On 15 October 2026, I spoke with Bea and Orla, who have backgrounds in history and are involved with the Oxford Review of Books, about AI, creativity, and research. Bea told me that she had never used AI for writing. In explaining her position, she attached importance to human effort. I asked why that effort mattered so much. She said that she needed to think about it further. I do not want to turn that unfinished exchange into a settled philosophy on her behalf. What stayed with me was the question it opened.
That same day, I had been reading a Chinese article presenting recursive self-improvement as an approaching historical rupture. Its urgency rested partly on Jack Clark’s revised expectation that full recursive self-improvement might become possible in 2027 rather than 2028. Clark’s own newsletter does contain that revision, but an expectation about a possible capability is different from an established deadline for humanity (Clark, 2026). Alongside reports of striking AI achievements in mathematics, the forecast made the distance between technological ambition and ordinary conversations about acceptable assistance feel unusually large.
My initial reaction was to ask whether our debate had already fallen behind the capabilities being discussed. My own starting image was “standing on the shoulders of giants”: research already depends on capacities and knowledge that no individual has produced alone. I saw AI as a possible extension of that inheritance, an amplifier and collaborator. But the analogy did not settle Bea’s concern. It explained why receiving assistance need not be discreditable; it said less about what happens to the person who receives it. I continued the discussion with AI because I wanted to understand that gap. I did not begin with a theory of authorship or an assessment framework. Those questions emerged as I found that each apparently useful distinction left something unresolved.
What are we defending when we defend effort?
When I asked what we were talking about when we valued human effort, the AI suggested that Bea might be concerned with the way sustained engagement changes the person doing it. I explicitly said that I valued this too. Reading something difficult, struggling to express an interpretation, and discovering that an attractive explanation fails can develop judgment. The AI then distinguished practical work, intellectual formation, and personal expression. Formatting references, learning to recognise a weak argument, and choosing words for something personally significant all involve effort, but reducing each kind need not have the same consequences. The distinction helped me ask what a particular activity makes possible. Yet I pushed back against a possible implication: why should intellectual formation be placed on the side of unaided work? I raised coding because working with AI still requires me to guide the system and judge whether its response addresses the problem. A historian might similarly challenge a proposed connection and return to the sources because of the exchange. My objection did not establish that every AI interaction is educationally valuable. It challenged the assumption that distinguishing clerical labour from formative effort already tells us where AI belongs.
Another AI formulation caught my attention: “effort alone does not establish intellectual or moral value.” I said that I found this interesting. The formulation complicated my own defence as well as the objection to AI. If I defended collaboration only by insisting that I still worked very hard, I would preserve the assumption that the amount of effort determines the legitimacy of the result. A laborious interpretation can be mistaken; a quickly recognised connection can be valuable. The more useful question concerns which activities develop capacities worth possessing, and whether assistance supports or displaces them. Difficulty can be productive without becoming a virtue in itself, while a reduction in difficulty can be either helpful or educationally costly.
The AI also offered a position it would reject: that a work loses value whenever technology reduces the labour involved. I asked whether this really described Bea’s view. That question mattered. The AI had framed the position conditionally, but a hypothetical position can still become a convenient opponent if the conversation proceeds as though someone actually holds it. Bea had said that she did not use AI for writing and that effort mattered; she had also left her explanation unfinished. I wanted to understand her concern without supplying a more easily refuted argument on her behalf. The possibility that she values remaining personally involved in understanding and creating survived this objection. Whether AI can support that involvement remained open.
What historical research might become
My example to Bea had been about an expanded possibility. What happens when a historian can investigate connections across a thousand books or ten thousand documents that would have been impractical to compare closely within one research project? I returned to that example in the AI conversation and asked for cases from biblical research or other fields that would make the possibility concrete. “AI can help historians” was too general to resolve the issue. I wanted to know whether the assistance could change the research that became possible, rather than merely shorten familiar tasks. Yet “impossible without AI” also needed care: teams, databases, statistics, and earlier computational tools can extend historical inquiry. The relevant question concerns the particular limitation a method overcomes.
A concrete biblical example helped. Popović, Dhali, and Schomaker used computational handwriting analysis to investigate the Great Isaiah Scroll, finding evidence supporting the involvement of two principal scribes despite the manuscript’s closely similar handwriting. Their study connected measurable differences in writing patterns with a question about ancient scribal practice (Popović, Dhali, and Schomaker, 2021). It did not determine who composed the Book of Isaiah. Keeping copying, composition, and interpretation separate is essential to understanding what the result establishes. Its significance lies partly in making a difficult palaeographical judgment available to quantitative investigation.
The Aeneas project offered another example. Its researchers assembled a corpus of 176,861 Latin inscriptions and developed a system for retrieving contextual parallels and assisting restoration and attribution. In their evaluation, historians working with Aeneas achieved better results on restoration and geographical attribution than either historians or the system alone (Assael et al., 2025). This is evidence about particular tasks under specified conditions, rather than a verdict on the whole profession. It nevertheless complicates the picture of AI as something that merely saves a historian from clerical work.
These examples gave my initial intuition a more precise basis. The issue was no longer simply how many documents one person could read. It concerned whether a new method could produce comparisons that changed the evidence available for a judgment. That was a more specific claim than suggesting that AI could generally “do history.” The handwriting study and Aeneas used specialised methods, datasets, and evaluations; neither licenses treating an ordinary chatbot as an equivalent research instrument. Their relevance is methodological. A retrieved parallel can alter an interpretation, and a classification can change the question being asked. The boundary between preparing information and doing intellectual work therefore requires examination.
The conversation developed a hypothetical example involving sermons on the biblical command to love one’s neighbour. A researcher might ask whether changing uses of “neighbour” indicate changing attitudes towards outsiders. AI could help locate passages and propose patterns across a large collection, while the researcher checks whether an apparent chronological change instead reflects sermon genres, audiences, or the documents selected for digitisation. That appeared to offer a reassuring division of labour. But I kept pressing: if creativity includes discovering connections, formulating questions, and proposing explanations, why assume that machines can contribute only to the preliminary stages? I asked whether their capacity to process material might allow them to do some of these things better. The scale of their input would not itself prove interpretive reliability, but the question prevented “human judgment” from becoming an unexplained reservation.
A system might itself suggest the genre problem, recognise a sampling error, or propose the better research question. The AI’s response that we should acknowledge such achievements when they are reliable seemed right to me, yet I immediately raised a further difficulty: people may still refuse to recognise machine creativity, especially where human guidance was involved. In that case, credit becomes harder to allocate rather than easier to dismiss. This was a change in the inquiry. I had begun by asking what AI could enable a historian to do. I was now asking how we recognise intellectual contributions when the activity already crosses the human–machine boundary. Historical creativity still has to answer to evidence; inventing an explanation and inventing the evidence for it remain very different acts.
The reports of AI mathematics made this problem harder to evade. I had brought a report of a claimed proof and a dispute over its relationship to an existing mathematical approach into the conversation, asking whether a new method would count as creation. I was not in a position to settle the mathematical claim myself. The conditional point was sufficient: if a system produces a genuinely new, valid, and significant proof, human guidance does not by itself erase the system’s contribution. Nor does the machine’s contribution erase the human work. That led back to my question about the historian’s distinctive role and dignity. Finding a task that machines cannot perform would be a precarious foundation for human worth; acknowledging that does not resolve the practical future of a profession.
When assistance becomes an idea
I then tested a more familiar boundary: AI may help with translation and grammar, but it should not help with writing. Was that distinction too simple? It seemed to classify the visible output more easily than the intellectual activity. A translation choice can affect how a theological concept is understood; a grammatical revision can expose an ambiguity in an argument. Conversely, a paragraph generated from a carefully specified argument might contribute less to its conception than one question asked during brainstorming. This returned me to Bea’s stated practice. Knowing that she had never used AI for writing did not yet tell me how she would understand a conversation that clarified a question or generated an objection.
The AI proposed asking whether Bea would still object if collaboration made me read more carefully, encounter more counterevidence, and understand the subject better. I found that formulation interesting, but I did not know her answer. It also raised a problem within my own position. The AI reminded me that “understanding and endorsing an idea does not mean that you originated it.” I responded with a more particular case: what if I raised the question and the idea took shape through our conversation? I was not claiming that understanding someone else’s finished idea makes me its inventor. I was asking about an idea that becomes identifiable through the exchange. My question might make a particular answer possible; an AI distinction might enable me to construct a counterexample; that counterexample might then require an explanation neither of us had previously supplied. Human agency here includes changing what the inquiry is about, rather than simply approving an output. “Who originated it?” remains a legitimate question, although we may first need to specify which idea, at which stage of its development, we mean. The contribution can be substantial even when it cannot be separated cleanly from the interaction that made it possible.
There is also a difference between producing research and demonstrating a particular capacity in an assessment. A researcher may appropriately organise a project around extensive collaboration. An assessment may need to establish whether a student can independently perform some part of the work. Neither purpose makes the other illegitimate. What troubled me was the possibility of a university invoking independence without specifying what must be independent, or invoking originality as though intellectual work normally emerged without encounters, conversations, reading, or inherited methods.
The institution also has something to explain
This was where the discussion moved towards academic integrity. Taylor and LaCroix argue that universities need to relate their AI policies to their purposes and examine whether their institutional practices are coherent with their declared values (Taylor and LaCroix, 2026). Their argument gave scholarly form to a concern I was already expressing: students should not carry the entire burden of making sense of rules whose educational rationale remains obscure. Their paper is a conceptual and rhetorical analysis, however, rather than an empirical demonstration that every university’s policy is incoherent.
I would also distinguish a genuine contradiction from a justified difference between activities. A university can encourage AI use in one setting while restricting it in another, provided that the different purposes explain the different permissions. An exercise in independent translation and a project comparing thousands of historical documents might reasonably require different arrangements. Institutional accountability consists partly in making that relationship understandable. Simply offering AI access does not entail permission to use it everywhere; simply naming academic integrity does not explain a restriction.
A factual correction in our own exchange mattered here. I had raised Oxford mathematics as a possible example of restrictions on AI brainstorming. The departmental policy examined in the discussion actually lists generating research questions, extensions, or examples among generally permissible uses, alongside requirements for declaration and checking (Mathematical Institute, 2025). The example therefore could not support a claim that Oxford mathematics simply prohibits idea generation. Preserving that correction matters more than preserving a convenient institutional target for my argument.
Oxford’s central assessment policy also explicitly links permitted assistance to the purpose of an assessment and requires clear explanations of permissions and declarations (University of Oxford, 2025). When the AI pointed to that requirement, I doubted how clear the practical position actually was. Citing a principle of clarity cannot demonstrate its successful implementation. A policy may leave a student uncertain whether a conversation counts as clarification, brainstorming, or substantive intellectual assistance. I was therefore asking more than whether universities had published rules. I wanted to know whether the categories in those rules adequately described the activity they were supposed to govern.
The AI qualified one answer by saying that its distinctions concerned what a sound policy should make, rather than a ruling on a particular disciplinary case. I pressed the objection again: a policy that creates its own ambiguity may harm both students and the university. The qualification about disciplinary rulings was reasonable, but it did not finish the normative argument I was pursuing. I wanted to examine whether the rule itself was adequately designed, before considering what happened when somebody was accused of breaking it. Ordinary conversations, indirect suggestions, and uncertain intellectual origins seemed to be cases a policy should anticipate, rather than leave entirely for students to resolve.
The demand for institutional clarity also has a procedural dimension. The Office of the Independent Adjudicator’s casework guidance calls for clear explanations of acceptable AI use, fair opportunities for students to respond to allegations, and evidence sufficient for the institution to substantiate its case. It cautions decision-makers about the limitations of detection tools (OIA, 2025). Those requirements do not resolve every philosophical disagreement, but they prevent uncertainty from becoming an unrestricted burden placed on the student. A student should be able to understand the allegation and the evidence supporting it.
The conversation behind a screen
The example that led towards my experiment concerned indirect influence. Suppose a student is forbidden from generating ideas with AI, but a friend offers a useful suggestion previously obtained from AI. What exactly has the student done? I raised this because permission to talk with another person seemed difficult to reconcile with a prohibition defined entirely by the ultimate origin of the suggestion. The immediate conversation is human, while part of its intellectual history involves a machine. Following that history further makes the problem more complicated: the AI may reproduce a scholar’s argument, and the scholar may have developed it through earlier conversations. The immediate speaker cannot, on their own, supply the whole genealogy.
I turned that example into a proposed interface. One connection would lead to a friend’s WhatsApp, another to an AI API, and a routing mechanism would choose between them without showing me the selection. Variable delays would obscure one clue to the source of each reply. I would receive the conversation through the same interface without knowing which connection was active. This was a thought experiment, not a reported implementation. The timing mechanism was also only a proposed way of concealing a clue; it would not establish that the content was indistinguishable. In the same message, I asked another question: I had just formulated this experiment myself, although it emerged during an AI conversation—was that still my original contribution?
My first formulation mixed several questions. Could I describe an answer as having a fifty per cent chance of coming from my friend? Would that uncertainty matter to whether I was permitted to use it? And what did the case reveal about a policy that classified assistance by source? The AI replied that a probability about origin is not a proportion of authorship. A reply does not become half human-authored because I am uncertain who supplied it, and a fair random selector does not guarantee an exact split across a finite conversation. Those corrections were useful. They did not, however, settle the objection about why a source should determine permissibility.
If the router makes an independent, equally weighted choice for each reply, the probability of receiving at least one AI reply across \(n\) exchanges is:
\[
P(\text{at least one AI reply}) = 1-\left(\frac{1}{2}\right)^n.
\]
Across ten exchanges, that is approximately 99.9 per cent. The calculation depends on repeated independent routing; it would differ if one selection determined the interlocutor for the whole conversation. It also addresses only one possible interpretation of my question. My proposed channel explicitly included AI, so demonstrating that it would probably produce some AI replies could not by itself answer the philosophical challenge. Exposure to AI, permission to receive assistance, and intellectual contribution are different variables. A single suggestion might contain the decisive insight while numerous other replies contribute little. The number clarified a feature of the mechanism; it did not measure originality or justify a rule.
A further distinction in the AI’s response did more of the conceptual work. Not knowing the source of a particular reply differs from knowing that I chose a mixed channel. Uncertainty about the first does not erase knowledge of the second. That challenged the suggestion that randomness could itself make otherwise prohibited assistance permissible. Yet it left room for the criticism that had motivated the design. A university might regulate the decision to use the channel, the incorporation of particular material, or disclosure of assistance. It would still need to explain why that object of regulation was appropriate. My proposed mechanism had made the question more concrete, but it had not yet isolated the question cleanly.
The same response distinguished ambiguity in a rule, uncertainty about the facts, and disagreement with the rule’s justification. That distinction became useful here because my experiment could be interpreted in all three ways. A university might state a clear restriction while the source of a particular reply remained unknown; the restriction might also be clear and still fail to serve its educational purpose. My criticism could survive this clarification, but it needed to specify the failure. The uncertain reply did not by itself prove that the wording was ambiguous, and a clear prohibition did not by itself supply a satisfactory reason for having it.
The arrangement also has a recognisable predecessor in Turing’s imitation game, where communication through an interface conceals characteristics of the interlocutor (Turing, 1950). The comparison, raised by the AI, helped locate the experiment without pretending that I had invented concealed human–machine conversation. My emphasis was different. I was interested in what a university could justify concerning learning, contribution, and responsibility when the recipient could not identify the source. The historical resemblance limits a claim to novelty while helping clarify the question I was asking.
Making the experiment answer a sharper question
I then asked the AI to design a better version of the thought experiment. That request changed its task. I was asking it to help preserve the worthwhile challenge while removing weaknesses in the first formulation. Its response proposed a different starting point: a university is deciding what its policy ought to permit, before there is any disciplinary case. The institution must explain the educational distinction it wants to make. This revision matters because an answer about knowingly choosing a mixed channel no longer disposes of the central issue. The inquiry now asks which difference should justify different treatment, rather than treating an existing prohibition as sufficient explanation of itself.
The revised case returned to historical research. A student independently asks whether changing interpretations of “neighbour” in sermons indicate changing attitudes towards outsiders. During discussion, they receive the suggestion that the apparent change may reflect differences in sermon genre. They investigate, classify the material more carefully, examine counterexamples, and revise the argument. Imagine two alternative versions of this history: a friend formulates the suggestion in one, and AI formulates the identical suggestion in the other. The student’s subsequent work and independently demonstrated understanding are held constant. An evaluator judges the argument and the student’s contribution before learning the suggestion’s source, then explains which judgments should change after disclosure. The historical evidence remains the same, and the student performed the same investigative work. Acknowledgment changes because the supplier differs. Someone might also argue that an educational relationship or a requirement of independence makes a relevant difference; the experiment requires that reason to be specified. It does not show that receiving a suggestion is equivalent to originating it. The student did not originate that suggestion in either version. The question concerns the status of the work the student subsequently performs, and whether the source changes that answer.
The AI also suggested distinguishing the origin of the particular suggestion from the channel through which it arrives. Four cases make that distinction visible:
| Case | Who formulates the suggestion? | Who communicates it to the student? |
|---|---|---|
| A | A friend | The friend |
| B | AI | The AI system |
| C | AI | A friend forwarding it unchanged |
| D | A friend | An AI system relaying it unchanged |
A rule about using an AI interface groups B with D. A rule about receiving an AI-formulated suggestion groups B with C. A rule about receiving substantive external assistance may group all four together. A rule about independently demonstrated understanding has to examine what the student subsequently does. For this comparison, we stipulate who formulated the particular suggestion; we are not claiming that either people or machines create without influences. The cases show why a policy must identify what it regulates before it can explain why.
My random channel can then return as a separate test of responsibility. One student reasonably believes the interface connects only to a friend; another knows it is mixed and has permission to use it; a third knowingly uses it for a task that requires the relevant work to be unaided. The resulting argument could be identical, but the students’ information and obligations differ. Alternatively, two students could make identical informed choices and receive different sources purely through random routing. Whether that difference should affect their treatment depends on what the rule is intended to assess or protect.
A final variation delays disclosure until the work is finished. The student understands the argument, then discovers that a pivotal suggestion came from AI. Their knowledge of the process changes, and acknowledgment may need correcting. Their feelings about the achievement might change too. Yet the disclosure cannot retrospectively remove the reading, criticism, or understanding that actually occurred. That observation does not prove that provenance is irrelevant. It distinguishes the history of a contribution from the capacities a person has developed through working with it.
There is an important limit to the refinement. By holding understanding constant, it isolates a possible objection to the source itself. It cannot demonstrate that different sources actually produce equivalent learning. If someone’s concern is that ordinary AI use encourages dependence or weakens judgment, the thought experiment does not refute that empirical claim; investigation of actual practices would be needed. Human and AI assistance may also differ in availability, reliability, and the relationships they sustain. The revised case therefore offers a way to distinguish objections that are often combined. It provides no general permission, and it should not be presented as experimental evidence for unrestricted AI use.
Where provenance still matters
The AI’s response also required a qualification to my suggestion that judging a resource by its origin was a bad approach. Historical research made the problem particularly clear. Suppose AI proposes an interpretation of the editing of Isaiah, and a researcher tests it against manuscripts and scholarship. The proposal’s origin alone does not determine whether it is defensible. But suppose the supposed evidence is a quotation presented as coming from an ancient manuscript. Whether that quotation is authentic or invented is central to what it can establish. I could retain the criticism of treating machine origin as an automatic intellectual disqualification without making provenance irrelevant to historical knowledge.
There is a parallel distinction in authorship. A sound argument does not become false because someone else first proposed it, but presenting another contributor’s proposal as my independent invention would still misdescribe the work. Truth, originality, learning, and compliance are connected without being interchangeable. The difficulty is therefore not solved by saying that only results matter. It calls for a more informative account of the result and its production: what was suggested, what was checked, what was changed, and what the researcher can explain.
I would consequently favour policies that identify the capacities students must demonstrate independently, describe permitted assistance through realistic examples, and require proportionate acknowledgment of significant help. Where independent ability matters, the assessment needs appropriate evidence of that ability. Discussion, supervised components, or selected process records may contribute, without becoming universal solutions. Requiring an exhaustive transcript of every influence would impose a different problem. A workable account of responsibility has to concern assistance people knowingly seek and contributions they can reasonably recognise.
An idea of my own inside a conversation with AI
The question about my own contribution had arrived before the improved experiment. I had asked it in the very message that introduced the random-channel proposal. That timing matters: I was already using the exchange to examine the conditions under which an idea could be called mine. The AI had not supplied the specific proposal before I raised it, although its earlier responses formed part of the conversation in which I developed it. The record supports saying that I formulated this particular challenge within our exchange. Establishing that no one had ever proposed anything similar would require a different investigation, and the later refinements should not be folded retrospectively into my original contribution.
Boden’s distinction between creativity relative to an individual’s previous thinking and creativity relative to the history of a field helps keep those claims apart (Boden, 1998). Discovering an idea through my own reasoning can matter without establishing historical priority. Conversely, merely encountering an unfamiliar idea would not establish that I created it. In this case, the contribution concerns the mechanism I proposed and the problem I used it to investigate. Its relationship to earlier thought experiments remains something to examine, rather than something a fluent AI assurance could settle.
The sequence makes a more informative account possible than assigning an overall percentage of authorship. Bea’s concern prompted a question about effort. I brought that question to AI, accepted the relevance of intellectual formation, and challenged the implication that collaboration would diminish it. The discussion of creativity led me to ask about originality in a conversation; the policy discussion led me to the friend example and then the random channel. AI supplied objections and refinements, while my request for a better experiment changed what it was being asked to produce. These contributions have different functions. A criticism that redirects the inquiry may matter more than a large amount of prose, and the significance of either can depend on what happens next.
The writing process has now repeated the problem on another level. I first asked AI to turn the conversation into an article, and then asked for a second pass that would preserve more of the history of thinking. A coherent first-person narrative can make the inquiry easier to inspect, but coherence can also conceal changes of direction. It can make an AI distinction sound as though I had always held it, turn a question I was testing into a belief I had confidently adopted, or represent a later refinement as part of my original design. The revision therefore has to preserve interruptions as well as conclusions. My objection to a plausible answer may be the event that makes the next stage possible.
Even this account has limits. The conversation records what I wrote and what AI returned; it does not provide complete access to every influence on either side. Reconstructing the exchange can identify a proposed mechanism, a correction, and a changed question without revealing the entire genesis of an idea. The phrase “thinking alone” in my title therefore needs care. I can work without a live collaborator while still thinking with inherited concepts, texts, and remembered conversations. That does not make all forms of assistance equivalent. It makes the relevant differences something to explain, rather than a distinction that follows automatically from whether another voice is visible on the screen.
I would like to return to Bea with a more precise question. If working with AI led me to read more carefully, encounter stronger counterevidence, and understand a subject better, would her concern remain? It might. She could value forms of making and human relationship that this account of learning does not exhaust. I cannot answer for her. Equally, I would want room to explain that accepting help need not mean abandoning participation. Our original disagreement remains open, but there are now several different things we could be disagreeing about.
What I can ask of a university is more precise than a general endorsement of AI: an explanation of what a restriction protects, why the relevant kinds of assistance differ, and how students can understand their obligations before being judged. What I can ask of my own work is an account of participation that remains open to correction. I began by defending the possibility that AI could extend what a researcher was able to do. I now have a more difficult question about what happens within that extension—how effort forms understanding, how contributions become identifiable, and how responsibility should be assigned when those processes interact. My random-channel experiment began by concealing a source. Its development has made the history of the thinking harder to leave out.
References
Assael, Y., Sommerschield, T., Cooley, A., et al. (2025). “Contextualizing ancient texts with generative neural networks.” Nature, 645, 141–147. https://doi.org/10.1038/s41586-025-09292-5.
Boden, M. A. (1998). “Computer Models of Creativity.” In R. J. Sternberg (ed.), Handbook of Creativity, pp. 351–372. Cambridge University Press. https://doi.org/10.1017/CBO9780511807916.020.
Clark, J. (2026). “Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space.” Import AI, 12 October. Original newsletter.
Mathematical Institute, University of Oxford. (2025). Departmental policy on use of AI, October 2025 version. Policy PDF. Consulted 16 October 2026.
Office of the Independent Adjudicator for Higher Education (OIA). (2025). Casework note: Complaints relating to AI and academic misconduct. Published with accompanying case summaries on 15 July. Casework guidance. Consulted 16 October 2026.
Popović, M., Dhali, M. A., and Schomaker, L. (2021). “Artificial intelligence based writer identification generates new evidence for the unknown scribes of the Dead Sea Scrolls exemplified by the Great Isaiah Scroll (1QIsaa).” PLOS ONE, 16(4), e0249769. https://doi.org/10.1371/journal.pone.0249769.
Taylor, T. B., and LaCroix, T. (2026). “Purpose before policy: academic integrity, generative AI, and rhetorical stance.” Higher Education. Published online 11 June. https://doi.org/10.1007/s10734-026-01706-1.
Turing, A. M. (1950). “Computing Machinery and Intelligence.” Mind, 59(236), 433–460. https://doi.org/10.1093/mind/LIX.236.433.
University of Oxford. (2025). AI use in summative assessment. Page last updated 27 October 2025. University assessment policy. Consulted 16 October 2026.