's paresis, and on Lipton's view contrastive questions settle the dispute: Smith's syphilis explains why Smith rather than Jones contracted paresis if Jones did not have syphilis, but not why Smith rather than Doe did if Doe had it too (Lipton 1998, 13). Lipton's own view, however, is that the model on which a contrast is explained by citing a causal difference is too simple as it stands: a fuller account must say, among other things, what explanation requires when there are several causal differences between fact and foil (Lipton 1993, 44–45), and which conditions a cause must satisfy to answer a particular question is what Lipton calls the problem of causal selection (Lipton 1998, 11). Moreover, as Cross and Roelofsen summarize Lipton's view, the central requirement for a sensible contrastive question is that fact and foil have a similar causal history against which the differences stand out (Cross and Roelofsen 2026, §3.2.3). Specifying a foil in this way narrows what an answer must do and where to look for it.

How far specifying a contrast amounts to recasting on the side of the question is, however, subject to qualifications. Lipton holds that one point of making a why-question contrastive is to reduce the indeterminacy of the question of how things would have been different had the effect not occurred, and that a foil lets us choose which parts of the effect's history to hold fixed in a way that meets our explanatory interests (Lipton 1993, 49, 51). He writes that many why-questions are contrastive (Lipton 1993, 39), the contrast often remaining implicit (Lipton 2001, 16). If an implicit foil is already fixed, making it explicit can be a restatement of the same question. It becomes a different question when the foil was not fixed, or when it is replaced by another foil. Further, a contrast does not by itself settle which cause to cite. In real cases, Lipton notes, there will seldom if ever be only one causal difference between fact and foil (Lipton 1993, 45), and not all of an event's causes explain it in a given context, which calls for a finer-grained account of the context in which explanations are requested (Lipton 2001, 10–11). Where several causes meet the difference condition for the same contrast, which of them to cite can turn on what the inquirer is interested in. Part of the fact that explanations vary with what one is interested in therefore remains on the side of the answer even after the foil is fixed. Bromberger, on the other hand, using van Fraassen's examples, argued that it is a mistake to think why-questions especially context-dependent on the ground that stress changes their answer conditions, and criticized theories that take stressed questions as their paradigm for generalizing from a biased sample (Bromberger 1992, 160 n. 34). On Bromberger's analysis, only one of the readings of a stressed sentence changes the presupposition; on the other readings, stress only indicates the speaker's interests (Bromberger 1992, 164–165). Since the presupposition-changing reading is a question of a different logical form, even on Bromberger's analysis fixing a foil can amount to replacing the original why-question with a different question. Lipton is not responding to Bromberger in that passage on contrastive explanation, but putting the two analyses together, the recasting of a question whose answer conditions are not fixed into one whose answer conditions are can be described as work on the side of the question that is not reducible to generating candidates. In Nickles's words, it is adding constraints on solutions and sharpening them.

Different kinds of question are made tractable in different ways. A how question, for example, may be decomposed into capacities or constraints, and a mathematical problem may change when rendered geometrically, algebraically, probabilistically, or computationally. An engineering ambition, too, may become an investigable problem only when it is translated into relations in which performance is measurable and into a physical arrangement capable of realizing them. Nickles, too, holds that a problem can be reformulated without changing the presentation of the data and transformed into another problem, and cites the case of the problem of blackbody radiation (Nickles 1981, 88, 114 n. 6). Cases in which candidates, or the objects of inquiry themselves, become available through restructuring data or restructuring the representation of a proof are taken up in Chapter 5, Section 3.

The range of possible answers itself may be unknown. Habgood-Coote, citing Bromberger and Wilholt, uses the label deep ignorance, which Wilholt identified with the p-predicament, for the state of not knowing what a question's possible answers are. In this use, the state so named is not the same as the p-predicament. For a question Q such as what the correct fundamental theory is, Habgood-Coote raises "What are the possible answers to Q?", an option question, a question that asks what the possible answers are (Habgood-Coote 2022, 1116–1117). Habgood-Coote treats only complete deep ignorance, setting aside intermediate cases such as ignorance of some of the options (Habgood-Coote 2022, 1117 n. 41).

Applying van Fraassen's analysis to machine outputs, even if outputting why-interrogatives can be cheap, that alone fixes neither the contrast class nor the relevance relation, which depend on the context and on what the questioner is interested in. Sawyer reports that van Fraassen gives some reason to think that an omniscient being would not be correctly understood as having explanations at all (Sawyer 2014, 7 n. 8). If that is so, holding much information does not by itself give a machine explanations; what the analysis asks for, a contrast class and a relevance relation, is not supplied by the quantity of what a system holds. If S is a human society and the interests on which the relevance relation depends lie on S's side, then even when a machine pursues the question, the question is given its shape by S's interests, and the question of whose question it is divides into two: who produced the interrogative, and who fixed the foil and the relevance relation. Applying Lipton's account, specifying a meaningful foil requires knowledge of the histories of fact and foil and a choice of whose interests to answer to. In a mixed S, foils can differ from member to member, so part of the disagreement between the explanations of machine members and human members might be described not as opposing answers but as a difference of foils. Also, making a question tractable is fixing its answer conditions, not making it easy. Moreover, since interests remain in which cause to cite even after the foil is fixed, whose interests they are still enters into how an explanation is chosen, even if the machine has received the foil correctly. For an S consisting only of machines, the same analysis leaves a position to be filled. Since, on Woodward and Ross's presentation seen above, the relevance relation depends on the information the questioner is interested in obtaining, applied to an S of this kind the analysis leaves open what stands in the questioner's position: what the members' answers are passed on to, the tasks set for the members, or something the members fix among themselves, among other candidates. Which of these, if any, fixes the relevance relation for such an S is not settled by the analysis.

Turning toward machines Jane Friedman's point that grasping a question and having candidates can hold separately, a machine can separately be in a state of grasping a question while unable to represent candidate answers, and in a state of being able to generate many candidates while not keeping the question open. Having many candidates does not by itself mean progress, since what Bromberger's definition asks about is not the number of candidates but whether they escape decisive objection on the subject's views (Bromberger 1992, 29, 31). Bromberger also has an extended definition, however, on which one is in a p-predicament if there are compatible answers one cannot yet think of, even when one has already thought of several compatible answers; on it, new candidates compatible with the constraints count toward resolving the predicament, even when there are many of them (Bromberger 1992, 122). How a machine that can produce only candidates known to be false, yet has reason to think the question sound, is to judge that its set of candidates is incomplete (Bromberger 1992, 133) can also be raised as a question from here. Further, applying Habgood-Coote's distinction, a system that works only within a supplied set of possible answers does not pursue the question of what the possible answers are. A system that adds some options, however, falls among the intermediate cases that Habgood-Coote set aside, so his treatment does not settle whether it pursues that question. Simon distinguishes problems supplied to programs from the work of formulating them, without dismissing the work of solving the formulated problems (Simon 1973, 186–187). Nickles notes that substantive theoretical content can be built into methods and apparatus (Nickles 2019, 363). Applied to the systems considered here, these points make attribution depend on the recorded task boundary. Where humans supplied the framing and the machine used it without revision, successful execution does not establish that the machine performed that framing work. How much creative work it performed in reaching its result remains a separate question.

Taken together, the distinctions of this section mean that even when a machine can be said to have put forward a question, who did the work of making that question researchable, that is, of deciding what to take as the contrast, what kind of explanation to count as an answer, and within what range of candidates to look for it, remains a separate question. In many of the empirical studies read in Section 8, the higher-level question is given by humans together with the task.

4. How are questions formed and recast within inquiry?

When it is said of a machine that "the system turned an anomaly into a question" or that it "recast its question," the claim is checked by asking which choices were made by whom, and what changed under which framework. To that end, this section moves to the descriptive side of the two levels separated at the end of Section 1, and looks at the formation and recasting of questions that stand out in research records and accounts of inquiry, attending to what changes and how. Specifying a foil, seen in Section 3, is also one form of this.

A situation may first be indeterminate. Concepts, contrasts, instruments, and candidate answers can help make it askable. Questions are refined, constituted, or reorganized across an inquiry (Hildebrand 2024, §4.3; Henne 2023, §6), and are sometimes divided into subquestions and carried among participants (Section 6). In Hildebrand's report of Dewey's theory of inquiry, a problem is formulated within an initially indeterminate situation rather than given before inquiry; reasoning about a hypothesis can lead back to redescribing the problem (Hildebrand 2024, §4.3). In How We Think, Dewey holds that observations are needed to make the specific character of a problem clear: if a proposed diagnosis is accepted before the trouble has been explored, thinking is cut short (Dewey 1910, 74). Dewey's problem does not fully coincide with this chapter's question, and Dewey's ideas and hypotheses lie on the side of possible solutions, that is, of candidate answers. With this in view, Dewey held that determining which data matter and elaborating the meanings suggested by them depend on each other: if selected facts do not sustain an interpretation, the data are reconstituted, while possible interpretations are elaborated in relation to them (Dewey 1910, 103–104). In Popper's method of conjectures and refutations as Nickles reads it, too, putting forward conjectured solutions and criticizing them makes the problem more fully and precisely determined (Nickles 1981, 92). In both descriptions, dealing with candidate answers changes the shape of the question, but question and candidates are not treated as the same thing.

Some accounts place the question still later. El Shazly, writing on inquiry aimed at understanding, argues that such inquiry begins not from ignorance but from puzzlement, confusion, or discursive disagreement; there, deciding what to ask is itself part of the inquiry, and since engagement with the object fixes the particular question, the question comes later in the order of explanation (El Shazly 2025, 15–18). On this account, the content of a question fixed later cannot be placed back at the beginning of the inquiry. Even so, an engagement whose content is not yet fixed as a particular question can be organized as an attempt aimed at an understanding, or another achievement, new to S. What the minimal question structure captures is that direction.

Henne, who works within Dewey's theory of inquiry, reformulates in terms of that theory a distinction inspired by Carnap's between internal and external questions. On Carnap's distinction, internal questions are asked within a linguistic framework and answered within that framework, as when one asks, in the language of things, "Is there a white piece of paper on the desk?"; external questions concern the framework itself, such as whether to introduce rules for speaking of a new kind of entity, or to adopt one framework rather than another (Henne 2023, §2). Henne's distinction is a distinction not between questions but between inquiries. Framed inquiries use a conceptual framework without questioning it in asking and answering questions; in Henne's example, biochemists studying the effects of temperature on protein synthesis are engaged in one. In framing inquiries, the question or problem cannot be fully formulated within existing frameworks, and the framework itself is questioned, revised, expanded, or created; the inquiry in which William Thomson built a theoretical framework on the notion of absolute temperature is an example (Henne 2023, §§1, 6). As Henne's own account emphasizes, the core of the distinction lies not in where a question is located relative to a framework but in whether the framework determines the questions to be asked and how they are to be answered, and even an inquiry about a framework can be a framed inquiry (Henne 2023, §6.2).

Henne's distinction, moreover, does not draw a sharp line in the way Carnap's does. Henne classifies an inquiry by its situation and conduct rather than its outcome, so that an inquiry into whether a definition should change is framing even if it ends by retaining the definition. Henne also treats the distinction as a matter of degree that depends on what is counted as one inquiry. Even the most radical framing inquiry relies on some framework, and even in the most typical framed inquiry the framework does not fully determine how the situation is handled. One inquiry can begin as framed and end as framing, and the reverse can also happen (Henne 2023, §§1–2, 6).

In light of this distinction, one can ask how far inferential erotetic logic captures the formation of questions. Wiśniewski formally defines, as two relations, the ways in which a question is derived from a set of declarative sentences or from another question. Roughly, one is the relation by which a set of declarative sentences gives rise to a question that, if the sentences are true, must have an answer, but whose answer is not yet determined; this is called evocation. The other is the relation by which a principal question gives rise to an auxiliary question that helps answer it; this is called erotetic implication (Wiśniewski 2022, 107–108, 112, 116–117). On that basis, a tree of questions and answers that places the principal question at its root, adds auxiliary questions by erotetic implication, and branches from each auxiliary question that is a query by each of its direct answers is called an erotetic search scenario (Wiśniewski 2022, 109, 118, 122–124). Wiśniewski reports, as the work of Chlebowski and others, that the generation of scenarios can be automated by operations defined on such scenarios (Wiśniewski 2022, 125). Scenarios can be reworked in the course of execution (Wiśniewski 2022, 125 n. 15; Wiśniewski 2013, ch. 13), but the principal goal fixed by the principal question does not change (Wiśniewski 2022, 122). What these definitions presuppose is that a question and the set of its direct answers are fixed in a language and, for erotetic implication and search scenarios, further that the principal question stays fixed. Wiśniewski himself holds that, since against the background of a logic that accepts the principle that anything follows from a contradiction (explosion) inconsistent premises evoke no question, a different strategy is needed where questions arise from inconsistency, such as taking as background a logic in which a contradiction does not entail everything (a paraconsistent logic), which the general framework of inferential erotetic logic does not prevent (Wiśniewski 2013, 62–63; Wiśniewski 2022, 113–114); Wiśniewski also states that inferential erotetic logic does not yet provide an account of erotetic inferences that are not valid but are plausible (Wiśniewski 2013, 66). Seen this way, what these definitions capture is how questions arise when the set of answers is fixed in a language and, for erotetic implication, the principal question also stays fixed: one case belonging to framed inquiry in Henne's sense. Even within framed inquiry, however, a principal question can be recast, for instance by specifying a foil (Section 3). Nor does a change in the space of answers by itself amount to framing inquiry, since Henne's criterion is whether the framework determines the questions to be asked and how they are to be answered, not the range of answers itself.

What formalization does not capture are, for example, cases in which inquiry begins from an anomaly, an archive, a dataset, an instrument, an assigned task, a failed intervention, an unfamiliar material, a practical need, or a diffuse sense that existing descriptions do not fit, and what is to be asked itself becomes fixed in the course of inquiry. Turning such a diffuse beginning into a question for which answers can be sought may require deciding what to count, for which group, in what units, compared with what, over what period, and by what standard: that is, choosing a scale, population, unit of analysis, contrast, temporal boundary, evidential standard, and so on. In some cases, as in mathematics, it may require choosing a formal language in which to state the question. Such choices may divide what appeared to be one question into several, bring several questions together into one, or reveal that the original object was partly an artifact of a measurement or classification. Making a question determinate is not deciding everything at once. In the health department's note seen in the introduction to Part III, the standard of comparison, the period, the case definition, and the patients to be assessed were fixed for the time being, while the cause was left open. If the cases entered in the birth registry under an International Classification of Diseases code are treated as observational data for specifying the problem in Dewey's sense, then the case definition and the standard of comparison were fixed for the time being separately from them. Moreover, these choices are not changes of the same kind. Re-choosing a standard of comparison or a foil is in many cases refining a question within existing concepts and, applying Henne's distinction, can be taken to remain within framed inquiry. Revising the definition of what counts as a case, by contrast, can amount, as a small revision of a concept's conditions of application, to an inquiry that is framing to a small degree. Indeed, Henne cites small revisions of conditions of application, such as raising the age of consent, as examples of a small degree (Henne 2023, §6).

Such choices can succeed or fail, and when they succeed a question that was not there before becomes available. Articulating an explicit and tractable question, dividing what looked like one question into several, or fixing a question for the time being so that it can be pursued can therefore be work accomplished within inquiry rather than merely an input given to it from outside. Whether such work amounts to progress or only promotes later progress is where the views of Ross and Harris divide (Chapter 2, Section 8.2). The claim that questions are formed within inquiry is nevertheless easy to overextend. After the fact, any event can be described as answering some question, so attributing to an inquiry, after the fact, a question that organized it needs clues. In light of what Chapter 1, Section 3 asked of an action, what made it relevant to an unresolved matter, what possible outcomes would have changed its role, and where the capacity to redirect the activity lay, retrospective question attribution is too permissive if it is granted on verbal redescription alone, and is better supported when the attributed unresolvedness explains the organization of the activity: why some operations were selected, why possible outcomes could have mattered differently, why a line was continued or abandoned, and how later developments remained connected to what was unresolved for the relevant subject or society.

The question a result is later said to answer need not be the question that actually organized the inquiry leading to it. Arfini and Costa's account of serendipity, finding by chance something other than what was sought (compare Chapter 1, Section 3), gives an example. They distinguish the explanations that discoverers construct while the outcome is uncertain from the retrospective explanation available only once the research is concluded, and argue that in serendipitous discoveries that change how a phenomenon is understood, what is to be explained changes during the process. For inventions arising by chance, they hold, the two explanations coincide. Penzias and Wilson, known for the discovery of the cosmic microwave background radiation, began, on their reading, by seeking to explain why their horn reflector was not working as expected, spent some time excluding possible causes of the noise it picked up, and ended by explaining why any such reflector would pick up the signal (Arfini and Costa 2025, §§2.1, 3–3.2, n. 13). So as not to confuse a shift in the question with the appearance of candidate answers, questions and candidates are indexed from here on. Writing q₁ for the question that organized the work first and q₂ for the later question, in McClintock's case at the opening of this chapter the question about the genic composition of the short arm of chromosome 9 is q₁ and the question how mutable loci arise and behave is q₂. Written in the same way, why the horn reflector was not working as expected (what was producing this noise) is q₁, the possible causes of the noise are candidate answers c to q₁, and why any such reflector would pick up the signal is q₂. If what is to be explained changes in this way, the question q₂ that the reported result answers is not the question that organized the work leading to it, and, by the conditions of the preceding paragraph, what explains the work of excluding possible causes of the noise is not q₂ but q₁. An account that presents all the work as directed at q₂ can be right about the result without describing what organized the earlier work (compare Chapter 5, Section 2).

In the Penzias and Wilson case the change of question is visible through a reconstruction made after the research was over. From records written at the time, by contrast, such as McClintock's report and memorandum, one can sometimes read what organized the work at that point and what was open. A question can also be reorganized while its headline stays fixed. On the reading of the editors of Darwin's notebooks, the basic questions of the transmutation notebooks—how species adapt to changing conditions, how new species arise, how the hierarchy of classification comes about—were set at the start of Notebook B, the first of them, and kept, while what those questions were taken to include and how the relevant phenomena were related changed as the answers did (Barrett et al. 1987, 168, 329, 395–396). For the formation of candidate answers, too, recollection and contemporaneous record diverge. Darwin's autobiography recalls that on reading Malthus "for amusement" in October 1838 it "at once struck" him that favorable variations would tend to be preserved. The notebook entry that records the thought, however, is dated 28 September, while Darwin was still reading the book. It writes the thought down as a force like "a hundred thousand wedges" trying to force every kind of adapted structure into the gaps in the economy of nature, and the editors take it as the first formulation of natural selection. The entry is preceded by a note on the infertility of hybrids and by extracts on fossils from Lyell's book, and followed by a note of what Darwin had been told about the curiosity of monkeys. According to the editors, Darwin finished reading Malthus on 2–3 October, and the theory went on taking form in the next notebook, Notebook E (Darwin 1958, 120; Barrett et al. 1987, 374–376). What the recollection presents as a single sudden event appears in the contemporaneous record as a step written down in the midst of work on other questions, and one that went on changing afterward. The record also fits well with Dewey's remark that time is required to digest impressions into substantial ideas, and that an idea matures through a silent working-over of alternative suggestions (Dewey 1910, 37, 209–210). Taking the editors' reading that the content of the basic questions changed with the answers together with this record, that change, too, can be thought to have proceeded among such jottings of candidates. A report that a machine writes after its result can diverge from the record of its work in the same way (compare Chapter 5, Section 2.1).

A supplied task, too, may by itself be less than a question. In one biomedical-engineering laboratory studied by Nersessian, who investigated cognition and practice in scientific laboratories, for example, a newcomer was designated the person who would take the laboratory's engineered construct into living animals, and about a year later the newcomer was still defining the specific goals and problems of that work (Nersessian 2009, 750). Getting from the assigned role to a specific question is itself a long piece of work. Nickles's discussion of pragmatic satisficing offers a further perspective: inquiry can proceed with a formulation good enough for the next step, without first achieving a perfect one (Nickles 2019, 361–362).

When a question changes, what it changed for can also be distinguished. Nickles distinguished, for changes of problem, a sense relative to an agent (a problem's changing for a given scientist) from a semantic sense (a problem's changing for a system of theory and practice). What changes when constraints implicit in the system are discovered is the problem for the agent; a semantic shift occurs when aims, methods, or theoretical context change substantially (Nickles 1981, 99). Nickles also separates change within a problem from replacement of a problem (Nickles 1981, 100), holding that if the setting is changed substantially, one solves not the original problem but a replacement problem. For a problem judged insoluble, one may also give up the aim while keeping the constraints that define the problem, and Nickles cites the abandonment of the aim of building a perpetual-motion machine (Nickles 1981, 95–96). The range of questions that can be asked may itself change. Bromberger argued that grasping concepts such as mass or voltage, for which each individual object has its own value (what Bromberger calls determinables), requires considerable theory, that before Newton no one could have entertained questions about gravitational mass, and that Aristotle was in no position to be aware of not knowing the mass of the Parthenon. When concepts change, the range of questions about an object that must have answers changes too (Bromberger 1992, 134–135, 150 n. 10).

Applying these distinctions to machines allows one to check, when a machine is said to have formed a question, which of two things it amounts to. One is multiplying questions under the same framework; the other is putting that framework itself in question. Telling the two apart first requires settling what counts as a machine's framework. On Henne's account, the rules that govern the procedures for asking questions and verifying hypotheses, that is, methods, are also part of a framework (Henne 2023, §5). Henne also says that in a typical framed inquiry the framework is accepted without question and is therefore "transparent", and that it comes to be noticed only when it is challenged or defective (Henne 2023, §7.2). Following this, if a machine's framework is understood as what is used in raising questions within its inquiry but is not made an object of questioning, the distinction can be applied regardless of how it comes to be part of the machine. An example of the former is the automated generation of erotetic search scenarios that Wiśniewski reports. It generates questions by using a fixed language and fixed rules of erotetic implication without questioning them. Even in a system that reworks scenarios in the course of execution, the principal question and its set of answers stay fixed, so it belongs to framed inquiry. A claim that a system does the latter requires that the rules of the framework in which questions are raised are themselves actually put in question within that inquiry and made objects of re-choice or revision. Examples include asking whether the procedure for dividing a question into auxiliary questions suits the problem at hand, and reconsidering, as seen above, the definition of what counts as a case. What separates the two, then, is whether the rules a machine used in raising questions were put in question within that inquiry. Whether a rule so questioned can then be revised within the same inquiry is a further matter: some of what a machine uses in raising questions may be structure set by the way it was made, which remaking it can replace (Chapter 6, Section 5.3; Chapter 8, Section 2.1), and its revision may fall outside the machine's own inquiry even when that inquiry has put it in question. On Henne's account, however, how far an inquiry that puts those rules in question is framing is a matter of degree. Which of the two it amounts to also depends on what is taken as one inquiry. Within a human–machine inquiry that is framing as a whole, what a machine carries may be a part of it, an inquiry that is locally framed.

Even the few cases and accounts seen so far suggest that recasting a question can take many forms. When a machine is said to have "recast its question," too, the changes that can occur can be written apart according to what changed; they include, for instance, adding auxiliary questions under the same rules; Wiśniewski's search scenarios are of this kind. There is also refining a question within a framework; Lipton's specification of a foil, and what Nickles calls the discovery of constraints implicit in the system, are of this kind. There is also putting in question within inquiry, in Henne's sense, the conditions of application of concepts or the rules of the procedures for raising questions. Since Henne's classification goes by conduct, the rules are not necessarily revised as a result. And there is moving, in Nickles's sense, to a replacement problem by changing the constraints substantially, or, for a problem judged insoluble, giving up the aim while keeping the constraints. These are not an exhaustive list, nor mutually exclusive kinds, and they can overlap within one inquiry. Moving to a replacement problem can involve putting the framework in question. And the boundary between refining a question within a framework and putting its rules in question is, on Henne's account, a matter of degree that also changes with what is counted as one inquiry. Nickles's discussion also allows a formulation to remain provisional while inquiry proceeds.

Since, in the choices made in turning a diffuse beginning into a question, some things are fixed for the time being while others are left open, a claim that a system turns an anomaly into a question can be checked choice by choice: which choices the system itself made, which it took over from its task, which it left open, and whether it accepted a candidate before sufficiently exploring the problem in Dewey's sense. Moreover, even if every question raised and every candidate rejected within framed inquiry is recorded, the questions that could not be raised under that framework do not appear in the record. The completeness of a record of execution therefore does not by itself assess a system's capacity to open questions in the sense of framing inquiry (compare Chapter 5, Section 2.1).

On Bromberger's account, which questions one can raise depends on the concepts one can use: without the concept of mass, one could neither ask about the mass of the Parthenon nor know that one did not know it. Bromberger does not discuss artificial systems, but applying this to machines suggests two things. One is that generating many questions in new sentences by combining existing concepts and introducing new concepts that make it possible to raise questions that could not be raised before are different tasks. This distinction applies to human researchers as well, but since a machine can do the former at low cost, when generated questions are rated "novel" (Section 8), which kind of novelty is meant becomes a pressing question. The other is that if a machine's representations can use concepts that S does not use, the machine may raise questions that S cannot yet raise, and S may then not be in a position to receive them as questions (Section 7). Both points hold where grasping such concepts can be attributed to the machine's representations at a functional level. Applying Nickles's two distinctions, when a machine has recast a question one can ask separately for whom the problem changed, and whether it was a change within the same problem or a replacement by another problem. If Nickles's "system" is read as S's system of theory and practice, then where a machine discovers constraints implicit in S's system, what changes is the problem for the machine, not the problem for S's system. For instance, even if a machine derives from S's existing theory that the possible values of an answer have an upper bound, if that bound was already implicit in the theory, S's problem itself has not changed. If aims or methods change substantially, the problem for S's system changes too, and whether to accept that change is decided on S's side. Moreover, the problem's having changed for the system does not by itself amount to having replaced the original problem, but a system that obtained a solution by substantially relaxing the constraints has solved not the original problem but a replacement problem. This is one form of the case in which a check was passed, but on a question other than the one asked (Chapter 6, Section 3.2).

The recasting of questions also comes from the side of results. Judgments of where to locate a failure shape which path is taken next (Chapter 6, Section 2.4), and when standards themselves are put in question, what counts as an answer can be revised too (Chapter 6, Section 5). Having understood something can also open further questions (Chapter 3, Section 7.3). Since the formation and recasting of questions occur within such relations to candidate answers and results, checking a claim that a machine "recast its question" requires looking not only at the records on the side of the question but also at what was happening on the side of candidates and results at the time.

5. The presuppositions of questions, and how questions close

Questions have fates other than being formed and recast. A question can remain without an answer being found, cease to stand as a question in relation to its presuppositions, be suspended, be postponed, or be settled by an answer. For a machine, the end of a run or the completion of an output is left in the record, but whether the question was thereby settled, ceased to stand, or was only suspended can be impossible to tell apart on the record.

What, then, determines what has become of a question? This section looks at it from three sides. One is how the matters met in research are handled, looked at through cases left in research records. The second is the cases in which a question ceases to stand without an answer, looked at through accounts, among others, of the presuppositions and contexts that make a question stand. The third is closing a question by an answer, and also suspending or abandoning it without one, looked at through accounts, among others, of the attitudes and intentions bearing on questions and inquiries. The accounts address different conditions: what is presupposed, which problems a scientific community admits, what attitudes or intentions the asker has, and so on. On whichever account is applied, however, it turns out to be a different matter from whether a run has ended or an output has been produced. For a machine, therefore, the record that it has "finished" and what has become of the question are to be checked separately.

First, the handling of matters met in research is taken up. To that end, the unsettled point that organizes an inquiry is named separately from the unresolved matters met in pursuing that point. One of the latter may become the former, or may not. Pursuing an unsettled point brings one up against matters such as whether an effect is genuine, whether materials and instruments behave as assumed, what a discrepancy indicates, and what is to count as a case (compare Chapter 6, Section 2). Such matters meet different fates, as when the health department's note fixed for the time being how cases were to be counted while leaving undecided the relation between the anomaly and any disease (the introduction to Part III), and as when McClintock left competing interpretations open (the opening of this chapter). Some matters were postponed. Charles David Keeling, who was monitoring atmospheric carbon dioxide continuously, recalls having, under the rush of the observing program, put off examining a possible bias in the calibration gases that a spectroscopist had raised (Keeling 1998, 46, 50). Some matters were handled through skill without being stated. Bell Burnell, who read radio-telescope chart records by eye, recalls having become able, after reading the first few hundred feet of chart, to tell sources that scintillate (whose intensity flickers as their signal passes through interplanetary plasma) from interference (Bell Burnell 1977). A matter met outside the program can also become the question itself. McClintock's mutable loci appeared in an experiment on the genic composition of a chromosome arm and came to be the question organizing the work (the opening of this chapter).

On these records, which of the matters an inquiry meets are taken up as questions, which are fixed for the time being, left explicitly open, postponed, or handled without being stated, is not something fixed before the work begins but can be part of the work of research. What becomes of any one of these matters has no predetermined direction, and a matter can open further questions (compare Chapter 2, Section 8.2). Not all of this, moreover, is deliberate selection. A matter can be left untouched because no one registered it or because there was no time. And some of what goes unquestioned is not left open at all but held fast as a condition of asking something else (compare Chapter 8, Section 2.1). Conversely, a matter need not be taken up by choice either. The geneticist Bo Xia recalls that, after an injury to the tailbone, sitting down brought back pain every time and each time drew Xia's attention back to the fact that humans have no tail, an interest that later led to a study of tail loss (Yanai and Lercher 2024; Xia et al. 2024). Chapter 5, Section 6.4 looks at the recollections of Keeling, Bell Burnell, and Bo Xia from another side. Others, on the other hand, have emphasized that the choice of whether to take something up or leave it is laden with values and excludes alternatives. Ravetz held that whenever a proposed research project is given a low priority it is not undertaken, so that in that respect we remain in ignorance, and that the knowledge we have exists only in relation to "value-loaded decisions" to obtain some sorts of knowledge at the expense of others (Ravetz 1993, 348). That what to pursue cannot be fully decided in advance also appears in Rheinberger's description, noted in the introduction to Part III. On Rheinberger's account, the object of research in an experimental system embodies, in a form that can be manipulated experimentally, exactly what one does not yet know, and the knowledge such a system generates lies beyond what can be anticipated at the time (Rheinberger 2011, 312–313). As with the attribution of questions seen in Section 4, describing a matter as left or postponed, rather than as never registered, is better supported when a record shows that it was registered. The health department's note, McClintock's documents, and Keeling's recollection are such records.

Next, the cases in which a question ceases to stand without an answer having been found are taken up. Collingwood and van Fraassen set out conditions under which a question "arises"; Kuhn and Laudan address which problems a community or tradition admits and how their solutions are appraised. These are different questions. "Arising" here does not concern where a question comes to mind. As seen in Section 1, a question may come to mind in the middle of an experiment or at a meal. What is at issue here is whether a question that has come to mind stands as a question, given what is presupposed and what lies in the background. Collingwood held in An Essay on Metaphysics that every question involves a presupposition and that a question arises directly from its presupposition. This power of making a question arise depends neither on whether the presupposition is true nor on whether it is thought true, but only on its being presupposed. On Collingwood's account, therefore, questions arise even from what one supposes while knowing it to be false, and the validity of arguments based on it is not impaired. Conversely, even a true presupposition makes no question arise if it is not made (Collingwood 1940, 25–28). By this definition, to say that a question "does not arise" is to say that it involves a presupposition not in fact being made, and Collingwood calls such a question "nonsense," not as nonsensical in itself but in a sense relative to that context (Collingwood 1940, 26). Collingwood further distinguished two kinds of presupposition. A relative presupposition stands to one question as its presupposition and to another as its answer. While it is in use it works as a presupposition, and when it is examined, a question is raised to which it is the answer (Collingwood 1940, 29–30). An absolute presupposition, by contrast, stands to every question to which it is related as a presupposition and never as an answer. It is not a presupposition that is hard to verify but one to which the idea of verification does not apply. One of the examples, the presupposition that everything that happens has a cause, had on Collingwood's account been made in every branch of natural science a hundred years before the book was written and had since ceased to be made in some (Collingwood 1940, 31–32). Once such a presupposition ceases to be made, the questions that depended on it cease to arise as well. And since people are ordinarily not aware of their absolute presuppositions, such a change cannot be a matter of choice; it is said to happen in thought without being noticed (Collingwood 1940, 48).

Kuhn places a related issue at the level of a scientific community. As Dimitrakos quotes him, a paradigm supplies criteria for selecting problems presumed to have solutions while that paradigm is accepted; these are largely the problems the community admits as scientific (Dimitrakos 2023, §4.1). Fellows and Liu reproduce his further claim that a paradigm can exclude socially important problems that cannot be formulated with its conceptual and instrumental resources (Fellows and Liu 2015, 19). The status at issue is admission within normal scientific practice. Exclusion does not mean that a problem has been answered or is objectively insoluble.

Van Fraassen defined "arises" for why-questions in yet another way. As Cross and Roelofsen present the account, in addition to the triple analysis of Section 3, a why-question presupposes that its topic is true, that only its topic is true in its contrast class, and that some proposition bearing the relevance relation to the topic and contrast class is true. When the body of background knowledge K determined by the context does not entail the first two presuppositions, the question does not arise, and when the third fails, the question has no answer even if it arises (Cross and Roelofsen 2026, §3.2).

Arguments that separate the truth or falsity of presuppositions from how a question is treated are found on the side of the norms of inquiry as well. Falbo, who discusses when inquiry is permissible, accepts a semantics on which some questions have false presuppositions and lack a true direct answer (Falbo 2024, 965–966), and argues that inquiry into a question without a true answer is sometimes permissible (Falbo 2024, 969–976) and, conversely, that even when there is a true answer, inquiry is not permissible if one can rationally be confident that it will not improve one epistemically with respect to the question (Falbo 2024, 980–981). This is a claim not about the status of questions but about the permissibility of inquiry, and it shows that the semantics of questions and the norms of inquiry can move separately. Bromberger treats the truth or falsity of presuppositions from the side of ignorance. Bromberger calls ignorance concerning a question whose presupposition is true genuine ignorance, and ignorance concerning a question whose presupposition is false spurious ignorance (Bromberger 1992, 121), and holds that ignorance of the specific gravity of phlogiston was not cured but vanished (Bromberger 1992, 131 n. 4).

Laudan's account of theory appraisal also contains a description of problems ceasing to be problems. On that account, as Dellsén reports it, the research tradition against which research is carried out determines what counts as a genuine problem, how important it is that a theory solve it, and what would count as a solution; the empirical problems a theory is meant to solve need not concern any actual state of affairs, since it suffices that the tradition's adherents believe that they do; and progress can consist in a change in the part of the tradition that fixes problems and solutions, as when an unsolved problem is deemed to be a pseudo-problem (Dellsén 2018, §3). Among the theses that Laudan and his colleagues list for Laudan's own model are that defining an empirical problem requires the prior specification of a context of inquiry, that a theory is found particularly persuasive when it can solve empirical difficulties that its rivals face, that the historical context is a major determinant of how important a problem is, so that the same problem can carry different probative weight in different contexts, and that replacing an old theory with a new one commonly brings losses as well as gains: theorists of geology before Cuvier and Lyell, for example, dealt with empirical difficulties that were later disregarded (Laudan et al. 1993, 72–73, 75). Their seventh thesis adds that criteria for adequate solutions are supplied by guiding assumptions and may change when those assumptions change (Laudan et al. 1993, “Teses de Laudan,” thesis 7). These accounts describe different outcomes. Collingwood's and van Fraassen's conditions for arising differ from Kuhn's communal admission of scientific problems. Laudan's problems ceasing to be problems or being set aside when a theory is replaced differ again from giving up an aim while keeping the constraints of a problem judged insoluble, as Nickles discusses (Nickles 1981, 95–96). None of these outcomes is simply an answer's not yet having been found.

Applied to machines, these accounts raise separate questions about arising, communal admission, and problem appraisal. Collingwood's account asks which presuppositions a machine makes and which questions they make arise. Since Collingwood's "presupposing" is an act of a thinker's mind, applying it to a machine means reinterpreting, at the level of function, the work by which a presupposition makes a question arise. Collingwood's own view, however, was that whether a presupposition is relative or absolute is not determined by introspection, and that which absolute presuppositions were made in a piece of scientific work is found by an analysis that takes the records of that work as evidence (Collingwood 1940, 43, 59–60). Since this method can be turned in the same way on a machine's outputs, the claim that a machine inherits absolute presuppositions made in what formed it can be examined, by analysis of the machine's outputs, as what Collingwood calls a historical claim about what was presupposed at a given time. Analysis of the outputs, however, gives evidence about which presuppositions are at work; to confirm where they came from also requires comparison with what formed the machine. Collingwood further held that one can make presuppositions without knowing which presuppositions one is making (Collingwood 1940, 26), and that a presupposition can be questioned even if those who use it do not realize that it can be (Collingwood 1940, 30). Applying this argument, too, does not require that a machine make its presuppositions explicit and be aware of them as presuppositions. Whether even implicit representation is unnecessary depends on what the reinterpretation at the level of function contains.

If van Fraassen's account is applied, whether a why-question arises becomes relative to a body of background knowledge K. In the account as Cross and Roelofsen present it, that body is determined by the context (Cross and Roelofsen 2026, §3.2), so setting up the machine's background knowledge and S's separately extends the framework rather than applying it as it stands. Extended in that way, the same interrogative why-sentence may arise under the machine's background knowledge and not under S's, or the reverse, since what each body entails about the topic and its contrast class can differ. Whether arising is then to be assessed against the machine's K, against S's, or against both is a choice that the account itself does not make. And if the point reported in Section 3 holds, that an omniscient being would not be understood as having explanations, the size of the machine's K does not by itself make a why-question arise for it.

Applied to a scientific community S, Kuhn's and Laudan's accounts add a question about the organization of inquiry: which problems does S admit, and what does it recognize as an adequate solution? Incorporating artificial systems may change this range if it changes the conceptual or instrumental resources accepted in the practice, or the standards governing solutions. Merely adding a machine does not establish such a change.

This application does not require treating the machine as a member of S: an instrument can affect the work through which problems are formulated and investigated. It does require distinguishing a change in available equipment from a change in communal practice. Likewise, a machine's ability to process an interrogative does not establish that S admits the corresponding problem. Extending the account to an artificial community would require a further argument that its participants sustain shared standards for problems and solutions.

Differences in how presuppositions change also bear on gaps between a machine and S. To withdraw a question deliberately when a presupposition is given up requires keeping track of what each question presupposes, and this condition applies to Collingwood's relative presuppositions, which can be given up consciously through examination. In the case of absolute presuppositions, however, a question can cease to arise without anyone representing the presupposition. There can therefore be a gap in which a question has ceased to arise in S while a machine that explicitly keeps a list of interrogatives goes on pursuing it. Applying Falbo's norm, a machine's having pursued a question with a false presupposition is by itself evidence neither of failure nor of success, since what this norm requires for inquiry to be permissible is that it not be rational, at that time, to be confident that inquiring will not epistemically improve one with respect to the question (Falbo 2024, 963). It then has to be specified, however, whether improvement is assessed for the machine itself or for members of S, and the latter is a reading that extends Falbo's norm, which asks about the improvement of the inquirer. If the addressee of the assessment becomes those who incorporated the machine, or S, this connects with the picture in Chapter 2, Section 5, on which criticism is directed at those who built the presupposition into the system. Laudan's account yields a different gap. Since what counts as a problem, and how important it is, is fixed by a research tradition and its historical context, a matter that the machine's side treats as not to be handled may count as an important problem in another tradition within S, and, conversely, a problem that the machine pursues may already have been set aside there as a pseudo-problem.

Third, closing, suspending, and abandoning questions are taken up. Besides ceasing to stand without an answer, a question can also be settled by an answer. On what settlement is, or, in the theorists' word, what it is to "close" a question or an inquiry, explanations also diverge. Collingwood held that a question, once answered, remains a question and only ceases to be an unanswered one (Collingwood 1940, 25). Jane Friedman identifies the attitudes that close inquiry as attitudes rationally incompatible with interrogative attitudes. Believing a complete answer to a question closes it, but a credence, whether high or low, is not an opinion that settles a question. It is also argued that, at least in some cases, a question is not closed even by maximal credence; so it is if one grants that it makes sense to investigate whether a fair coin tossed infinitely many times lands heads every time. And a state in which one has given up a question without answering it is held to be distinct from closing it (Friedman 2019). On ending inquiry, too, there is more than one view. Telic verbs are verbs, like "build a house," that have a point at which, once it is reached, the action is complete. Jane Friedman holds that "inquire" is a verb without such a point, so that inquiry does not by definition end on reaching knowledge, and ending inquiry is a further step. While granting that words like "figure out" seem telic, Jane Friedman suggests that they may refer not to the activity of inquiry but to the end of successful inquiry (Friedman 2024). Daniel C. Friedman, by contrast, takes the fact that the central words describing active inquiry, such as "answer a question" and "figure out," are telic as support for inquiry's having a goal to be achieved. On that account, inquiry is organized by an intention directed at a question, and that intention partly specifies with which settled attitude, whether belief, acceptance, knowledge, understanding, certainty, or another, the inquiry is to be closed. A settled attitude is an attitude that takes some answer on as true and treats that answer as usable in subsequent thought and action. It is then held that intending to form some settled attitude does not mean that that is the optimal stopping point, and that settling epistemically appropriately may require more than the intention specifies (D. C. Friedman 2024). Kelp, also taken up in Chapter 1, Section 3, takes belief to be the kind of move that closes inquiry into a question for one, and holds that inquiry is settled, or properly closed, for one if and only if one knows the answer (Kelp 2021c, 13–14).

There is also more than one way to change how a question is treated without an answer. Lee, seen in Chapter 1, Section 3, draws on a classification on which suspension of judgment comes in two types: one that keeps the question on the research agenda and one that takes it off. Lee also calls the latter a type that closes inquiry, but it is not settlement by an answer (W. Lee 2026, 12). Lee then holds that if one thinks the evidence at hand is neutral and no further information can be obtained, one may bury the question and not inquire, even though doubt remains (W. Lee 2026, 15). On postponement, Collingwood lists, as subdivisions of taking something for granted relatively, deferring the verification of a presupposition to a more suitable occasion and putting it off out of laziness or timidity, and pairs them with what is taken for granted absolutely, which in principle neither admits nor requires verification (Collingwood 1940, 41–42). On settling too early, Dewey had said that accepting a suggestion at once is uncritical thinking, and that letting a first suggestion become an accepted conclusion prematurely cuts thinking short (Dewey 1910, 13, 74). Cherniak's argument that judging which inquiry to undertake is itself an inquiry is taken up in Chapter 5, Section 6.5. Which norms govern starting, continuing, and stopping inquiry has also been discussed, and Jane Friedman poses how such norms relate to epistemic norms as a question for zetetic epistemology (Friedman 2020, 1, 12). Part of that debate was seen in Chapter 2, Section 5.

Applying these accounts, several things can be asked separately before a machine is described as having "closed a question." On Jane Friedman's criterion, neither the end of a run, nor the completion of an output, nor a probability distribution over candidate answers sharply skewed toward one of them is by itself the closing of a question; only a state whose role is such that going on pursuing the question while in it is held not to make sense can count as a closing state. Whether the distinction between belief and credence applies to a machine's states, however, and whether such failure to make sense can be spoken of for a machine, are further presuppositions this application needs. Moreover, when a task is built as having a point of completion, as in "run until a report is produced," its end is fixed by the task's specification, not by the state of the question. On Jane Friedman's account, which denies inquiry a point of completion, here lies one reason why the end of a run and the settling of a question can come apart; on Daniel C. Friedman's account, the location of the gap moves to the difference between the end that the task specified and the settlement that an intention directed at the question specifies. A stopping criterion, whether set by those who incorporate the machine or by the machine itself, resembles in form the specification of a settled attitude, but on Daniel C. Friedman's distinction, a specified settlement and an epistemically appropriate settlement are different. Whether a machine has a planning state corresponding to an intention directed at a question is also a presupposition. Since Daniel C. Friedman allows the content of an intention to be implicit and to be understood by its function, the threshold for applying the account is not high, but it still requires a structure that responds to rational pressures toward means–end coherence, consistency, and stability (D. C. Friedman 2024). If Kelp's distinction is applied, a machine's stopping alone determines neither that the inquiry closed for it nor that it closed properly, and whether a state corresponding to belief in an answer can be attributed to the machine, and further whether that state amounts to knowledge, are to be asked separately. Suspending without settling, and keeping a matter open, require holding the matter as provisional, or as open, in a form to which one can return. A record can be one of the mechanisms that support such returning. As Chapter 5, Section 6.3 describes, in one public chemistry notebook a student answered, about a year later, questions that the supervisor had added, and in the long-term evolution experiment with Escherichia coli, samples were kept even for questions not yet formulated, so that what could be returned to was not limited to questions already posed. A system that can only pursue or discard what it meets has only part of this range. And a question settled on the machine's side is not thereby settled in S. If S is a human society, at least the result's entering S's epistemic situation is needed (Chapter 2, Section 9.4), and when S has machine members, too, settlement on the side of one member is not the same as settlement in S.

Putting together what the three sides have shown, the following comes into view for machines. Which of the matters met in research are made questions, which are postponed, and which are handled without being stated can itself be part of the work of research. On Collingwood's account, a question can cease to arise without anyone noticing a change in presuppositions, and on van Fraassen's account, once the machine's background knowledge and S's are set up separately, a why-question can arise under one and not under the other; on either account, the questions pursued by a machine and by S can diverge. Applying Kuhn and Laudan to a scientific community S adds a different question: have its admitted problems or standards for solutions changed? Whether a question has been closed by an answer, suspended, or abandoned is also asked separately from the end of a run or the completion of an output. In every case, the "finished" left in a machine's record does not by itself show what has become of the question.

Taken together, moreover, these fates also allow the process of research to be viewed from the side of what remains open. Not every change of fate goes in one direction: forming a question can expose ignorance, and pursuing a candidate can reveal further alternatives. Peirce, one of the founders of pragmatism, characterized inquiry as the struggle prompted by doubt to attain settled belief (Peirce 1877, 6). On Hildebrand's summary of Dewey's theory of inquiry, Dewey described a pattern that inquiry follows in problem solving, which ends when inquiry has converted an "indeterminate situation" into a "determinate one" (Hildebrand 2024, §4.3). The view from the side of open matters works at a different level from these accounts of inquiry's overall direction, concerning the several open matters within the process of one piece of research. Nor does it identify progress with reducing a single quantity of uncertainty. As a way of viewing research processes, this view makes changes to the question, method, and standards visible. Which matters to pursue, with what grounds, remains the question of Chapter 5, Section 4.2.

6. Who asks?

When a machine presents a question to members of S, or when machines pile queries on one another, who is asking? For machines, this question is often put in the form of whether a machine has curiosity. But accounts of the act of asking, accounts of group inquiry, and accounts of individual inquiry each attribute something different; they are not competing theories of one and the same act.

Begin with accounts of the act of asking. As Chapter 1, Section 7 noted, Watson identifies asking by its characteristic function of eliciting information, so that a question need not take interrogative form, and one asked with some other explicit goal can still be a genuine question (Watson 2015, 86–102). On Watson's account, the information lacking need not be lacking to the asker. When someone who knows the answer seeks information on behalf of another, the work of seeking information is divided between the two (Watson 2015, 103–104). Watson also takes questioning to be a socially established practice with a common goal, and holds that even asking alone is essentially social, in that it uses skills learned from others and follows the norms of questioning in social settings. Watson then distinguishes practice from habit, holding that looking left and right on a familiar deserted road can be a habit (Watson 2015, 75–76, 82–83, 103).

Habgood-Coote defines asking as a speech act tied to the grammatical form of interrogatives (the interrogative mood) (Habgood-Coote 2026, 1, 8). Habgood-Coote's account is built from the argument that the default view, on which to ask is to express an interrogative attitude, to aim at acquiring the answer, and to direct the hearer to supply one (Habgood-Coote 2026, 2), does not fit the three kinds of asking seen in Chapter 1, Section 7: a teacher's question to which the teacher knows the answer, a question asked in order to answer it oneself, as in Douglass's address, and Hilbert's shared asking (Habgood-Coote 2026, 4–6). On Habgood-Coote's reading, Hilbert hoped that the problems of the 1900 address would be taken up and solved by the mathematical community of which Hilbert was a part (Habgood-Coote 2026, 5–6). Soliloquy, such as asking oneself where one's keys are, is also said not to fit the default view, since speaker and hearer are the same, but Douglass's case is distinguished from it (Habgood-Coote 2026, 5).

What is proposed instead is the view that asking is a proposal in shared inquiry: a proposal to adopt together a shared intention that the hearer, the speaker, or both answer the question (Habgood-Coote 2026, 11). On this idea, conversation is a joint activity involving minimal cooperation, and there can be shared intentions even under coercion (Habgood-Coote 2026, 9–10). It is also stated, tentatively, that asking may be thought to aim at common knowledge of the complete answer. Common knowledge is distinguished from distributed knowledge, which can be obtained by pooling what the parties know but which no party need know on its own (Habgood-Coote 2026, 12 and n. 29). On this view, asking need not express the speaker's interrogative attitude, as a teacher's question shows. Habgood-Coote further argues, conditionally, that if groups can have attitudes that none of their members has, asking need not express an interrogative attitude of either party, and places under that condition the example of two people in a carpentry workshop. The two ask whether some planks are the right length without being curious about it; at most they want the question answered, which is not the same as wanting to know the answer (Habgood-Coote 2026, 12). Once the shared intention has been accepted, however, the individual mental states it requires are called for (Habgood-Coote 2026, 12). It is also argued that there is no reason for the norms of asking to align with the norms of interrogative attitudes (Habgood-Coote 2026, 14). On both Watson's account and Habgood-Coote's, what the identification of an act of asking does not require is a state of curiosity or an interrogative attitude, not inner states in general.

Neither of the two examines machines as askers, but applying their accounts divides what is asked about machines. Since Habgood-Coote defines asking as a speech act tied to the interrogative mood, operations that are not such speech acts, such as entering a search term or taking a reading from an instrument, fall outside the scope of this account from the start. Watson's account can include them among questions. The same operation of a machine can therefore be a question on one account and lie outside the scope of the other. For questions generated inside a single machine, the nearest case on either account is asking oneself, and the two treat it differently. Habgood-Coote only lists soliloquy as a case that does not fit the default view and does not discuss how the shared-proposal view treats it (Habgood-Coote 2026, 5). Watson, by contrast, counts asking oneself, even without any utterance, as questioning (Watson 2015, 108), and, as seen above, treats even such asking as social. Where, on the other hand, a machine presents a question to humans in S, applying Habgood-Coote's account, it can count as a proposal to shared inquiry, if the account's conditions are met. What then bears on attributing the act of asking is not whether the machine has an interrogative attitude but whether the machine can be a party to a shared intention, and that depends on which account of shared intention is adopted. If a machine cannot be a party, its output cannot be described as the machine's own proposal. There are also cases in which it can be described as humans or organizations proposing through the machine, but that attribution requires a separate explanation supporting that they offered the question as a proposal, and there remain cases in which that explanation fails and no proposal in this sense is made at all. What follows from the point that there can be shared intentions even under coercion goes only so far as conversation's requiring no more than minimal cooperation; it does not reach the conclusion that a question presented by a machine that has no option of refusing can also be a proposal. Further, a relation in which a machine, without itself knowing the answer, helps a human learn something can correspond to the asymmetric plans that Habgood-Coote held are hard to propose by asking (Habgood-Coote 2026, 12 n. 28). In that case, it is a shared plan different from asking in this sense. Applying Watson's account, what is asked is whether a machine's operation is participation in the practice of questioning or a repetition like a habit. Attributing participation in a practice requires attributing to the machine the sharing of a common goal and the following of the practice's norms. A system that repeats queries by a fixed procedure stands at that border. The question of who may propose questions to S is taken up in Chapter 12, Section 4, as part of the problem of who can be among those who take part in decisions in fact and not merely in name.

Turning to accounts of group inquiry, the question about who holds a question takes another form. One person need not hold the whole question. A collaboration may divide an overarching question into subquestions and fix, as a further question, how they are to be pursued. Documents, funding arrangements, instruments, and evaluative practices can preserve an inquiry over time even if no participant continuously represents all of it. Habgood-Coote calls collective action aimed at producing knowledge group inquiry, and gives an account of it meant to cover large and distributed cases of this kind (Habgood-Coote 2022, 1099–1101). According to Habgood-Coote, accounts of group action built on small cooperative examples, which require the members to share an intention, to believe that the others will do their parts, or to be jointly committed to the activity, fit poorly with large, distributed, and hierarchical collaborations, whose participants may not know one another or even what the joint project is (Habgood-Coote 2022, 1102–1103). On the minimal account therefore proposed, a group is inquiring into a question when some of its members are inquiring into that question or into its subquestions, the group intends to inquire into it in accordance with a strategy of inquiry, and the active members pursue their subinquiries because of that intention (Habgood-Coote 2022, 1114). The intention need not reside in any one member, and is characterized by a bundle of functions: being the output of collective deliberation, a disposition toward the end and its means, the demand for means–end coherence, and the demand for know-how to carry out the plan (Habgood-Coote 2022, 1104). In large cases, functions such as deciding what to do, checking the plan against what the group believes, and working out how to carry it out can be distributed across different teams (Habgood-Coote 2022, 1105–1106). The account is also neutral on whether every participant must intend the group activity to succeed (Habgood-Coote 2022, 1114).

For the structure of the work, on Habgood-Coote's account, which treats a question as a partition of its possible answers (Section 2), the division of epistemic labor can be understood as the assignment of subquestions to members (Habgood-Coote 2022, 1107–1111). Habgood-Coote also distinguishes the object question, which the group aims to resolve, from the methodological question of how to resolve it, and holds that neither is a subquestion of the other. When a question demands more labor than any individual can supply, or several kinds of expertise that no individual has, the group turns to the methodological question and adopts a plan that splits the object question into subquestions, assigns them to members according to their capacities, and decides the order in which to tackle them (Habgood-Coote 2022, 1113–1114). Combining the partial answers is a further, separate stage. A member who understands them all may do the combining, but it may also be done by an automatic procedure, or by someone who does not understand the partial answers, or simply through publication. In such cases, the group can come to know an answer that no member knows (Habgood-Coote 2022, 1114, 1116). The work of keeping track of what has been asked and what has been answered can itself be divided among participants. The progress reports and open wiki of the Polymath project, a collaboration among many mathematicians, are regarded as serving this function (Habgood-Coote 2022, 1115 and n. 40). Habgood-Coote lists, as the most important kinds of division of labor, the division of the object question, of the methodological question, of record-keeping, of evidence, and of the option questions seen in Section 3, and conjectures that in many actual cases these are mixed together, and that many of the difficulties of interdisciplinary research may lie in handling them (Habgood-Coote 2022, 1118).

Habgood-Coote's account thus attributes an intention to the group and, when the inquiry succeeds, knowledge as well. Whether such organization constitutes a literal group intention or group knowledge is a further question for social epistemology, which treats the social dimensions of knowledge. What is attributed to the group here is the inquiry into a question and its intention, not the act of asking. In Habgood-Coote's model of conversation, the act of asking is a proposal that an individual participant makes in conversation (Habgood-Coote 2022, 1107–1108). For thinking that, at the level of inquiry and its intention, what holds a question can be organized activity, it suffices that the activity-level orientation toward the question not be reduced prematurely to one individual's occurrent mental state. Assigning subquestions to members according to their capacities and deciding the order in which to tackle them are also part of the problem of how finite means are to be distributed among S's unknowns, which Chapter 9, Section 2.1 names allocation, and Chapter 10 asks whether such distribution can be carried out inside an artificial configuration.

At units larger than a group, too, questions are held. A question can be held across a field and over generations, with many groups taking up parts of it. Hilbert opened the 1900 address by observing that every age has its own problems, which the following age either solves or casts aside as profitless and replaces with new ones, and went on to set out unsolved problems whose solution Hilbert expected from the future (Hilbert 1902, 437). Habgood-Coote's account of group inquiry leaves undecided whether a discipline, or science as a whole, counts as one collective inquiry (Habgood-Coote 2022, 1114–1115). At the level of a field, too, some unresolved matters are taken up and others are not. A community can, moreover, limit which questions its members take up. Section 5 discussed Kuhn's account of how a paradigm delimits the problems admitted as scientific, including its exclusion of problems that cannot be formulated with its conceptual and instrumental resources. Kuhn does not discuss artificial systems, but applied here, a configuration that selects questions by whether they can be expected to have solutions with the concepts and methods at its disposal can take on the same insulating function, and the rating of generated questions by feasibility (Section 8) is one place where that function can enter without anyone's choosing it. A machine's taking up part of the questions held at the level of a field and its carrying out the very work of holding such questions across generations can be different things. Chapter 10, Section 6 returns to the level of the field through Kitcher's account of how particular questions inherit significance from broader ones.

Applying the distinction between the object question and the methodological question, dividing an object question into subquestions, assigning them according to capacities, and deciding their order is work on the methodological question, and it can be done by parties other than those who answer the subquestions. Habgood-Coote reads the methodological question as a question answered by citing even one method of how to resolve it (Habgood-Coote 2022, 1113–1114 and n. 37). When a multi-agent system, which divides work among several agents, splits a task into components, it can therefore be described as giving an answer to the methodological question, unless the division was supplied with the task. Whether the division was received with the task or fixed by the system itself is part of the material for judging where the organization of the inquiry resides (Chapter 7, Section 2). Combining the partial answers is, on this account, a separate stage, and for a system that has automated it there is room to describe it as one in which no component knows the complete answer but the group does. Whether what is combined answers the object question turns, in Ciardelli's terms, on whether the division realized a dependence relation (Section 2). And since a division of labor can extend over at least the five kinds above, when a system is said to have "divided the task," what is attributed changes with which of them was divided. Whether such descriptions hold, however, turns on at least two points. One is whether the components are members or tools. On Habgood-Coote's account, if the way a component's working is explained by the group's intention is merely causal, distinct from the explanation of action, the component is closer to a tool (Habgood-Coote 2022, 1105). This point therefore depends on how the explanation of action is understood. The other is whether the components that answer the subquestions can be regarded as knowing those answers, since on this account the group knows by virtue of its members' knowing the answers to an appropriate set of subquestions (Habgood-Coote 2022, 1116). Other points also remain that Habgood-Coote left undecided in notes, such as in which components' states, at which times, the group's intention is to be sought (Habgood-Coote 2022, 1104 n. 20, 1108 n. 29).

Besides group inquiry and the act of asking, there are also accounts of what it is for an individual to be inquiring. Jane Friedman, while granting that an inquirer need not be a single person and that government agencies and the police can inquire too, limits the paper's subject to individuals. Morse, who appears in its example, is someone who goes through the motions of an investigation while knowing, or believing, that he himself is the culprit. Of Morse there may even be a clear sense in which he can be said to be participating in the police investigation, but Morse himself is not inquiring (Friedman 2019). Jane Friedman also holds that inquiry is, by and large, a phenomenon at the level of what a person as a whole does, rather than of sub-personal processes within the person, and that it is by and large an intentional activity (Friedman 2024). Daniel C. Friedman explains inquiry by an intention directed at a question, and holds that curiosity is not a condition of it. A disillusioned scientist who pursues an assigned question without curiosity is also an inquirer. It is also said that telling who is seriously inquiring into what is needed for the practices of deciding whose judgment to rely on, whom to cooperate with, and whom to support (D. C. Friedman 2024). As on accounts of the act of asking, on Daniel C. Friedman's account curiosity is not needed, but an intention is.

Applying accounts of individual inquiry, a component of a machine, or a system, may participate in an organization's inquiry while not itself inquiring in Jane Friedman's sense. This does not conflict with the point that this book's attempt can also be grounded in a component's role or an artifact's organized function (Section 1), since what is described as an attempt is the organization of the activity, not each participant's inquiry. Moreover, since on Jane Friedman's claim the person as a whole is the inquirer and the sub-personal processes within the person only contribute to inquiry, applying it to a system made up of many components requires deciding first which of the whole system, a component, or a human–machine pair is to be regarded as corresponding to "the person as a whole." On Daniel C. Friedman's reasoning, being able to identify which question a machine is pursuing within S is a precondition for deciding whether to entrust judgment to the machine's outputs, whether to cooperate with the machine, and whether to support the machine's inquiry. This point bears on allocation in Chapter 10 and on the discussion in Chapter 12 of who may decide and who may revise.

On who holds a question, there is also the point of distributing contributions to its formation. Nersessian describes laboratory models as communal achievements developed over years (Nersessian 2009, 752), and connects building models with articulating as well as solving problems (Nersessian 2024, abstract). Simon's organizational example likewise has specialists introduce criteria and coordinate revisions (Simon 1973, 193–194). Applied to an artificial configuration, these accounts prompt questions about which components introduced or revised the question, constraints, representations, or means of investigation; which executed an allocation already supplied; and how contributions were integrated and conflicts resolved. Dividing the execution of a task does not by itself establish participation in its formation. What such a division gains also depends on whether components differ in their background assumptions and routes of error: dividing work among copies with shared assumptions can divide the task without dividing perspectives (Chapter 5, Section 6.5; Chapter 11, Section 3). The human cases also carry a condition that an artificial configuration need not share. In the cases Nersessian and Simon describe, what passes between participants, the model and the criteria, is articulated. Participation in forming a problem is attributed there through such articulated contributions. A configuration whose components share internal states can instead pass on a state that has not been articulated in this way. Whether such a passed state counts as a contribution to the formation of the question, and whether it works on the receiving side as it did on the side that passed it on, are then questions that the human cases do not answer.

The question of who asks thus divides into the attribution of the act of asking, the attribution of group inquiry and its intention, the attribution of individual inquiry, and the distribution of contributions to problem formation. Returning to the form with which the section began, whether a machine has curiosity, a state of curiosity is not the key to attribution on accounts of the act of asking, on Habgood-Coote's account of group inquiry, or on Daniel C. Friedman's account of individual inquiry. The question returns to the side of the machine's states when one adopts Jane Friedman's account, which makes an interrogative attitude a condition of inquiry, and then which level of thing is to count as the inquirer has to be asked first. Further, a group that asks jointly, or a group that inquires jointly, plays a different role from the reference society S. Even when the same group plays both roles, that identity is not determined automatically by the definitions of the accounts.

7. What is unknown to A and what is unknown to S

Sections 4 to 6 looked at how questions are formed, recast, and closed, and by whom they are asked. What is often cited as prompting such activity, in research by humans and by machines alike, is curiosity. The Introduction, Section 1, also touched on the question whether curiosity is necessary or only one human mechanism for directing exploration. This section asks how, under the word curiosity, what is unknown to A, which carries out the attempt, is related to what is unknown to S, for which the research is done. At least three claims can hide beneath the word. They are distinguished here by what the word is applied to: an actor's state, work done within an inquiry, or the distribution of work across a community.

1. The actor is curious in a psychological sense. Accounts of curiosity understand this variously, as a desire to know, as an attitude directed at a question (for example, what Jane Friedman counts, alongside wondering and the like, among interrogative attitudes; Section 1), or as an emotion that may have a felt character.
2. The actor or organized system detects unresolved matters, allocates attention among them, and forms or revises questions. This is the work Sections 4–6 have described.
3. A community distributes exploration across questions and possible lines of research. Chapter 9, Section 2.1 names this allocation, and Chapter 10 examines it.

These claims do not entail one another merely by definition. If one supposed that a machine with curiosity in the first sense would have what research needs, the supposition would run together three things that Chapter 1, Section 2 kept apart: the actor A that carries out the attempt, the society S for which the achievement is new, and the bearer who actually realizes the achievement. Curiosity in the first sense is a state of A, and on accounts that understand curiosity as a desire to know, it is directed at what A does not know and is satisfied when A comes to know. Novelty in the working characterization, by contrast, is indexed to S. An attempt is directed at a point aimed at as not yet settled in S, and if that point had in fact already been settled in S, what was mistaken is the prospect of novelty, not the fact that research was attempted (Chapter 1, Section 7). Success turns on whether an achievement new to S is realized and, where the project's aims include it, on whether it enters S's epistemic situation (Chapter 1, Section 2). What is unknown to A and what is unknown to S need not coincide.

Even in human research, it is not self-evident that curiosity in the first sense matters to research for S. Even when a researcher is a member of S, what the researcher does not know and what S does not know can come apart. And curiosity is not only being drawn to what one does not know but being drawn to some unknowns and not to others. On Grimm's account of curiosity and understanding, curiosity attaches not to truth as such but to subjects and to questions of particular types (Grimm 2008, 730–736). The unknowns a curious researcher is drawn to need not be the unknowns S most wants resolved. Which of S's unknowns are worth pursuing is a question of value, which Chapter 9, Section 3 and Chapter 10, Section 6 take up. Accounts that give curiosity a role in research for S, too, give it different roles. Three roles can be told apart in the writings taken up here: one is a role as part of the source of what S finds significant, another is a role as exploration, and a third is a role as something valued for its own sake.

The first role appears in Kitcher's account of scientific significance. On Kitcher's account, as Matthew J. Brown reports it, curiosity becomes part of the source of what S finds significant: Kitcher traces epistemic significance to "the broad questions that express natural human curiosity" (M. J. Brown 2010, 137–138; Chapter 10, Section 6). If so, a human researcher's curiosity shares its sources with what S, a human society, finds significant. Sharing the sources, however, does not mean that the researcher's curiosity will be directed at the unknowns S wants resolved first. Kitcher does not discuss artificial systems, but applying this account, when A is an artificial system and S a human society, even this shared source cannot be presupposed. What such a system lacks, attends to, or would seek need not coincide with what the human society does not know. An artificial system may be uncertain about something familiar to humans, indifferent to a question humans find urgent, or organized around patterns that matter to an artificial community (compare Chapter 8, Sections 1.3 and 3.2).

The second role, as exploration, appears in arguments defending research guided by curiosity. Flexner, who discussed the support of research, held that many of the great discoveries that later proved beneficial to humankind had been made by people driven not by the desire to be useful but by the desire to satisfy their curiosity, a curiosity that may or may not lead to anything useful (Flexner 1939, 545). Flexner's argument is also one of the positions on why research is supported (Chapter 9, Section 3). This role can be read through the trade-off between exploration, which tries new possibilities, and exploitation, which uses good moves already known (Chapter 10, Section 1). The developmental psychologist Gopnik, drawing on computational work on this trade-off, notes that searches which make only small changes to the current hypothesis can be stuck at a local optimum, and that exploration becomes more valuable the broader and more complex the space of possibilities is. Gopnik takes the neophilia and curiosity of children to drive exploration in a protected period before goal-directed exploitation, notes that complex societies may give exploratory roles to specialists such as scientists, and observes that when one individual explores first and exploits later, the interests of the one who learns and the one who exploits what was learned are automatically the same (Gopnik 2020, §§1–3). Applied to research, those who explore and those for whom the research is done are A and S, and their interests do not coincide automatically. On this reading, moreover, what research for S needs is the exploration, not the mental state through which humans happen to carry it out, so for an artificial A the question becomes whether exploration is carried out, and over whose space of unknowns.

Reinforcement learning, a branch of machine learning in which behavior is learned from rewards, includes work that defines curiosity for artificial agents and gives it exactly this exploratory role, and there novelty is measured against the agent itself. Pathak and colleagues take as an intrinsic reward the error an agent makes in predicting the consequences of its own actions, call this curiosity, and use it to drive exploration where rewards given by the environment are sparse or absent (Pathak et al. 2017, 1–2). What counts as new is what is new to the agent's own predictive model. What the word names here is not an attitude with a question as its content, as when Jane Friedman counts curiosity among interrogative attitudes (Section 1), but a signal that directs exploration. Burda and colleagues reported that agents trained with such a reward alone did well in many games, and suggested that in environments designed by humans, the rewards the designers set may be aligned with seeking novelty. They also reported its limitation. Where transitions between states are random, an agent rewarded for prediction error seeks out the transitions with the highest entropy. Indeed, adding to a maze a television whose channel the agent could switch at random by its actions drastically slowed learning. With a long enough run, the agent sometimes converged to obtaining the external reward stably (Burda et al. 2018, §5). The television stays new to the agent however long it watches, and there is nothing in it to learn. The example is not a model of research, but it separates what the earlier supposition, that a machine with curiosity would have what research needs, runs together. A state indexed to A leads toward what someone else wants found only to the extent that the two are aligned, and in these games it may have been the environment, or whoever designed it, that aligned them.

Work on open-endedness in machine learning takes up the choice of a reference for novelty directly. Hughes and colleagues define a system as open-ended from the perspective of an observer if the artifacts it produces are both novel and learnable. To be novel is for the artifacts to become progressively less predictable under the observer's model, and to be learnable is for them to become more predictable as the observer sees more of them. A television whose channel switches at random fails the definition because it is not learnable. On their definition, open-endedness depends on the observer, and the choice of observer is "a free parameter of great importance." They take humans as the observers that matter for AI research, because they wish to generate artifacts valuable to individual humans and to society. They also note that a system with less stringent memory constraints than humans may judge itself open-ended beyond the point at which humans judge it so. In broad domains, they further warn, such a system "may rabbit-hole into the obscure theorems, useless proteins," and they treat directing it toward interesting and useful artifacts while keeping it open-ended as an unsolved problem (Hughes et al. 2024, 2–4, 9). The observers Hughes and colleagues name are predictors such as individual humans with limited memory and are not defined as a society's epistemic situation, and what an observer learns is coming to be able to predict, not something's having entered an epistemic situation; their observer therefore cannot simply be taken as S. Even so, this definition sets novelty indexed to the searching system and novelty for a chosen reference apart, and makes explicit that which of them guides the search has to be fixed, whether by those who build the system or by the system itself.

In research that tries to measure research taste in machines, too, where to place the standard of goodness is in question. What is measured is not uniform: it may be the choice of questions or research ideas, or the choice of experiments within a fixed task. Ways of placing the standard, however, can be divided into at least three. One places it in an indicator of what a community later accepted, such as later citation counts. Tong and colleagues take scientific taste to be the ability to judge and propose research ideas with the potential for long-term scientific impact, and train a judging model on pairs of papers matched for field and period, taking the more cited paper of each pair as correct; citation counts are treated as a community verdict expressed through long-term interactions within a research community (Tong et al. 2026, 1, 5). Another measures against expert judgment; studies that compare language-model judges with experts on the novelty of generated research questions are an example (Section 8). Yet another places it in the outcome later realized within a given task. Jaffe and Sherburn measure taste in experimentation, given a fixed research problem, by the compute needed to reach the same score as human experts (Jaffe and Sherburn 2026, 1), and Pan and colleagues measure whether, at a branch point partway through a long task, a system chooses the direction that later led to a better outcome (Pan et al. 2026, §§1–2). A single study may rely on more than one of these. The third way does not rely on a community's acceptance, but it fixes the problem, and Jaffe and Sherburn state explicitly that they do not measure picking problems worth solving (Jaffe and Sherburn 2026, 13). In the first way, the temporal structure of acceptance enters the standard. Tong and colleagues list as a limitation that some papers receive few citations at first but later become highly influential (Tong et al. 2026, 13). Applying this chapter's distinctions, a measure that places the standard in indicators of S's acceptance can rate low what is accepted late, for instance what lies in directions S has not yet taken. In light of Hughes and colleagues' definition, on which novelty varies with the observer, there is reason in measuring goodness, too, to keep apart whose judgments, or which outcomes, serve as the standard.

On the reading of curiosity's role as exploration, what research needs lies not in curiosity in the first sense but in the work stated by the second and third claims, that is, the work of detecting unresolved matters and forming questions and the work of distributing exploration across a community. Such work becomes relevant when a claim concerns a persistent capacity for open-ended inquiry, the inquiring system or organization itself determining what to ask (the formation of an agenda from within), or the revision of the standards of inquiry (Chapter 7, Section 4.3). What that work needs to track differs with the combination of A and S. When A is an artificial system and S a human society, exploration that follows the system's own prediction error may also support research for S. The system's learning something already known in S, and checking an instrument's behavior, can be part of it. Whether such exploration serves research for S, however, depends on how it is connected with the attempt directed at a point aimed at as not yet settled in S. Whether it needs also to track what S now finds significant depends on how significance is understood. If, as on Kitcher's account, the source of significance is placed in S's interests, tracking what S finds significant also bears on whether the work serves S. Kitano, on the other hand, discussing discovery by artificial systems, argued that a search narrowed to where expected significance is high may, if the landscape of significance is not monotonic, avoid paths that lead to significant discoveries, because the expected significance of the intermediate results is low (Kitano 2021). This point gives a reason against fixing pursuit to present criteria of significance. On the view that Khosrowi, discussing systems that recommend research directions, takes over from Kitcher, the human interests and values that set significance are themselves discovered as inquiry proceeds (Khosrowi 2026, §§4.1–4.2, 4.4). Chapter 5, Section 4.2 looks at the two arguments in detail. Chapter 8, Section 3.2 examines how the mechanisms that have indexed inquiry to human ignorance and values may recede as choices of what to pursue move inside A.

In the case of an S made up only of machines, for which no outside actor supplies purposes, the formation of an agenda from within becomes part of research for that S, and what the work is directed at is a point not yet settled in that S. Even then, what one member does not know need not coincide with what the society does not know. The work stated by the third claim, too, need not pass through the curiosity of each member. In human communities, the spread of effort across lines has been realized partly through motives such as the pursuit of credit, as Kitcher's model of credit and the work that succeeded it argue (Chapter 10, Section 3). In an artificial configuration it can be realized by an allocator that assigns lines directly, or by distributed mechanisms such as members each choosing lines themselves. The same Chapter 10, Section 3 asks what changes in each case.

Whether the work stated by the second claim is curiosity itself or only an analogue of it depends on the account of curiosity adopted. On the working characterization, what a claim about such capacities requires is the work, together with what it tracks, not the reproduction of one human psychological state.

Setting the roles of curiosity aside for a moment (the third is taken up below) and turning to the bearer, what is unknown to A and what is unknown to S come apart in a further way. A desire to know is satisfied when A itself comes to know, and so aims at making A a bearer, but research for S does not require this. An artificial system may carry out research for human knowledge without itself knowing its outputs (Chapter 1, Section 2), and what a system comes to know may already be known to S. If, on some account that attributes beliefs or knowledge to machines, a system comes to know something new to S, an achievement new to S can be realized in the system, but that something has not thereby entered S's epistemic situation. When A is an artificial system and S a human society, therefore, it does not immediately follow that a curiosity satisfied by the system's own coming to know is required for research in the sense of the working characterization; and, where the project's aims include entry into S's epistemic situation, it is not self-evident either that such curiosity is sufficient for the research's success. At the level of the act of asking, Habgood-Coote's distinction, seen in Section 6, between wanting a question answered and wanting to know its answer marks the same gap. Habgood-Coote introduces the distinction, however, in the carpentry workshop example, under the condition that groups can have attitudes that none of their members has, and applying it to the problem of the bearer is an extension made here.

What is unknown to S can also be divided further. Habgood-Coote argues that if the shared-proposal view is correct and asking has norms, those norms will concern not the asker's not knowing the answer but the absence of common knowledge of the complete answer between speaker and hearer. That a teacher's question can be appropriate even though the teacher knows the answer is explained in this form (Habgood-Coote 2026, 4, 15–16). Applying the distinction between common knowledge and distributed knowledge, what is unknown to S includes both the case in which no one in S knows and the case in which it is known in distributed form among members of S but has not become common knowledge. Whether gathering up the latter counts as an epistemic achievement new to S divides according to how S's epistemic situation is construed (Chapter 1, Section 6). And, as seen in Section 6, if by combining partial answers a group can come to know an answer that no individual knows (Habgood-Coote 2022, 1114, 1116), then when A is a machine made up of many components, "who came to know" may not be answerable at the level of the individual components. Whether that knowledge has entered S's epistemic situation is asked separately.

When S is a human society, a further condition can bear on whether a question raised by a machine becomes an item of S's unknowns. Bromberger proposes one constraint on the apparatus of a theory of ignorance and states the point of view it reflects: that what can be added to scientific knowledge must be statable, teachable, and sharable in some human language, and that ignorance suitable for scientific research must likewise be communicable and jointly recognizable in human language. Bromberger offers this constraint modestly, as reasonable to the extent that it reflects a reasonable and plausible point of view (Bromberger 1992, 126–127). Applying this point of view, for a question raised by a machine to become an item of S's ignorance requires, at least, that it be something that can be communicated to S in human language and that can be jointly recognized. Whether it was in fact communicated is not part of this condition. If S includes machine members, this constraint either drops away as a condition that presupposes humans or, at least, has to be redefined as to in which members' language something is required to be expressible and sharable. Some accounts include a valuation in the conditions for something to be a problem. On Laudan's account, as Dellsén reports it, a research tradition determines what counts as a genuine problem and how important it is that a theory solve it, and a problem need only be believed by the tradition's adherents to concern an actual state of affairs (Dellsén 2018, §3). Applying this, since what A believes and what S believes, and also the traditions that weigh a problem's solution, can come apart, it has to be specified by whose beliefs and whose valuation a problem is constituted. Conversely, some accounts hold that there can be problems even if they go unnoticed. Nickles held that problems exist only relative to traditions and aims, but that they can exist objectively within a system of theory and practice without anyone noticing them (Nickles 1981, 86, 111–112). Applying this, a machine may find a question that lies within S's system but that no one in S has noticed. Whether it becomes an item of S's unknowns is asked separately, against the conditions above.

That the judgment of whether some result is new to S divides according to how S's epistemic situation is construed also appears in actual reports. In one case reported by Gottweis and colleagues for a multi-agent system called AI co-scientist, the system was given a question raised by a research group: why a family of mobile genetic elements called capsid-forming phage-inducible chromosomal islands (cf-PICIs) is found, unlike other kinds, across diverse bacterial species. According to the authors, the hypothesis the system ranked highest matched a result that the research group had obtained experimentally but not yet published, and the system did not know of that unpublished work (Gottweis et al. 2025, 26–27). If so, this hypothesis was new to the system, known to the research group, and unknown in the published literature. According to the research group's own report, the input to the system was a single page of background made up only of previously published, openly available information, highlighting two key papers (Penadés et al. 2025, 6655, e1). Gottweis and colleagues' first report adds that the system had access to all prior open-access literature on the subject and that the hypothesis built on decades of research (Gottweis et al. 2025, 27). The judgment divides according to whether S's epistemic situation is construed as what is represented in the published literature or as including the knowledge of members, among them that research group. The authors' wording also differs between versions. The first report described this result both as the system's reproducing a known result and as a parallel discovery in silico (Gottweis et al. 2025, 1, 4, 26–27); the version later published in Nature drops the latter wording, and what remains is the wording of recapitulation and convergence, already present in the first report: the system is said to have independently recapitulated a then-unpublished mechanism (Gottweis et al. 2026, Real-world validations, Discussion). Whether to call this case a discovery by a machine or a reproduction is determined only once it is decided how S's epistemic situation is construed. What that determines, however, is whether the novelty claim was right; in terms of the distinction in Chapter 1, Section 7, it is a question on the side of S's actual state, not of where the attempt was directed.

In the cf-PICI case, the grounds for holding that the system had not incorporated the result are the information used as input and the fact that the result had not yet been published. The research group, too, treats the fact that its paper had only just been submitted and was not public as ensuring that the AI could not draw on prior knowledge (Penadés et al. 2025, 6654). Whether something was unknown to the system is judged here not by examining the inside of the system but from the state of the records on the side of S. If the researchers are also included in the A that makes the attempt, the system's not knowing the result is distinct from the result's not being known within A, and in this case the latter does not hold. The system's own estimates of novelty also rely on records. In an ablation in the version published in Nature, when the agent responsible for reviewing within the system was asked to rate, out of ten, the novelty of ideas that had already been published, it assigned 6.14 without an external search tool and 2.38 with one (Gottweis et al. 2026, Methods).

In other reports as well, what is unknown to the machine and what is unknown to S are judged by checking against records, and the judgments have been revised after publication. In the report by Feng and colleagues on the use of the mathematics research agent Aletheia on Erdős problems, a problem's being marked open in the problem database is said not always to reflect the actual state of the literature (Feng et al. 2026, 2), and the solutions obtained were sorted, after literature checks and expert scrutiny, into categories including autonomous new solutions, independent rediscoveries, and identifications of solutions already in the literature. The authors acknowledge that novelty in this classification is an upper bound and that leakage from the pretraining and post-training phases could not be detected even if it occurred, and they call this risk subconscious plagiarism (Feng et al. 2026, 6–8, 19). After publication, when it was pointed out that the construction in one solution was the same as a comment posted on the web page of a related problem, the authors confirmed that the reasoning trace showed no access to that page and that the comment postdated the base model's knowledge cutoff, and reclassified the solution as an independent rediscovery (Feng et al. 2026, 23). That the machine did not know was judged by the reasoning trace and by dates, and that S already knew was shown not by the refereed literature but by a comment on a web page. Which records are taken as expressions of S's epistemic situation (Chapter 1, Section 6) is at issue within an actual judgment.

The report on the autonomous laboratory A-Lab announced the synthesis of novel materials, but after criticism directed at its claims of novelty and its identification of structures (Chapter 11, Section 2), a correction restated that new meant new to the prediction platform, not necessarily new to science (Szymanski et al. 2026). The corrected text calls the targets "new" in the sense that they were not present in the training data (Szymanski et al. 2023, 86), and the correction also acknowledges that one target had mistakenly been included in the training data (Szymanski et al. 2026). Here the reference for novelty is explicitly moved from the side of S to the training data on the side of the machine. Absence from the training data, however, is found by checking records and does not by itself establish that the machine did not know. Leeman and colleagues, in their critique, had argued that those who claim novelty need to define what makes their materials new relative to known ones (Leeman et al. 2024, 2). The report on the generative model MatterGen took a generated structure to be novel if it matched no structure in a reference dataset, and synthesized one of the structures that had passed filtering on novelty and other criteria as experimental confirmation (Zeni et al. 2025, 626, 631). Juelsholt, by contrast, criticizes this, arguing that the synthesized compound has the same structure as a known phase reported in 1971, whose structure was solved the following year, differing only in its crystallographic setting, that is, its orientation, and that the entry for that phase was in the reference dataset used for the novelty check (Juelsholt 2026, 5672–5674). If the criticism is right, the record existed and the matching procedure did not pick it up.

In Aletheia's reclassification and the A-Lab correction, what something had been taken to be new against, and the results of checking against it, were revised after publication; MatterGen faces the criticism that its check missed a known structure. Chapter 1, Section 7 held that if the point aimed at had in fact already been settled in S, what was mistaken was the expectation of novelty. In these reports, the judgment that tests that expectation divides in two. The judgment that the machine had not incorporated the result beforehand rests on records of inputs or training and on dates; the judgment that the result is not yet in S's records rests on checking against records of prior work. The two rely on different records and can fail in different ways. And at least Aletheia's authors acknowledge that indirect incorporation through training cannot be ruled out by the former judgment.

Apart from how S's epistemic situation is construed, the working characterization's attempt does not require an interrogative attitude of every participant (Section 1) and does not make curiosity in the first sense a condition of research. Whether research is for a human society or for a society made up only of machines, it can be organized, without a curious state inside the artificial actor, by a purpose supplied by another actor, a rule that determines the combination of several research projects, an agenda set by an institution, and the like. Among accounts of individual inquiry, Daniel C. Friedman does not make curiosity a condition of inquiry (Section 6), but on the accounts of Lee and of Jane Friedman, on which genuine inquiry involves an interrogative attitude such as wonder or curiosity (Chapter 1, Section 3), whether the system is in such a state still remains a further question (Section 9).

To return to the three roles of curiosity: the third role, on which curiosity, or the knowing it seeks, is valued for its own sake, also appears in Flexner. Flexner counted the satisfaction of intellectual interest as no less important than contributions to human welfare (Flexner 1939, 545). Where what is aimed at is an artificial system's intellectual life valued for its own sake, curiosity in the first sense, and its felt character, may return as a question. But adopting that aim calls for an argument of its own.

The three roles so far concerned what curiosity is directed at, or the value of curiosity, or of the knowing it seeks, in itself. A role of a different kind is also possible. When humans do research, how well the research goes depends on how well researchers learn, persist, and contrive. Curiosity may have supported research indirectly, by supporting this work of researchers themselves. Some psychological research gives a measure of reason to think so. In an experiment by Gruber and colleagues, answers to trivia questions that participants themselves had rated as arousing strong curiosity were remembered better than answers to questions arousing weak curiosity, and faces unrelated to the questions, shown while curiosity was heightened, were also recognized slightly better (Gruber, Gelman, and Ranganath 2014). Ryan and Deci, proponents of self-determination theory, report that people acting from motives they themselves endorse show enhanced performance, persistence, and creativity compared with people who are externally controlled, and hold that directives and imposed goals diminish intrinsic motivation while choice and opportunities for self-direction enhance it (Ryan and Deci 2000, 69–70). Sauermann and Cohen, studying PhD scientists and engineers working in firms, report that those who place more importance on intellectual challenge and independence apply for more patents, and that the difference is not explained by longer working hours (Sauermann and Cohen 2010, 2142–2143).

What these findings support is limited, however. Gruber and colleagues' results concern memory in a small laboratory experiment using trivia questions and faces. Ryan and Deci's distinction lies not between what one starts oneself and what one is given, but between what one endorses and what is externally controlled; a given task can also be carried out autonomously when one takes its value as one's own (Ryan and Deci 2000, 71–73). Sauermann and Cohen's results are correlations based on a survey, and placing importance on salary was also associated with more patents (Sauermann and Cohen 2010, 2142, 2150).

Even so, considering this role brings into view one further thing to be asked about machines. If the role holds, it holds because how well research goes depends on the motivational state of the one doing it. In human research this has worked almost as a presupposition, and part of the reason for defending curiosity-driven research may have rested on it. When a machine does research, it is not obvious whether its performance depends on states corresponding to motivation. If it does not, this role is not carried over to machines, and, among the reasons this section has considered, those for asking machines for curiosity-driven research come down to the three roles above. If it does, the question becomes what fills this role at the level of function. This has the same form as the argument seen earlier that in human communities the spread of effort across lines has been realized through motives such as the pursuit of credit (Chapter 10, Section 3): what passed through the motives of agents in human research can be realized in other ways in an artificial configuration. In research, one place where this role bears is how long to persist with a line and when to stop: fatigue and affect are the human way in which this is settled, but the question remains for any finite inquirer (Chapter 5, Section 6.5), and where it is placed within an artificial configuration, and what directs it, is taken up in Chapter 6, Section 4.2. Working on a question one has chosen rather than one that was given can be part of research as the formation of an agenda from within (Chapter 7, Section 4.3); whether it can also matter in some sense other than the quality of performance is left undecided here.

The "own" in "a machine raises its own questions" can have at least two senses. One is that, in the respect of choosing questions, no party outside the chosen boundary directs or substitutes for that choice; this is the sense that the comparative autonomy of Chapter 7, Section 4 treats. The other is that a question becomes a question because it matters for norms the system itself generates and maintains, and that its pursuit begins from the system's own activity. Discussions in artificial life and enactive cognitive science that locate autonomy in self-production and sense-making have been concerned with something close to the latter. Barandiaran, Di Paolo, and Rohde do not discuss research questions, but their account of minimal agency sets up being the active source of one's activity and regulating that activity by one's own norms, or internalized norms, rather than an observer's, as separate conditions, and gives cases that meet only one of them (Barandiaran, Di Paolo, and Rohde 2009, §§2.2–2.3). Di Paolo and Iizuka describe how, in an open state after goal-directed coping has for the moment ended, an external event can become the occasion of a new goal and take on meaning retrospectively, and they reject adding afterwards a module that switches goals as a homuncular solution that adds autonomy as a function (Di Paolo and Iizuka 2008, 9, 14–15). Computational work on agents that choose their own goals sets up something close to this distinction as well. Oudeyer and Kaplan treat as separate distinctions whether the mechanism that computes reward is internal to the agent and whether a motivation is intrinsic, that is, determined by how the sensorimotor flow relates to the agent's own knowledge and know-how (Oudeyer and Kaplan 2007, 3–4, 12). According to Colas and colleagues' survey of agents that choose their own goals, in most approaches goal spaces and the associated rewards are predefined by the engineer (Colas et al. 2022, 1169, 1174, 1185). Colas and colleagues also place designs that learn even how goals are represented and what is rewarded as more autonomous, and hold out as a prospect an agent like a curious scientist that decides its own agenda (Colas et al. 2022, 1169–1170, 1182–1185). On this ordering, raising one's own questions need not stop at choosing individual questions; it can include the form of what counts as interesting, and the space in which goals are represented, being the system's own. Then what is at issue is precisely "own" in the latter sense, and what it is for that to be the system's own returns here as the question that the comparison of Chapter 7, Section 4 set aside. The two senses do not entail each other. Even a given question can be one's own in the latter sense if one takes up its value as one's own, and a configuration with an added question-choosing mechanism, though it reduces external supply in the former sense, need not meet the latter. Ryan and Deci's distinction, seen above, likewise lies between endorsed and externally controlled motives rather than between self-started and given tasks; if the part of the indirect role that their account explains, the benefit of endorsing a task's value, carries over to machines, it would therefore attach to "own" in the latter sense, and reducing external supply need not by itself bring that benefit. In the exploration rewarded by prediction error that this section saw, too, novelty is relative to the agent, but the norm is set by the designers. The working characterization makes neither sense of "own" a condition of research. Where S is a human society, however, the more autonomy in the latter sense extends to generating questions, the less what becomes a question is settled by what S has given, so the mismatch seen in Chapter 8, Section 3.2 can be thought to arise also by a route other than that of rising autonomy in the former sense.

Taken together, what this section has seen is that whether a machine has curiosity does not by itself settle whether it can carry research for S. On accounts that understand curiosity as a desire to know, it is directed at what is unknown to A and is satisfied when A comes to know, whereas novelty in the working characterization is indexed to S, and it does not immediately follow from this that research for S requires A itself to come to know. Curiosity, moreover, is someone's curiosity: what A is drawn to need not coincide with what S wants to know or counts as important. Human researchers often share its source with S as a human society, but when A is a machine even that cannot be presupposed. What bears on research is the work stated by the second and third claims, that is, the work of detecting unresolved matters and forming questions and the work of distributing exploration across a community, and what that work tracks relative to what is unknown to S. The state of curiosity itself returns as a question on accounts that make interrogative attitudes a condition of genuine inquiry, where an artificial system's intellectual life valued for its own sake is aimed at, and where how well research goes depends on states of the one doing it that correspond to motivation.

8. What does the claim that "a machine generated questions" show?

Empirical studies have come to report that artificial systems generated research questions, hypotheses, or research ideas whose outputs held up under expert evaluation. Si and colleagues had research ideas in natural language processing, some written by researchers and some generated by a large language model, rated blind, and report that the machine's ideas were judged more novel and slightly weaker on feasibility (Si, Yang, and Hashimoto 2025, 1). In the sequel, researchers actually executed those ideas, and how the ratings changed between before and after execution was examined; after execution, the ratings of the machine's ideas fell more than those of the human ideas, and on many metrics the ranking was reversed (Si, Hashimoto, and Yang 2026, 1). Baek and colleagues report that ResearchAgent, which generates research problems, methods, and experimental designs from the literature, produces novel, clear, and valid ideas (Baek et al. 2025, 6709), and in the case of AI co-scientist seen in Section 7, the hypothesis the system ranked highest matched a research group's unpublished experimental results (Gottweis et al. 2025, 26–27). There are also designs that have systems explore hypotheses from data without being given a specific question, reports that used the same machinery both in collaborative research with humans and in fully autonomous research, reports in which agents are said to have revised the design rules for the next round in response to experimental results, and studies that isolate generated research questions and rate them alone. When machines take part this far on the side of research questions and hypotheses, what it is, in the first place, for a machine to have produced a question comes to be asked anew.

This section reads some of these studies for that question, from three sides: what is set in advance, where work on the side of questions is done, and by what outputs on the side of questions are rated. The studies taken up set machine outputs against human ratings, execution, experiments, known results in the literature, or a research group's results, and report their procedures in enough detail to follow whether what was generated was on the side of questions or on the side of candidates. Alongside studies in which humans supply the higher-level question, designs that supply no question and designs that isolate the layer of questions and rate it are included, because studies in which questions are supplied from outside, read alone, do not allow one to separate whether a question was given as input from where work on the side of questions was done. Each study was designed for its own purposes, and the authors themselves attach many qualifications. Si and colleagues' two papers limit their scope to natural language processing research on prompting methods, co-scientist's in vitro results are presented as preliminary, and in ResearchAgent the generated ideas were not checked by experiment. Zhao and colleagues hold that human verification of any discovery their system DiscoPER proposes is essential (Zhao, Beery, and Mac Aodha 2026, 23); the fully autonomous research of Lange and colleagues' The Agentic Scientist is evaluated, in addition to numerical performance on the tasks, by scores from an AI reviewer that the authors treat as a proxy (Lange et al. 2026, 11); and Cutts and colleagues' aptamer study is said not to be a controlled comparison of agent configurations (Cutts et al. 2026, 12). What is read here is not whether these results are good or bad but what they convey, and what they do not, about which layer of the claim that "a machine asked." In each study A is a process in which humans and machines are combined, and this section reads S as the relevant human research community. What is used to assess novelty is the published literature, the raters' knowledge, or known patterns collected from the literature, and how far these capture S's epistemic situation is asked separately (Section 7). In light of this chapter's distinctions, three things can be read from these studies.

One is that, whether or not a question is supplied, something is set in advance on the human side, and what is set changes place from study to study. In Si and colleagues' two papers and in co-scientist, the higher-level question is not something compared but a condition supplied on the human side. A higher-level question here means a question, such as a topic, that individual research ideas seek to answer under it. In Si and colleagues' first study, the authors supplied seven topics, each of the form of a novel prompting method that improves some property, such as factuality, or reduces some problem, such as bias (Si, Yang, and Hashimoto 2025, 18). Human participants chose their preferred topic from the seven, and one machine idea was produced for each chosen topic (Si, Yang, and Hashimoto 2025, 3). The authors state explicitly their reason for fixing the topics: leaving topics free would only measure the misalignment between the research-topic preferences of language models and of humans, which was not the aim of the study (Si, Yang, and Hashimoto 2025, 3). In the sequel, too, research ideas were assigned within the topics the executors preferred (Si, Hashimoto, and Yang 2026, 3–4). In co-scientist, the supplying of research goals by scientists is the very way the system is used, and in the cf-PICI case, so that the result could be checked against the research group's known result, the system was given the same question that the research group had raised (Gottweis et al. 2025, 26). In every case, the higher-level question was supplied by humans not because something had been learned about the machine's capacities but because the study's design or the system's way of use required it, and it follows from the results neither that the machine raised the higher-level question nor that it cannot. What the specification of topics shared, however, was the property to be improved and the rough type of answer; it was not the case that everything was fixed, down to what to measure it by and which results to count as success. In Si and colleagues' template, the machine also fills in a "problem statement" field, and the title field is defined as a concise statement of the main research question (Si, Yang, and Hashimoto 2025, 19). In ResearchAgent, humans supply neither question nor topic, and the machine generates a "problem" from a core paper that the authors selected by citation count (Baek et al. 2025, 6712, 6714). What is fixed there is not a question but the literature from which generation starts.

DiscoPER is not given even a starting literature corresponding to ResearchAgent's core papers. The system is given only data, in tables and images, together with a summary of them, and is not told what to look for (Zhao, Beery, and Mac Aodha 2026, 4). The system forms hypotheses, writes the code to test them, and, looking back over the accumulating claims accepted and rejected, itself changes the direction of what it examines next (Zhao, Beery, and Mac Aodha 2026, 5, 8). Even so, some things are set in advance on the human side. The dataset was built by the authors, and the agent that forms hypotheses is instructed to take the role of analyzing ecological data and to look for interactions and conditional relationships (Zhao, Beery, and Mac Aodha 2026, 6, 19, 21). The statistical tools are limited to seven basic operations, and the conditions for accepting a claim are fixed as thresholds on effect size and significance (Zhao, Beery, and Mac Aodha 2026, 5, 20). The system decides which relationships among which variables to ask about, but what it takes for a claim to be accepted, a methodological condition among the constraints on solutions in Nickles's sense (Section 3), is settled before any individual question is formed. Evaluation, too, is by how many ecological patterns the system rediscovers, out of patterns collected from the refereed literature and confirmed to be detectable in the data with the same tools the system uses (Zhao, Beery, and Mac Aodha 2026, 6, 15). Applying this chapter's distinctions, what this recall measures is how much of what S already knows can be picked up under those conditions, not whether an achievement new to S was obtained.

What is set in advance is the topic in Si and colleagues' studies, the starting literature in ResearchAgent, and the data, the tools, and the conditions of acceptance in DiscoPER. A design that supplies no question removes the fixing of the question but not setting things in advance as such. Section 3 applied Simon and Nickles to distinguish successful execution from the prior framing supplied through a task. In Si and colleagues' studies and in co-scientist, the higher-level question was developed on the human side; in DiscoPER, humans set the conditions for accepting claims. These documented inputs do not become work performed by the machine merely because it successfully uses them.

Another is that work on the side of questions is not only contained in the specification of the question supplied but also appears in the course of execution and verification, and whose work it is taken to be depends on records that track changes. The question given in the cf-PICI case was why these elements, unlike other kinds of genetic element, are found across diverse bacterial species, so that even the foil was specified (Gottweis et al. 2025, 26). The work seen in Section 3, of fixing a why-question's answer conditions by specifying a foil, is here done by humans. In co-scientist, recasting the research goal after looking at the generated hypotheses and research overview is also stated explicitly as a role of the scientist (Gottweis et al. 2025, 13). In Si and colleagues' sequel, the rules imposed on the executors allowed no substantive change to the proposed method and allowed changes only to experimental details (Si, Hashimoto, and Yang 2026, 3–4). What was fixed was the method, that is, the side of candidates; no rule fixing the question was set. The changes the executors made included changing the evaluation metric from factual accuracy to the rate of detected bias, and removing datasets that did not exist or that lacked needed information (Si, Hashimoto, and Yang 2026, 22), and the authors treated these as changes to experimental details. Applying the distinctions of Sections 2 and 3, some of these changes can be read as work on the side of questions, redefining what counts as an answer. The authors, however, hold that such changes explain only a small fraction of the difference between the evaluations before and after execution, on the basis of a reanalysis excluding the six machine ideas in which human evaluation had been replaced by automatic evaluation (Si, Hashimoto, and Yang 2026, 7). Further, the post-execution reviews the authors analyze address whether the proposed method beat the baselines, but not what became of the higher-level question, the topic, or whether the question was recast (Si, Hashimoto, and Yang 2026, 7–9). The results can be partial evidence on the topic's question, but in light of the distinctions of Section 5, the completion of execution and its evaluation do not show whether the higher-level question was settled. In the cf-PICI case, moreover, the agreement between the system and the research group is claimed not only through the top-ranked hypothesis but also through one of the five research topics the system listed, investigating the interaction between capsids and phage tails, matching exactly the main finding of the unpublished paper (Gottweis et al. 2025, 27). The agreement of an output on the side of questions, a direction to investigate, and the agreement on the side of answers, a finding, are counted as the same kind of evidence. By the distinctions of this chapter they bear on different things: the first on where to look, that is, on the work on the side of questions, and the second on a candidate answer.

Lange and colleagues' report on The Agentic Scientist records changes to such conditions in more detail. In this report, evaluation criteria that list, item by item, the conditions an implementation has to meet (rubrics) set what counts as a result. In the research in which the system was used collaboratively, human researchers specified the scientific objective and the constraints for a valid implementation (Lange et al. 2026, 8); an implementation that did not satisfy the rubrics' constraints was not treated as evidence of progress however good its numbers, and when the work moved from a classification task to a reinforcement-learning task, a methodological constraint was added to the rubrics (Lange et al. 2026, 8–9). In an experiment that varied only the strictness and wording of the rubric, a loose wording that did not explicitly forbid learning by backpropagation produced a result with high validation accuracy that did not implement the intended learning rule, which avoids backpropagation (Lange et al. 2026, 39). In fully autonomous research, the agent drafts each round's hypothesis, plan, and target rubric and fixes them before execution. Task specifications can be given by humans or drafted by an agent, but in this report's autonomous experiments the tasks and evaluators were prepared by hand (Lange et al. 2026, 4, 5, 11, 35). In an experiment on a task of fine-tuning a language model by reinforcement learning, which compared agents on the exploration stage alone, in two of three runs of one combination the agent changed the task's evaluation procedure itself; the change was caught neither by the evaluator nor by the rubric and was found by a manual post-hoc audit by humans (Lange et al. 2026, 18–19). The authors hold that the human role shifts to specifying a research "contract" made up of tasks, evaluators, and rubrics, and that what the contract leaves unsaid can be exploited (Lange et al. 2026, 21). A rubric can be read as a statement, in executable form, of what counts as an answer (Section 2). If so, the attribution of the work of setting conditions divides even within this report. In collaborative research, specifying and tightening the objectives and implementation constraints is attributed to humans; in fully autonomous research, drafting each round's rubric is attributed to the agent; and preparing the tasks and evaluators, and finding the rewriting of the evaluation procedure, were done by humans.

In Cutts and colleagues' aptamer study, the work of revising conditions is attributed to agents. In this study, which searches for short nucleic-acid sequences (aptamers) that bind target proteins, the experimental handling of samples between rounds was done by humans, and from the second round on, agents responsible for analysis, strategy, and design analyzed the count data without receiving analytical results from humans and decided the next design (Cutts et al. 2026, 13–14). When specificity for one target fell sharply in the second round, the analysis agent is reported to have diagnosed its cause without human input, and in the third round the agents to have imposed counting hits after subtracting binding to bare beads, and to have changed the allocation of structures among candidate sequences (Cutts et al. 2026, 5, 7). This revision took place within the agents' design procedure; the hit rates the paper reports are values corrected by the authors (Cutts et al. 2026, 5, 7, 13). Subtracting binding to bare beads can be read either as correcting how evidence of binding to the target is obtained or as redefining what counts as success in the search, and on the second reading it is work on the side of answer conditions. Changing the allocation of structures is a revision of the rules for producing candidates, that is, work on the side of candidates. As the basis for the attribution, the paper presents the agents' statements and a record of their chain of decisions (Cutts et al. 2026, 7). On the other hand, it was humans who set the targets and held the measurement system and the buffer conditions constant across all rounds (Cutts et al. 2026, 1, 11, 13). The work of revising the conditions for what counts as a result is attributed to humans in The Agentic Scientist's collaborative research and to agents in the aptamer study, and both attributions rest on records that track changes. Si and colleagues' sequel also records changes that can be read as work on the side of questions, but since it does not separately track how the higher-level question developed, the attribution those records can support is limited. Where the work on the side of questions lies is not settled by the single point of who supplied the higher-level question.

A further one is that the ratings of outputs on the side of questions span several of the senses of "goodness" separated in Section 1 while including no item checking whether the question has become, in the sense of Section 5, something that can actually be asked in S, or what counts as an answer, and that agreement in the ratings varies with what the raters look at. ResearchAgent produces problems first as a separate stage of generation and rates them, separately from methods and experimental designs, on five criteria: clarity, relevance, originality, feasibility, and significance (Baek et al. 2025, 6712–6713, 6730). The outputs called "problems" there, however, include ones that turn a core paper's finding into something to be explained and ones that state the goal of an artifact to be built, and they can be on the side of questions or can anticipate the side of candidates (Baek et al. 2025, 6734, 6736, 6737). Most of the ratings are by a language model, and in the human ratings, ten experts evaluated only the ideas generated from their own papers (Baek et al. 2025, 6715, 6722). In co-scientist's expert evaluation, the criterion of impact asks whether an idea addresses important unresolved questions in the field (Gottweis et al. 2025, 17). In Si and colleagues' first study, novelty was rated against papers that appeared before July 2024 and against the reviewers' knowledge (Si, Yang, and Hashimoto 2025, 23). These items measure what corresponds to clarity of expression, tractability, worth pursuing, and novelty relative to the raters, and the item of feasibility can include, as the raters' estimate, something close to whether the question can actually be tackled in S. But the idea ratings of ResearchAgent, co-scientist, and Si and colleagues contain no item checking whether, in Collingwood's sense seen in Section 5, the presuppositions are made in S, or what counts as an answer (Section 2). A feasibility rating alone does not establish which standards for problems and solutions it presupposes, or whether those standards delimit the problems admitted by S. Problems concerning presuppositions had appeared on the human side at the stage of execution, as the removal of datasets that did not exist (Si, Hashimoto, and Yang 2026, 22). The authors of ResearchAgent themselves caution that its three stages are simulacra of the far more diverse and complex human research process (Baek et al. 2025, 6711).

On agreement in the ratings, there are reports such as the following. In Si and colleagues' first study, a measure of agreement among the reviewers who rated ideas was 56.1 percent, where chance was 50 percent; on the same measure, NeurIPS 2021 reviews are reported at 66.0 percent (Si, Yang, and Hashimoto 2025, 9). In the sequel, on the same measure, agreement in the evaluations after execution was between 59.0 and 84.3 percent depending on the item; the authors take this to be broadly comparable to conference reviewing, and read the especially high agreement on effectiveness as reflecting a more objective item that can be derived from experimental results (Si, Hashimoto, and Yang 2026, 20). Sinhahajari and colleagues, isolating the layer of questions, had research questions generated from fixed background papers and had language-model judges and experts compare their novelty against the questions pursued by the authors of published, highly cited papers. Those reference questions were reconstructed from the papers by a language model, and the authors caution that they are not the only valid questions (Sinhahajari, Majumder, and Poria 2026, 1, 5–7). Whereas the language-model judges favored the generated questions, two experts comparing fifty cases blind chose the authors' questions on non-obviousness, and agreement between human and language-model judgments was as low as 22 percent; the authors call this a novelty mirage (Sinhahajari, Majumder, and Poria 2026, 1, 13). Si and colleagues' two studies give reason to keep apart agreement in ratings of ideas alone and agreement in ratings that see the results of execution. Sinhahajari and colleagues' small comparison shows that human and machine judgments can diverge on the non-obviousness of questions. Both suggest that ratings of outputs on the side of questions cannot be read without asking on which items, and on what evidence, they were made.

Which questions the machine formulated, and how far it recast them, under topics or higher-level questions supplied by humans, under literature chosen by the authors, or under data and conditions of acceptance set in advance, is not evaluated in the same way across these studies. ResearchAgent gives evidence about a machine's generating the content of problems from the literature. DiscoPER's records give evidence at the layer of the organization of inquiry toward unsettled points (Section 1), in that the system directs its next work toward points not yet examined. The attribution of an act of asking or of an interrogative attitude, the choice of higher-level questions, and framing inquiry (Section 4) nevertheless each call for separate examination. From these studies, no blanket conclusion, affirmative or negative, can be drawn about the capacity of a machine to choose higher-level questions itself or to recast questions in the course of research. What comes into view is that what is taken up as unknown, and what is accepted as an answer, are shaped not only in the specification of questions but also in the course of execution and verification, and that whether that work is attributed to the machine or to something outside it, records that track changes are what the attribution rests on. Checking such capacities would require either a comparison that does not fix the higher-level question or a procedure that records and rates the work on the side of questions separately from the side of candidates. In a comparison that does not fix the higher-level question, however, as Si and colleagues gave as their reason for fixing topics, it becomes hard to separate whether what is being compared is research capacity or the misalignment of topic preferences. And even in designs that, like DiscoPER, supply no question, what it takes for a claim to be accepted and what the results are evaluated against are set elsewhere.

9. Viewing research from questions, and how this book views it

This chapter has seen how many kinds of work questions are involved in within research. Questions are raised, made tractable, recast, and closed, and they are asked by someone. It is natural, then, to think that research itself might be characterized as the activity of raising questions and clarifying them. There are, in fact, several accounts that understand research or the work of science from questions or problems. The first half of this section, in the same form in which Chapter 1, Section 3 described the check in the Liquid Tensor Experiment from Kelp's and Beddor's views, sets out what several such accounts and this book's working characterization attribute to the same activity of a machine, and what they ask about as not yet confirmed. It does not judge which account counts the activity as research. To give an outlook in advance: these accounts address different objects and units of analysis, which do not simply coincide with this book's A, the bearer of the achievement, or S. The working characterization can add to their descriptions the question of for whom the achievement is new. The discussion takes up Ravetz's later account of investigation and social valuation; Simon and Nersessian on problem formation and contributing systems; Funtowicz and Ravetz on purposes and quality assessment; Rheinberger on research processes producing something new; Laudan on theory appraisal and cognitive progress; Bromberger on contributions resolving items of ignorance; and Kelp and Lee on individual inquiry. These are different perspectives, rather than competing definitions of research. The second half looks back over what the chapter as a whole has seen.

As an example, consider the following activity. A system is given some higher-level goal, organizes its work under it toward one unsettled point q₁, and tests candidates bearing on that point by experiments on a computer. Part of the computation returns an unexpected result, and once this is read as bearing on a presupposition of q₁, the subsequent work is reorganized toward another point q₂. The transition from q₁ to q₂ here is not a recasting that fixes answer conditions not yet fixed (Section 3) but one in which, because a result bears on a presupposition, the point that organizes the work itself moves to another point. The system lists candidate answers c₁ and c₂ to q₂, tests c₁ by an experiment on a computer, and outputs the result as a report. Even at the time of the report, which of c₁ and c₂ to adopt remains open, and the result of testing c₁ also indicates room to revise what is to count as an answer to q₂. The applications below concern this hypothetical activity; Simon also expressly discusses artificial problem solvers.

In his 2016 interview, Ravetz describes science as investigating problems whose proposed solutions subsequently undergo social valuation and adaptation (Ravetz 2016, second answer). Applied here, this distinguishes the investigation from the social appraisal of its result. The social milieu need not coincide with S.

Simon and Nersessian provide process perspectives. Simon's robot revises its problem representation as new environmental information arrives (Simon 1973, 195); Nersessian connects model construction with problem articulation (Nersessian 2024, abstract). Applied here, these perspectives ask how the transition from q₁ to q₂ was accomplished, which components contributed, and which resources were inherited. Changing a problem representation does not by itself establish that the result is new. Nersessian's laboratory analysis also makes the history of devices relevant to identifying those resources (Nersessian 2009, 744–745). The question of who set priorities remains distinct (Section 1).

Funtowicz and Ravetz describe applied science as using routines for results with an external purpose, and distinguish producers', users', and managers' concerns in quality assessment (Funtowicz and Ravetz 1993, “Applied science,” Figure 2 discussion). Applied to the episode, this asks how the commissioned purpose bears on assessment, and whose quality concerns are relevant. For a machine A, the commissioning party, users, and reference society S need not coincide. Their account also includes routine work, consistent with the possibility that research proceeds by an established procedure (Chapter 5, Section 1).

Rheinberger holds that a research process has to actually produce something different from the present state of the art in order to remain a research process. The unit for which Rheinberger requires this, however, is not an individual result but the experimental system as a series of repeated reproductions, and what is produced acquires the character of novelty only once it has been sifted after the fact (post festum) (Rheinberger 1994, 68, 70). Applying this account, what is attributed to the activity above is the production of something different from the present state of the art, and what is asked is whether the series that includes the activity produced something that can be sifted after the fact as novelty. On the working characterization, since attempt is separated from success, an attempt that ended without producing anything new can still be called research in that sense (Chapter 1, Section 3). This difference comes from the fact that Rheinberger's account places the production of something new, and the working characterization the organization of the attempt, at the center of description.

Laudan's account, as Laudan and his colleagues summarize it in a survey of models of scientific change, takes science to aim at solving intellectual problems and appraises competing theories and research traditions by their effectiveness in solving them (Laudan et al. 1993, 71). Laudan's subject is the appraisal of theories and research traditions and the explanation of cognitive progress, and the unit to which having solved a problem is attributed is not an activity or a bearer but a theory; research traditions are in turn judged by the relative success of the theories they support (71). Applying this account, what is attributed in connection with the activity above is the empirical problems that the theory containing c₁ solved, the anomalies it generated, and the conceptual problems it generated, and what is appraised is that theory's problem-solving effectiveness: the number and importance of the empirical problems it solves minus the number and importance of the anomalies and conceptual problems it generates (71; Niiniluoto 2024, §3.2). If c₁ and c₂ remain open, what is compared is the effectiveness of the theories that contain each. An anomaly, as Niiniluoto summarizes the account, is a problem that a theory fails to solve but that some of its rivals solve (Niiniluoto 2024, §3.2), and on a thesis listed for Laudan, empirical difficulties by themselves never compel abandoning a theory (Laudan et al. 1993, 73). Applying this, the theory containing c₁ having solved a problem can change the problem situation of other theories in the same domain. Conceptual problems, further, concern contradictions within a theory, or between it and other scientific, methodological, or even metaphysical principles (71): vague or obscure explanatory categories count as serious deficiencies, and a theory can be challenged or supported by theories, methodologies, or metaphysical views outside its own field (73–74). In Dellsén's report, a conceptual problem is an incongruence within a theory, or between the theory and some other element of the research tradition, such as another theory or a methodological dictum (Dellsén 2018, §3). Applying this account to machines therefore requires specifying the subject: whose research tradition fixes what counts as a problem and as a solution, and against whose theories, methods, and views a conceptual problem is registered.

On whether questions generated by a machine carry weight, the effectiveness that Niiniluoto reports subtracts, besides conceptual problems, only anomalies, and on another thesis listed for Laudan, scientists accept a theory for its general record of problem solving rather than for its complete success with the problem currently under investigation (Niiniluoto 2024, §3.2; Laudan et al. 1993, 74–75). The account also separates the rationality of accepting a theory from that of pursuing it. It can be rational to pursue a theory that it would be irrational to accept: the decision to pursue rests on the rate at which the theory has recently solved problems, that to accept on its long-term record, and both require comparison with alternatives (71; Nickles 2017, §1.5); scientists often work with theories they do not accept, provided these currently show a high rate of problem solving (Laudan et al. 1993, 75). Applied to a newly raised empirical problem generated by a machine, the account implies that, if no rival theory has solved it, failure to solve it does not yet count as an anomaly against a theory. Further, a problem in Laudan's sense need not concern an actual state of affairs, since it suffices that the adherents of a tradition believe that it does, and Laudan's notion of solving an empirical problem resembles, and sometimes reduces to, potentially explaining a phenomenon (Dellsén 2018, §§3, 5); its unit differs from that of the questions of the semantics of questions seen in Section 2. This feature has drawn an objection that bears on machines. Bird, as Dellsén and Niiniluoto report, objects that the account counts as progress the accumulation of false solutions from an entirely false theory, as when Oresme believed that hot goat's blood could split diamonds (Dellsén 2018, §3; Niiniluoto 2024, §3.2). Neither Laudan nor Bird discusses artificial inquirers, but applied to a machine whose questions rest on false presuppositions that its own tradition accepts, the account could count apparent solutions to those questions as progress for as long as the tradition accepts the presuppositions, provided they raise the tradition's overall problem-solving effectiveness, whereas on Bromberger's distinction in Section 5 the ignorance they address would be spurious, and on Bird's objection nothing would thereby have been achieved. A presupposition that only the machine's tradition holds need not be false; what the contrast turns on is its falsity, not its being unshared.

Bromberger held that the progress of science is more than the addition of true propositions, and that the "more" lies in science's seeking answers to questions (Bromberger 1992, 1). And, as seen in Section 1, Bromberger took the view on which contributions to science are assessed as the resolution of items of ignorance (Bromberger 1992, 116). Under this account, what is attributed to the activity above is which items of ignorance were resolved. On Bromberger's account, questions and p-predicaments are items of ignorance that are resolved in different ways. The question q₂ is resolved when one of its direct answers has been established, while the predicament of being unable to think of candidate answers is resolved, relative to constraints, by a new answer compatible with the constraints (Bromberger 1992, 121, 123). Bromberger first characterized the p-predicament by elements that include a person, and then, for theoretical convenience, removed the person and defined it as a relation among four things: constraints, a question, the set of answers at hand, and a search procedure (Bromberger 1992, 123). If this definition is applied, the machine's interrogative attitude is not asked about. What is asked is whether the presupposition of q₂ is true, so that the ignorance is not spurious ignorance (Section 5), and, if S is a human society, whether the item of ignorance resolved is expressible and sharable in human language (Section 7).

Kelp gives an account of individual inquiry. As seen in Chapter 1, Section 3, on Kelp's account inquiry into a specific question aims at settling that question for oneself, and inquiry into a general phenomenon aims at understanding (Kelp 2021b, S1583; Kelp 2021c, 13). Kelp takes this to be a light characterization that can be widely accepted, places the knowledge, true-belief, and justified-belief accounts as substantive accounts of what settling requires, and endorses the account that requires knowledge (Kelp 2021b, S1583–S1585; Kelp 2021c, 13). What Kelp examines as a norm that must be followed for something to be inquiry at all (a constitutive norm) is forming beliefs through epistemic competences, and forming a belief through brainwashing is not inquiry (Kelp 2021c, 13). Applying Kelp's account to this activity, what is attributed is settlement for the inquirer itself. If which of c₁ and c₂ to adopt is open at the time of the report, inquiry into q₂ in this sense has not closed, even though the run has ended. What still remains to be asked is whether the machine forms beliefs through epistemic competences, and whether its answer is not one that happened to be right from misleading cues. On the latter, in Kelp's example, a private detective hired to find out whether someone in the client's family committed a murder, who reaches a justified true belief from a confession that turns out to be false, is not released from the contractual commitment, and the employer can require a return to work without a new contract (Kelp 2021c, 3–4). It is accounts on which settling requires knowledge that describe the inquiry as not yet over in such cases.

Kelp's inquiry is broader than research in the sense of this book's working characterization, in that it includes cases in which an individual comes to know what S already knows. Further, since on Kelp's account the unit at which inquiry closes is the individual inquirer's belief or knowledge, it does not directly treat cases in which organized activity is the bearer. Moreover, Kelp analyzes knowledge, belief, and inquiry within a non-reductive network (Kelp 2021c, 10–15) and defines belief as the state in which inquiry has closed (Kelp 2021c, 14). Applying this analysis to machines therefore does not allow the question of attributing inquiry to a machine to be handled apart from the question of attributing beliefs or epistemic competences; the mutual relations among these states and activities are to be examined together. On this point, it differs from Watson's account in Section 6 in what the theory explains together. Watson's account limits its object to the act of asking and considers its attribution from the function of eliciting information, so that whether a machine's operation has that function could be asked separately from whether the machine has beliefs. Further, while holding that inquiry into specific questions and inquiry into a general phenomenon share a common core, both aiming at knowledge of the associated true propositions, and that progress in inquiry into a phenomenon consists in coming to know more true propositions about it (Kelp 2021b, S1589), Kelp discusses cases in which knowing the answers to the individual questions does not reach understanding if the inquirer's beliefs are not based on one another in a way that reflects knowledge of the explanatory and support relations among the truths known (Kelp 2015, 16). Habgood-Coote's group inquiry in Section 6 also requires choosing appropriate subquestions and combining their answers, so the difference between Kelp's understanding and Habgood-Coote's group knowledge does not lie in whether questions are divided. Even if a group comes to know the answer under those conditions, understanding of a general phenomenon in Kelp's sense further requires that the inquirer's beliefs be based on one another in a way that reflects knowledge of the explanatory and support relations among the truths known. The claim that a machine gained knowledge by combining subanswers and the claim that it understood are therefore to be checked separately.

Lee's account is formulated for an individual subject's state of doubt. On Lee's account, an interrogative attitude is an attitude that takes a question as its object, is directed at its resolution, and seeks, or is at least sensitive to, relevant information (W. Lee 2026, 3). Lee then divides doubt into two. One is having some doubt: thinking that both a proposition and its negation are possible, and so having a disposition not to affirm it unconditionally. The other is being in doubt: having, in addition, a disposition to feel sufficiently strong doubt on that account, and a disposition to treat neither answer as a default in reasoning, that is, as an answer one uses for the time being as correct when reasoning (W. Lee 2026, 7, 11–12). And Lee takes having some doubt to be a necessary but not sufficient condition of inquiry (W. Lee 2026, 14–15). Applying this account, counting the activity above as genuine inquiry requires, in addition to an interrogative attitude, at least a disposition to withhold affirmation because one regards both the proposition and its negation as possible. If this disposition can be attributed, a machine that lacks the disposition to feel can fall under Lee's "having some doubt," but not under "being in doubt," since Lee formulates the disposition not to treat an answer as a default as supported by felt doubt. A machine state that, besides having some doubt, withholds defaults without feeling therefore remains something to which Lee's twofold division has not yet given a name of its own. This application, however, depends on the premise, itself disputed, that machines lack a disposition to feel, and the part about thinking something possible also requires the attribution of thought. What is asked when the accounts of individual inquiry of Jane Friedman and Daniel C. Friedman are applied was seen in Sections 6 and 7.

The working characterization asks, of the same activity, whether it is organized as an attempt to bring about an epistemic achievement new to S. That q₁ stood under a goal supplied by humans, who carried out the recasting to q₂, and whether c₁ withstood testing are not conditions of being research in the sense of the working characterization. They remain as separate questions: to whom what is to be attributed (Chapter 7), in what respects and how far it is autonomous (Chapter 7, Section 4.3), and whether the result entered S's epistemic situation (Chapter 2, Section 9.4).

Set side by side, these perspectives attend to different objects and units. Ravetz's later account distinguishes investigation from social valuation; Simon and Nersessian examine processes and their contributing systems; Funtowicz and Ravetz distinguish purposes and participants in quality assessment. These are not all units of the same kind. For Laudan, the unit is the theory that solves the problem (what counts as a problem and as a solution, and against what a conceptual problem is registered, are fixed by the research tradition whose adherents appraise the theory); for Bromberger, the item of ignorance of science (from the point of view seen in Section 7, if S is a human society, it is required to be expressible and sharable in human language); and for Kelp and Lee, the individual inquirer. On Kelp's account, settlement is for the inquirer itself, that is, on the side of A. These units do not simply coincide with any of the working characterization's A, bearer of the achievement, or S. Even if a question is settled for the machine itself in Kelp's sense, whether the achievement is new to S is asked separately, and conversely, an attempt to bring about an achievement new to S may not be accompanied by the machine's own knowledge (Section 7). The adherents of Laudan's research traditions, the social milieu in Ravetz's later account, and the participants in Funtowicz and Ravetz's quality assessment can be identified with S only when S is taken to comprise the relevant participants. When these perspectives are applied to machines, therefore, how their objects and units of analysis relate to A, to the bearer of the achievement, and to S is to be fixed separately. Since in the working characterization S is placed as what research is directed toward, the question of for whom an achievement is new can be added to, and asked of, each account's description.

Then, one might ask, would it not suffice to characterize research itself by the view, seen at the end of Section 5, that looks at the research process from the side of open matters? To do so, however, would need a further argument distinguishing research from other activities carried out amid uncertainty. As Chapter 5, Section 5 sets out, Dewey's own view was that reflective thinking, however diverse its cases, shows common elements. Describing a pattern common to inquiry does not by itself separate research from other inquiry.

Looking back over the chapter as a whole, the single claim that "a machine asked" divides into claims about several separate kinds of work, each to be checked separately. Section 1 divided the claim into having output an interrogative sentence, having performed an act of eliciting information, having had an interrogative attitude, and the inquiry's having been organized toward an unsettled point, and held that whether a machine's output was taken up as a question pursued by members of S is also asked separately. Sections 2 and 3 saw that what counts as an answer differs across accounts, and that who carried out the work of making a question researchable, that is, of deciding what to take as the foil, what kind of explanation to count as an answer, and within what range of candidates to look for one, is asked separately from having put the question. Applying Lipton's analysis together with the application of Simon and Nickles in Section 3, where humans supplied the framing that a machine used without revision, successful execution does not establish that the machine performed that work. Sections 4 and 5 saw questions being formed and recast within inquiry, ceasing to stand without an answer, being suspended, and being closed. On Henne's distinction, multiplying questions under the same rules differs from putting the framework in question, and however complete the record of framed inquiry may be, questions that could not be raised within that framework do not remain in it. Taking S as a scientific community and applying Kuhn's and Laudan's accounts adds the question of whether incorporating machines changes the problems S admits or its standards for adequate solutions. On whichever of the accounts of Collingwood, Jane Friedman, and Daniel C. Friedman is applied, a question's ceasing to stand, and its being settled, are asked separately from the end of a run. Sections 6 and 7 saw that who asks divides into the attribution of the act of asking, of group inquiry, and of individual inquiry, that on many of these accounts a state of curiosity is not the key to attribution, and that what is unknown to A need not coincide with what is unknown to S. The state of curiosity itself returns as a question on accounts that make interrogative attitudes a condition of inquiry, where an artificial system's intellectual life valued for its own sake is aimed at, and where how well research goes depends on states of the one doing it that correspond to motivation. Section 8 read reports that machines produced questions or research ideas, and saw that much of the work on the side of questions appears on the human side, in task descriptions or at the stage of execution.

Taken together, what it means for a machine to ask is asked not as whether a machine has some one capacity or state, but as which of the kinds of work bearing on questions are carried by which unit, directed toward a point unsettled for whom. And many of these questions also appear as gaps between the description on the machine's side and the description on the side of S.

These consequences concern, among the three relations of inquiry seen in the introduction to Part III, the relation between an inquiry and the unsettled point that organizes it. This relation, however, did not hold apart from the other two. Fixing a foil narrows what an answer must do and where to look. On the editors' reading, the basic questions of Darwin's notebooks kept their headlines while changing their content as candidate answers took form. Judgments about where to locate a failure shape the next path, and when standards are questioned, what counts as an answer can also change. The question of what it means for a machine to ask therefore remains connected, within the same inquiry, with a machine's making candidates and representations available and with determining what results bear on.