Whose Thinking Is Being Assessed? The Unsettled Boundary Between Assistance and Authorship in AI-Assisted Education

Abstract

This article examines the unsettled boundary between intellectual assistance and the requirement that assessed work express a learner’s own thinking. It reconstructs an iterative human–AI inquiry and develops it through close analysis of original research on language models, writing, feedback, authorship, learning and assessment validity. A graduate admissions policy provides one institutional example, rather than a verdict on higher education generally. The argument distinguishes the origin of an idea, the learner’s understanding, intellectual contribution, expression and disclosure. Direct reading sharpens the analysis: tutoring outcomes depend on instructional arrangements as well as models; observable interaction can be an imperfect indicator of learning; and valid interpretations of performance do not automatically justify assessment decisions. Research on epistemic dependence and Value Sensitive Design further complicates ideals of independence and the transfer of frameworks between disciplines. The article develops a question for theological pedagogy, especially Science and Religion, about how assistance supports or substitutes for the capacities being assessed. Its contribution is a more precise account of the evidence and judgments required, while leaving unresolved when comparable human and AI assistance should receive different treatment.

Keywords: generative AI; theological pedagogy; intellectual authorship; academic assessment; human–AI interaction; policy ambiguity; Science and Religion

I kept returning to a comparison that the answers I received did not quite address. A friend suggests a distinction, asks a difficult question, or points out a weakness in my argument. I follow the suggestion, examine the sources and develop my thinking. An AI system offers the same kind of assistance, and I undertake the same intellectual work. What, exactly, would justify accepting the first process while rejecting the second?

The question arose while I was exploring possible research in AI and theological pedagogy. My previous master’s work at KU Leuven had examined AI-use disclosure, authorship and assessment through human-computer interaction and Value Sensitive Design. I wanted to continue that inquiry while giving the technical workings of AI a more substantial role. I was also trying to understand my own practice of developing ideas through conversation with AI, publishing some of that work on my blog, and returning to it later as a reader. Finding a manageable research topic gave these questions a practical urgency, but narrowing the topic repeatedly exposed assumptions that I was not yet ready to accept.

Building on my technical background meant finding a closer connection between the technical and theological parts of my work. In the subsequent exchange with AI, I considered possibilities including alignment and theological understandings of sin, but repeatedly returned to pedagogy and the framework I had already developed. I wanted to deepen an existing inquiry rather than change subjects whenever a new technical possibility appeared. The later criticism of the admissions guidance developed through my exchange with AI.

At first, I thought a better understanding of how language models learn and generate explanations might help clarify which educational uses should be permitted. Reading guidance on AI use in graduate applications made the difficulty more concrete. The discussion gradually moved from how an AI system produces language to how an institution identifies a person’s intellectual contribution. By the end, I wanted to leave the question open. We had become more precise about the problem without resolving it.

The further reading incorporated here belongs to a later stage of that inquiry. With AI assisting the literature search and comparison, I have brought the exchange into conversation with research that sometimes supports my objection and sometimes limits it. This is a selective critical review, not an exhaustive systematic review. I distinguish empirical findings from conceptual frameworks and my own proposals, and I do not treat results from mathematics, physics or general writing tasks as direct evidence about theological understanding.

Obtaining the original papers made a further pass possible. With AI assisting extraction and comparison, I examined the relevant methods, operational definitions, tables and limitations, rather than relying on summaries of their conclusions. This changed several emphases below. A successful tutor turned out to include more instructional structure than the label “AI tutor” conveys; an engagement measure depended on a particular definition of superficial interaction; and the supplied edition of a foundational design chapter included a critical commentary that complicated my proposed use of its framework.

What I hoped technical understanding could explain

In What Counts as Progress in AI-Assisted Theology, I had explored the relationship between producing a strong theological argument, developing personal understanding and contributing to scholarship. Those achievements can accompany one another, but none guarantees the others. A valuable argument may exceed its author’s understanding; a student may acquire profound understanding without producing an original scholarly result. AI makes these possible separations particularly pressing, although it did not create them (Yin, 2026a). My first reaction in the conversation had been to describe the three achievements as fundamentally separated by AI: someone might produce a good paper without understanding what it said. The more careful formulation preserved that concern while recognising that the separations were already possible. What needed investigation was how AI changed their relationship and the evidence available for assessing it.

I wanted to give that distinction a technical foundation. One AI suggestion was to choose a theological assignment, compare original passages with AI-assisted revisions, and examine whether measures of textual or semantic change agreed with theological judgments. Such a project could be precise and useful, but I did not think it reached the question I was trying to ask. Measuring how much a text changes would still leave open what kind of assistance had occurred and what the learner had come to understand. I wanted programming or visualisation to investigate something that required technical analysis. More fundamentally, I wanted to understand how training produces linguistic and conceptual capacities, and how a model uses those capacities to generate an explanation that can become part of someone’s learning.

In the conversation, the AI proposed contextual adaptation and in-context learning as a relevant technical focus. That was a reasonable connection to iterative dialogue, but I found it too narrow. I had been asking about statistical foundations, attention and the role of numerical representations because I wanted to understand how the capacity to produce coherent explanations arises. The discussion became clearer when the AI distinguished learning during training, generating an explanation, and the human interaction through which an explanation might support learning. I was particularly interested in the first two as a foundation for examining the third. In the GPT-3 study, for example, tasks and demonstrations were supplied through text without further gradient updates to the model. That is a different process from changing its parameters through training (Brown et al., 2020). My interest extended to the capacities that made such responses possible in the first place.

The original GPT-3 paper also preserves an uncertainty that the phrase “in-context learning” can conceal. Its authors distinguish recognising a task encountered during training from acquiring a new task through the examples in a prompt, and leave open where different performances fall along that spectrum (Brown et al., 2020, pp. 3–4, 34). That helped clarify why the earlier answer had felt incomplete. Naming an observable capacity does not finish explaining its mechanism. Nor does a conversation’s improving coherence show that the model and the learner have undergone the same kind of change.

Research on internal representations offered a more substantial starting point. Geva and colleagues examined how transformer feed-forward layers associate input patterns with contributions to output predictions, and how those contributions combine across layers. Their findings concern particular models and mechanisms; they do not settle whether an AI possesses knowledge in a philosophical sense (Geva et al., 2021). They nevertheless show why a technical account can be more informative than simply observing that a model predicts the next token. A prediction objective, the structures learned in pursuing it, and the educational significance of the resulting explanation are distinct objects of inquiry.

The details of Geva and colleagues’ investigation make that restraint concrete. Their analysis included a sixteen-layer model trained on WikiText-103, examination of input prefixes that strongly activated sampled feed-forward components, and human annotation of recurring patterns. The description of these components as key-value memories concerns a mathematical and functional correspondence within the model (Geva et al., 2021, pp. 5485–5487). It does not identify a collection of consciously held beliefs or establish the provenance of a particular theological claim. Technical interpretability can explain part of a prediction while leaving the claim’s justification and its reception by a student unresolved.

That distinction becomes sharper when read alongside Bender and Koller’s argument about the difference between linguistic form and meaning. Their position paper challenges the inference from success in processing language to understanding grounded in communicative intentions and the world (Bender and Koller, 2020). This is an argument about what training on form alone can establish, not a settled verdict on every subsequent multimodal or interactive system. It also changes how I should express my earlier interest in AI “thinking” through high-dimensional vectors. Numerical representations are mechanisms to investigate; their dimensionality is not evidence of a particular kind of subjectivity. For pedagogy, the model’s capacities, the warrant for its claims and the learner’s understanding remain separate questions.

I initially wondered whether such an account could supply a kind of legitimacy for learning from AI. The assistant cautioned that explaining a mechanism could inform judgments about educational use without deciding them. That correction limited what I could claim: explaining the causal production of a sentence cannot establish its truth, justify every use of it or demonstrate what a student has learned. Equally, describing its production as statistical does not establish that the sentence is meaningless or educationally worthless. What survived was my interest in examining the assumptions behind permission and prohibition. If a restriction depends on a view about what AI supplies or what learning through it involves, technical and educational investigation could test that view. It could not supply the whole normative argument.

This also brought two uses of “generation” into view. In a technical description, a language model generates linguistic output. In an authorship judgment, saying that AI generated an idea attributes an intellectual contribution within a larger process. The first description does not settle the second. A generated question might express a distinction the user already supplied, introduce one the user had missed, or prompt a further idea that neither the initial question nor the response explicitly contained. Understanding the mechanism matters, but assigning contributions requires examining the interaction as well.

The boundary in the guidance

One concrete example that sharpened my concern was a university’s guidance on using AI in a graduate application. Its immediate scope matters: it concerns admissions materials, rather than providing a complete policy for every activity undertaken during a degree. I read it as a concrete example of a wider educational difficulty, while keeping those different settings separate.

The guidance permits AI-assisted initial research, including exploring a field and understanding a topic, provided applicants then “develop your own thinking” and write the application themselves. It also prohibits using AI to “generate the ideas or arguments in your application.” Its stated purpose is to assess applicants’ understanding, reasoning and creativity (Graduate admissions guidance, n.d.-a). The transition between the permitted and prohibited activities was precisely what I could not locate with confidence. The same guidance also asks applicants to explain their initial research use of AI in a footnote. Disclosure is therefore part of the permitted process, although disclosure alone does not define where that process ends.

The wording also suggested a sequence: first use AI to explore, then develop one’s own thinking, then write. I could recognise that sequence as a possible way of working. My difficulty was that the process I was describing repeatedly moved back through those activities. A response led to a question; the question sent me to a source; the source changed my objection; the objection changed what I asked the model next. A permitted attempt to understand a topic might itself bring a research idea into view. A policy could still distinguish acceptable and unacceptable contributions within that process, but the distinction could not be established simply by locating an initial research phase before the thinking began.

Research on writing makes that concern less dependent on the peculiarities of my own workflow. Flower and Hayes describe composing through interacting processes of planning, translating and reviewing, with goals themselves capable of changing during composition (Flower and Hayes, 1981). Their model predates current generative AI. It gives reason to question an overly linear reading of research, thought and expression, while leaving open what assistance is appropriate within each process. If revising a sentence can expose a conceptual weakness, and reviewing an argument can generate a new objective, a rule that separates language support from intellectual contribution needs to explain how it handles those transitions.

On closer reading, the significant point in Flower and Hayes is that writers can change their governing goals through composing. An objection discovered while revising may change what the text is trying to establish (Flower and Hayes, 1981, pp. 366, 377–379). That gives intellectual agency a temporal dimension. It can appear in recognising that an initially plausible objective is inadequate and redirecting the inquiry. My repeated refusal to leave pedagogy behind, and later refusal to replace the comparison with a clearer case of outsourcing, belong to that history. They still require evaluation as judgments; their persistence alone does not establish their quality.

I understand why an admissions process needs an application to provide evidence about the applicant. If someone else supplies the reasoning, research design and expression that are being assessed, a polished document can misrepresent the candidate’s capacities. A rule aimed at preventing that substitution has an intelligible purpose. Yet an intelligible purpose does not automatically produce an adequate account of the cases that fall under it.

The phrase “my own thinking” can refer to several different achievements. I might have first proposed an idea. I might understand an idea someone else proposed. I might have substantially revised it, found reasons to reject part of it, or developed an argument that neither I nor my interlocutor initially envisaged. I might also have written every sentence myself while making little substantive contribution to the reasoning. These descriptions concern different relations between a person and a text.

HCI research adds a useful caution about relying on a person’s sense of ownership. In two studies, involving 30 and 96 participants, Draxler and colleagues found discrepancies between participants’ reported ownership of generated text and their public authorship declarations; greater influence over the text increased their sense of ownership (Draxler et al., 2024). These are findings about perceptions and declarations in particular text-production tasks. They do not determine rightful academic credit or demonstrate learning. They do suggest that “this feels like my work,” “I claim authorship” and “I can justify my contribution” cannot safely stand in for one another. That distinction applies to my own account as much as to an institution’s suspicions.

The second Draxler study is particularly close to my comparison. Participants received postcard texts drawn from the same pool of AI-generated material, presented under human or AI ghostwriter conditions. They attributed more ownership to themselves, and credited the supposed ghostwriter less often, under the AI condition. The study therefore gives evidence that a source label can affect attribution even when the actual production method is held constant. Its setting matters: these were imagined postcards published through an experimental interface, not assessed research proposals (Draxler et al., 2024, §§5–6). The result neither vindicates unequal treatment nor proves that source should never matter. It makes the influence of expectations something to investigate rather than assume away.

The declaration interface also deserves attention. Explicit authorship options in the second study were accompanied by much more frequent AI attribution than the open text field used in the first. Other features differed between the studies, so this comparison does not isolate an interface effect. It nevertheless suggests a question for disclosure design: does a form enable people to describe the contribution they received, or require them to force it into an unsuitable category? A declaration is produced through an institutional interaction too. Its wording and options can affect the evidence an assessor subsequently receives.

The three prohibitions I had singled out also concerned different things: generating the application’s ideas or arguments, writing its materials, and presenting generated text as one’s own. A restriction on idea generation concerns intellectual contribution even when the assistance is disclosed. A restriction on drafting concerns the production of the submitted expression. Misrepresentation concerns what the applicant claims about the work’s production. These can overlap, but an answer about one does not automatically settle the others. Rewriting something myself, for example, would not establish that I originated its argument; acknowledging a contribution would not by itself establish that the contribution was permitted.

The assistant initially responded by distinguishing origin, understanding, contribution, expression and transparency. That was useful. It also cautioned that understanding an idea does not retrospectively make me its sole originator. I could accept that distinction without accepting that it answered my question. I was asking how an external contribution might enter an authentic process of intellectual development, including a process that eventually produces work I can explain and defend. What remained unanswered was why receiving and developing an idea should be treated differently according to whether the initial assistance came from a human or an AI system.

The guidance’s suggestion that AI-produced application material is unlikely to display sufficient depth added another issue (Graduate admissions guidance, n.d.-a). That is a judgment about likely quality, rather than a demonstration that every such process lacks personal intellectual development. My earlier inquiry had deliberately asked what would matter even when the AI’s theological contribution was good. An excellent result could still provide poor evidence of its submitter’s capacities; an excellent external suggestion could also become the starting point for substantial thinking. The quality concern therefore could not, by itself, resolve the ownership and learning questions that had brought me to the policy.

At one point I wondered whether the policy was legally or epistemologically mistaken. The legal claim would require a separate examination of applicable rules and their implementation. Nothing in this discussion establishes legal invalidity. The epistemological question remained available without that conclusion: what relationship between receiving assistance and possessing understanding does the guidance assume, and how well does that relationship describe actual inquiry?

The comparison the conversation kept missing

To its credit, the assistant checked another relevant document. The same university’s research-proposal guidance encourages applicants to seek advice and feedback from supervisors and friends, while also requiring the proposal to be their own work (Graduate admissions guidance, n.d.-b). This supported my comparison, but it also prompted a qualification. Encouraging human feedback does not amount to authorising a friend to devise an entire proposal on the applicant’s behalf.

That qualification was correct. It still did not answer the case I meant. AI can provide advice and feedback too. If a friend may question my assumptions while leaving the substantive thinking to me, the relevant comparison is an AI interaction that does the same. Moving from my example to a machine producing a complete proposal changes the amount and function of assistance. It makes the distinction easier by changing the case under discussion. What bothered me was that the reply contained a sound caution while leaving the comparison untouched. Repeating the caution did not make the unanswered part disappear. I needed to identify that mismatch before the discussion could move forward.

I therefore kept asking the question. The assistant eventually acknowledged that its earlier correction had not addressed the comparison directly enough, and stated that different treatment of equivalent assistance required an additional justification. That was closer to the issue. Yet the qualification about human feedback and AI-generated arguments resurfaced, and I again objected that AI could supply the same kind of feedback. I was not asking the assistant to ignore a difference in the learner’s work; I was asking it to hold that work constant long enough to explain why the source should decide the judgment. Only then did the exchange reach the more consequential formulation that feedback can itself generate ideas.

The assistant then supplied a hypothetical example from Science and Religion:

Does your argument distinguish methodological naturalism from metaphysical naturalism?

No actual friend or professor had said this to me in the encounter being described; it was an illustration developed in the AI exchange. Its usefulness lay in how much such a short question could do. I might investigate the distinction, discover that my argument relies on an assumption I had not recognised, and reconstruct the argument. The question would have contributed to a new intellectual direction. My response could still involve substantial judgment, reading and reasoning.

This exposed the difficulty in treating feedback and idea generation as cleanly separated activities. Feedback can introduce the distinction around which a later argument is organised. A question can suggest a possibility. An objection can make a different research problem visible. The person receiving the intervention may then do considerable work, but the intervention remains part of the history of the resulting idea. What changed here was my diagnosis of the weakness. I had initially asked where the policy drew its boundary. I now had reason to ask whether the activities named on either side of that boundary could be separated in the way the wording seemed to require.

Later reading gave that formulation a stronger educational basis. Carless and Boud understand feedback through the learner’s work of making sense of information and using it to improve subsequent work or learning. Their account explicitly includes information from people and automated systems, and attends to judgment, emotional responses and action (Carless and Boud, 2018). It therefore directs attention to uptake rather than the delivery of a comment alone. My naturalism example can be examined in those terms: what did the learner notice, investigate and change? The framework does not automatically license AI-generated proposals. It does make a policy category called “feedback” incomplete if it considers only who supplied the comment and overlooks what receiving it enabled the student to do.

An intermediary example that I had already raised in my first response to the policy now became more revealing. Suppose a friend passes on a suggestion they obtained from AI, and I develop it in exactly the same way as I would have developed the suggestion directly. The act of passing it through a person does not itself improve the argument or increase my intellectual contribution. If the friend evaluates, corrects or contextualises the suggestion, something relevant has been added. An unchanged relay does not add that work.

This thought experiment is about the justification of a distinction. It is not a procedure for concealing assistance. Knowingly arranging for someone to supply prohibited work and receiving advice whose earlier origins I do not know are different situations. Their difference also reveals a question for policy: is it regulating the applicant’s conduct, the assistance knowingly received, or the full ancestry of the ideas that eventually influence the application? Those objects of regulation require different evidence and impose different responsibilities.

A philosophical account of collective creation broadens the question of contribution without making it disappear. Khosrowi, Finn and Clark propose assessing the different roles of users, systems, developers and creators of training material through features including relevance, control and originality (Khosrowi, Finn and Clark, 2025). Their framework helps resist the assumption that one output must have one exclusive creator, although its normative claims remain proposals rather than an agreed academic marking scheme. For my purposes, a further distinction is essential: contributing causally to a text, deserving credit for it and acquiring a capacity through producing it are different relations. A student’s decisive objection may leave little wording in the final text while substantially changing its argument.

Their account of control includes a further complication: someone may have been able to intervene competently without needing to do so on a particular occasion. The authors also acknowledge how difficult such counterfactual judgments are to establish (Khosrowi, Finn and Clark, 2025, pp. 4498–4499). This prevents me from equating agency with the number of corrections I make. Accepting a sound suggestion can involve judgment; altering an accurate explanation can introduce an error. Moreover, dividing control over an output does not divide learning into corresponding shares. Several people can learn from one exchange, and substantial learning need not confer a claim to having originated its central idea.

There may be relevant differences between human and AI assistance. An adviser may bring accountable expertise, knowledge of a student’s development or a sustained educational relationship. A university may have reasons to limit particular tools in a task designed to assess performance under specified conditions. Such considerations deserve examination. They do not establish that every human suggestion has educational value while an equivalent AI suggestion lacks it. The justification needs to identify the relevant difference and connect it to the purpose of the assessment.

Hardwig’s account of epistemic dependence also prevents me from making independence an impossible ideal. Much scholarly knowledge depends on expertise and evidence that no individual can reproduce in full; responsible reliance is part of inquiry (Hardwig, 1985). This does not give an AI system the authority of a trustworthy expert. It makes the grounds for reliance a substantive question. A teacher’s disciplinary competence, a source’s evidential history and a model’s demonstrated reliability are different possible grounds, each requiring scrutiny. It also distinguishes relying on someone for a justified claim from allowing someone to perform the specific intellectual task an assessment asks me to demonstrate.

Hardwig’s argument is more demanding than a general endorsement of collaboration. He allows that someone may have rational grounds for accepting an expert’s conclusion without being able to understand or assess the expert’s evidence (Hardwig, 1985, pp. 338–339). This qualifies my emerging emphasis on being able to explain and defend everything. A student in Science and Religion might reasonably rely on specialist physics while being expected to scrutinise the theological inference drawn from it. What must be understood directly, and what may be responsibly taken on trust, depends on the educational claim. Reliance can be epistemically justified while leaving a particular learning objective unfulfilled; an assessment must specify which objective it is testing.

Holding the assistance constant in a thought experiment does not establish that all actual human and AI interactions are equivalent. A learner might respond differently to the same words depending on their source, relationship or expectations. Those differences would need investigation. My objection concerned the step from identifying AI as the source to deciding that the learner’s contribution was therefore deficient. The technical possibility of offering comparable feedback does not force permission in every setting, but it requires a justification that addresses that use rather than substituting a more obviously problematic example.

A remaining objection to my comparison is that institutions make rules for classes of cases, not only for idealised pairs of identical interactions. They might justify different procedures because the reliability, availability or supervision of assistance differs across those classes. My thought experiment does not eliminate that possibility. It asks them to identify the difference, explain its relevance and consider whether the resulting restriction is proportionate to the educational purpose. An administrative reason for requiring a particular process should also be distinguished from a claim that learning through another process is impossible. Those claims would require different arguments, even if they led to the same immediate assessment rule.

Nor should the policy be read as a simple declaration that all human contributions are acceptable and all AI contributions are forbidden. It permits some AI-assisted inquiry. My criticism concerns how it classifies the difficult cases within that permission. A consistent reading is possible: support is allowed, substitution is prohibited. The practical problem is deciding how that reading applies when the support changes the applicant’s ideas.

Returning to the same document supplies a useful contrast. For interviews, the guidance permits AI in general preparation but restricts AI and other external assistance, including help from another person, during the interview itself unless an appropriate adjustment has been agreed. The stated aim is to assess the candidate’s own reasoning and understanding (Graduate admissions guidance, n.d.-a). Here the restriction identifies an assessed event and applies across human and AI assistance. That does not resolve the earlier process of developing application ideas, but it shows that the guidance can distinguish preparation from performance without treating every difference as a difference between human and machine origins.

Returning to something I helped produce

My own blogging practice made the problem harder to dismiss as a hypothetical puzzle. I described a process in which I develop a text through iterative AI conversation, publish it, later study it, and eventually explain the ideas in my own words. The question concerns what happens during that return. Can a text that I helped produce become material from which I subsequently learn?

In Who Creates Whom?, I had already used three positions to describe this temporal difference: the person who initiates an inquiry, the person who critically develops what emerges through AI dialogue, and the person who later returns to the preserved work as a reader. They are positions occupied by one continuing, responsible person, rather than three separate selves (Yin, 2026b). That earlier account made the present question more specific. An assessment might concern what I could do during composition, what I can explain later, or how my judgment developed between those moments. A later demonstration of understanding cannot by itself establish independent production of an earlier text; neither does the production history exhaust what can happen during later study.

Earlier replies had concentrated on the publication’s status as a citable record and on the distinction between attribution and assessment permission. Those points addressed real issues, but they did not reach the centre of my question. I returned to the example because the object I wanted to examine was the learner’s development over time. Publication does not by itself make a use permissible. The person who returns to the text may nevertheless understand something that remained unclear during its production. I was asking whether the earlier history of assistance could determine the educational status of that later encounter before the encounter itself had been examined.

A related exchange had already revealed how easily the discussion could return to permission before examining learning. When I raised the possibility of studying AI itself, the assistant emphasised the need for an explicitly agreed research method rather than an automatic exemption. That reply addressed how a project might be authorised. I was also asking how its governing assumptions could be examined, since the adequacy of those rules could itself be the research question. If an institution’s restrictions prevented the investigation of whether the restricted process develops understanding, the rule would risk insulating an assumption from examination. I raised that as a conditional concern about research and academic inquiry; no refusal of such a project had been documented.

An account of learning has to allow that I can come to understand an explanation I did not independently originate. Otherwise, much of what it means to study a book, attend a lecture or work through someone else’s argument becomes difficult to describe. The harder issue is what I must do with the explanation, and what evidence would distinguish understanding from familiarity with its wording. This was the point of my question about the “ontological difference” between studying an AI-assisted blog and studying an article written by someone else. The sources can differ in authority and reliability; I still wanted to know what would make the learner’s subsequent act of understanding fundamentally different.

Remembering a passage or paraphrasing it fluently may coexist with limited understanding. That possibility applies to human-authored texts as well. Conversely, identifying an unsupported premise, explaining why a distinction matters, or recognising how a new objection changes an argument may give better evidence of intellectual development. These are candidate indicators, rather than an infallible test or a description of something I have already established empirically about my own practice.

Scholarship consulted while preparing this article helps keep those differences in view. Chi and Wylie’s ICAP framework distinguishes passive, active, constructive and interactive forms of engagement, proposing different learning processes and outcomes for them. Its discussion of activities such as self-explanation gives resources for examining what a learner does with material, rather than treating exposure to material as equivalent to learning (Chi and Wylie, 2014). The framework does not establish that a long chatbot conversation is educationally constructive, or that its findings transfer unchanged to theological inquiry.

Reading ICAP’s definitions more closely changed what I could infer from the dialogue format itself. “Interactive” requires substantive, constructive contributions and responsive exchanges; it is not established by alternating messages. The framework explicitly allows a computer agent as a partner, but distinguishes a learner’s substantive contribution from simple agreement with a dominant speaker. It also acknowledges that observable behaviour can misrepresent internal activity: someone quietly reading may be thinking constructively, while someone assigned a generative task may merely copy (Chi and Wylie, 2014, pp. 223–224). For my blog practice, this means asking what I contributed to the reasoning rather than treating the presence of conversation as evidence of learning. Even that interpretation remains provisional: the authors describe the knowledge-change processes associated with their categories as hypothetical, rather than directly verified by the framework itself (Chi and Wylie, 2014, p. 239).

I therefore cannot settle the issue by saying that I have “digested” an idea. That word needs an account of the activity and change it describes. But the reverse inference is equally inadequate: learning cannot be ruled out simply because AI participated in producing the material studied. Its origin, my subsequent understanding and my contribution to a later argument can all matter without becoming interchangeable.

There is also a danger in treating the archive as a transparent mirror of my development. A blog post preserves selected formulations, and a conversation record preserves exchanges; neither necessarily records every hesitation, abandoned reading or reason for a decision. On returning to the text, I may learn from it, but I may also mistake recognition for mastery. The appropriate question is consequently temporal as well as conceptual: what can I explain or do now, under what conditions, and how does that compare with what I could do earlier? An account of intellectual formation needs room for that change without rewriting the history of contribution.

Flower and Hayes supply a methodological caution here as well. Their use of concurrent think-aloud protocols responded to the unreliability of retrospective accounts of writing (Flower and Hayes, 1981, p. 368). A saved AI conversation is not such a protocol: it contains what I communicated to the system, alongside the system’s responses, while much reading and reflection can occur elsewhere. Reconstructing the exchange now also changes it into a narrative. The archive can constrain that narrative by showing which questions appeared when, but it cannot make the account a complete record of cognition. Preserving mistakes helps resist hindsight without eliminating its influence.

What studies of learning add to the question

The distinction between performance and learning has a substantial history in educational psychology. Soderstrom and Bjork review evidence that success during practice can be a poor guide to durable retention and transfer; conditions that make practice harder can sometimes improve later learning (Soderstrom and Bjork, 2015). This gives my original threefold distinction a firmer context without collapsing theological understanding into a memory score. It also challenges an attractive interpretation of my own experience: a dialogue becoming clearer and a text becoming better do not establish that I have acquired a capacity I can later use. The improvement is real at the level observed, but its educational significance requires further evidence.

The review also complicates the idea of a later test as a neutral inspection of what earlier study achieved. Retrieval practice can itself change subsequent retention; asking someone to recall and explain may be another learning event (Soderstrom and Bjork, 2015, pp. 185–187). This matters for the sequence I described: producing a blog, studying it later, and articulating its ideas again. Understanding could develop during the return and the explanation, rather than having been fully present at publication. That possibility supports investigating the sequence, while preventing me from using later competence as proof of earlier competence or sole origination. Research would need to distinguish when a capacity is demonstrated from when it may be developing.

Bastani and colleagues make that divergence concrete in a field experiment involving nearly a thousand high-school mathematics students. Access to a relatively unrestricted GPT-based assistant improved practice performance, yet the group later performed worse than the control group when AI access was removed. A tutor with teacher-informed safeguards largely removed that disadvantage, without establishing an independent-performance advantage over the control group (Bastani et al., 2025). The study concerns particular systems, students and short-term outcomes. Its relevance here is that the same broad category of technology supported different patterns of assistance. “AI was used” would be an insufficient description of the educational intervention.

The experiment’s sequence gives the result a more precise meaning. Students reviewed previously taught mathematics, practised with resources determined by their assigned condition, and then took a closed-book, closed-laptop examination. GPT Tutor received teacher-prepared solutions, common mistakes and hints. Its substantial advantage on assisted practice also included the opportunity to check answers before submission; control students saw solutions afterwards. The authors explicitly discuss this difference (Bastani et al., 2025, pp. 2–4). A higher assisted score therefore cannot simply be read as a proportional increase in learning. The guarded tutor’s examination result was statistically indistinguishable from the control result, which neither demonstrates an added benefit nor establishes exact equivalence between the conditions.

A different result appears in Kestin and colleagues’ randomised study of undergraduate physics teaching. Students using a carefully designed AI tutor achieved higher immediate post-test scores than those in the comparison active-learning classes. The intervention included structured activities, expert-prepared solutions and guidance for the model; its results do not isolate the contribution of a language model from that surrounding educational design (Kestin et al., 2025). Nor do immediate post-tests establish long-term retention, theological judgment or the capacity to formulate an original research proposal. The result matters because it gives evidence against treating all AI assistance as inherently opposed to learning, while specifying far less than a general endorsement of unrestricted chatbot use.

Kestin and colleagues describe a design failure that is especially relevant to my architectural interest. A system prompt alone did not reliably keep the tutor’s explanations in the intended sequence for multipart problems. The researchers therefore built that sequence into the platform and supplied detailed solutions in the prompts. Platform development took several months, beyond the preparation of individual lessons (Kestin et al., 2025, pp. 4–5, 8). The successful intervention consequently incorporated human disciplinary and pedagogical work at several levels. Its result supports examining the whole educational arrangement, including where the software constrains the model, rather than attributing the outcome to the language model alone.

These findings do not cancel one another. They ask related questions in different settings, with different systems, comparison conditions and measures. Together with the older distinction between performance and learning, they change what I would want a policy to explain. Does the assistance preserve opportunities to practise the capacity being assessed? Does it supply a finished solution, or help a learner engage with a difficulty? What remains when the assistance is withdrawn, and what becomes possible when it remains available? These are separate educational questions. A programme might value both independent competence and competent collaboration, but it would need evidence appropriate to each.

The original papers also locate the interventions at different moments in learning. Bastani and colleagues studied review and practice after prior teaching; Kestin and colleagues studied students’ first substantial engagement with the selected material. The latter authors identify their objectives mainly with understanding, applying and analysing, and explicitly limit their claim about complex synthesis and higher-order critical thinking (Kestin et al., 2025, p. 6). These differences do not identify the cause of the contrasting results. They do prevent a straightforward transfer from success in structured introductory tutoring to success in independently framing a theological research problem.

A less conspicuous detail in Bastani and colleagues concerns their analysis of engagement. They classified requests for answers or repetitions of the question as superficial; a conversation counted as nonsuperficial only when it contained none of those messages (Bastani et al., 2025, p. 6). That is an explicit operational choice for studying practice behaviour. It would need reconsideration before being used to judge my workflow. A student might request a complete argument precisely to examine its assumptions and construct an objection. This is my proposed counterexample, not a result from their experiment. It shows why an activity label cannot travel unchanged from one pedagogical purpose to another. The work done with an answer, and the capacity being developed, remain consequential.

They also return my interest in architecture to a practical level. A model’s training, the instructions governing a tutoring interaction, the sources supplied to it, and the interface’s sequence of activities can all affect what work is left for the learner. Technical design is therefore part of the pedagogical argument. It would be premature to infer that a particular design succeeds in theology because it succeeds in physics. It would be equally premature to treat the word “AI” as if it already specified the learner’s role.

What the evidence would have to show

The policy question eventually returned me to assessment. I had earlier described the essay or examination as a proxy: an institution encounters work produced by a person and uses it to draw conclusions about that person’s capacities. This was already an indirect relationship before generative AI. A written submission is evidence, not direct access to its author’s intellectual history. AI made the possibility of divergence more pressing in my inquiry, because improvements in the submitted product need not track improvements in personal understanding. The earlier distinction now had a sharper application: which conclusions about the applicant does this particular piece of work justify?

That is a familiar problem in assessment scholarship, rather than a difficulty discovered for the first time by generative AI. Messick’s account of validity connects the meaning and use of assessment results with supporting evidence, theoretical interpretation and the consequences of decisions made on that basis (Messick, 1994). Applied to my concern, this suggests asking what capacity an application is intended to demonstrate and how permitted or prohibited assistance changes the evidence for that capacity.

Messick distinguishes construct underrepresentation from construct-irrelevant variance, including irrelevant difficulty and irrelevant easiness (Messick, 1994, pp. 4–6). Applied here, assistance might make a task easy by supplying the reasoning it is supposed to test. Conversely, forbidding all language support might make an assessment of theological interpretation depend unnecessarily on another capacity. Neither outcome follows automatically: expression may itself be an intended part of the assessment. The same difficulty affects a move to oral examination. It can provide valuable evidence of reasoning, but its demands and scope also need justification. Restriction and redesign should both be examined for what they reveal, what they obscure and what they inadvertently reward.

Kane’s argument-based approach makes the evidential demand more explicit: claims based on an assessment should be supported through the inferences and assumptions connecting observed performance to its interpretation and use (Kane, 2013). Applied here, a sound theological paragraph might support a judgment about the paragraph’s quality. Inferring the student’s understanding, their capacity to produce a similar argument independently, or their readiness for research each adds a further claim. AI does not make all such inferences impossible. It changes which assumptions need support. This suggests that some policy disputes concern an unstated assessment purpose as much as an unclear definition of assistance.

Kane’s fuller argument adds a distinction between a defensible interpretation and a justified decision. Even an accurate measure can be used badly; supporting one link does not repair a weak link elsewhere (Kane, 2013, pp. 64–65). This reaches beyond the question of whether a student understands an argument. An institution might correctly establish some AI involvement and still need a separate justification for the consequence attached to it. Equally, evidence of present understanding might be insufficient for a task specifically assessing independent origination. The inquiry therefore concerns the warrant for a particular judgment and its use, rather than a single universal verdict on a student’s intellectual authenticity.

A task assessing independent formulation of a research question may reasonably place different limits on assistance from a task assessing the ability to understand and develop a question already supplied. Receiving a central idea from a friend could compromise evidence of independent formulation. Critically examining an externally supplied idea could, in another task, demonstrate the capacity being assessed. The educational purpose is doing work in these judgments; the source label alone cannot carry the argument.

The technical difficulty of establishing provenance adds another complication. Different histories of reading, conversation, revision and composition can lead to similar final prose. Research by Sadasivan and colleagues examines weaknesses in text-detection methods and relates detection performance to the statistical distinguishability of human and model output distributions. It does not establish that every detection method is useless under every condition (Sadasivan et al., 2025). Its subject is the detection of generated text, which is already different from determining who developed an idea or whether someone understands it.

The detection paper’s theorem is conditional: it bounds discrimination by how far apart the relevant text distributions are. The further expectation that stronger models will make those distributions converge is an assumption investigated through particular experiments, not a theorem about every future system (Sadasivan et al., 2025, pp. 11–12). This limits an easy argument on either side. Failure to detect prohibited assistance does not make it permissible. Successful detection would not establish absence of learning. A statistical classification, an assessment interpretation and a policy decision answer different questions, even when an institution connects them in one procedure.

There is also evidence that the errors of detection can be distributed unfairly. Liang and colleagues found that several detectors disproportionately misclassified writing by non-native English writers as AI-generated in the samples they examined (Liang et al., 2023). This does not establish a failure rate for every present-day detector or every student population. It does show why a technical accuracy claim needs to be examined alongside whose writing is most likely to be treated with suspicion. A classifier can appear to offer a convenient answer while shifting the evidential burden onto students whose language already differs from an assessor’s expectations.

The original Liang paper compared 91 TOEFL essays with 88 US eighth-grade essays across seven detectors, and also examined how changes in linguistic expression affected classifications (Liang et al., 2023, pp. 1–3). The comparison is consequential without being a matched experiment isolating native-language status from every difference between the datasets. Its relevance is a demonstrated vulnerability in those samples, not a timeless error rate. For my argument, it makes “sounds like AI” an especially weak substitute for evidence about a learner’s conduct. A judgment sensitive to linguistic predictability can become a judgment about a writer without establishing the intellectual history that the institution actually wants to know.

My further inference is that a classifier of textual origin cannot, on its own, settle a claim about intellectual contribution. Even accurate evidence that a model produced a particular sentence would leave questions about how the sentence was questioned, revised, adopted or understood. An interview can offer evidence of present reasoning and understanding, but it cannot conclusively reconstruct the first appearance of every idea. Drafts and conversation records may help, although their interpretation also requires judgment.

Consider an even stronger record: suppose every exchange, source consultation and revision were available. We might then know that the model first suggested a particular distinction and that the student later altered it. We would still have to interpret what those alterations demonstrated. A visible rejection might show informed judgment, or it might rest on a misunderstanding. An accepted suggestion might have been carefully examined, or adopted because it sounded convincing. Better records could reduce uncertainty about the process without automatically resolving uncertainty about understanding. This is why the problem remains even if the technical difficulty of tracing the text’s production is greatly reduced.

I began to wonder whether a rule intended to protect the relationship between a submission and its author might introduce new uncertainty into that relationship. Applicants may avoid useful learning activities because they cannot tell where permission ends. Assessors may classify similar processes differently. A detailed account of iterative assistance may invite scrutiny that less visible assistance escapes. These are plausible risks raised by the wording, not findings from a study of applicants governed by this particular policy or claims about the motives of its assessors.

The guidance can address clear cases of substitution while remaining weak on collaborative intellectual development. Describing that weakness does not require assuming that every restriction is unjustified. It does require asking whether the boundary is sufficiently clear, educationally defensible and supported by evidence to bear the decisions placed upon it.

When uncertainty becomes an institutional practice

The original conversation raised possible tensions between teachers and students without establishing their prevalence. Further research allows that concern to be stated more concretely. Jiang, Xie and Cao studied survey and interview data from 124 undergraduates and seven faculty members at two US research universities. Students reported recognising classroom rules while finding institutional guidance unclear, and faculty described practical constraints and insufficient training (Jiang, Xie and Cao, 2025). The limited sample and particular institutional settings matter. This is evidence that interpretation and implementation can be difficult, rather than proof that the admissions policy I examined has produced those same effects.

The methods sharpen the limits of that conclusion. The student survey and faculty interviews were analysed separately, with data collected in spring 2024; the study does not observe every stage by which the policies were created. The authors explicitly distinguish their investigation from an empirical explanation of institutional policy formation (Jiang, Xie and Cao, 2025, pp. 4–5, 13–14). I therefore cannot use reports of uncertainty to establish that a rule is conceptually incoherent, or infer a motive for adopting it. Unclear wording, poor communication, inconsistent application and a contested educational purpose are different possible problems. Their remedies may also differ.

Tsao’s qualitative study, drawing on interviews with 58 students and 12 teachers at a university in Hong Kong, identifies the labour involved in interpreting ambiguous AI policy, uneven enforcement and differences in access to tools. It also describes how ambiguity can allow experimentation while leaving individuals to bear responsibility when things go wrong (Tsao, 2025). This adds a dimension my initial comparison did not fully express. A boundary can be unclear conceptually and also distribute practical uncertainty unequally. Students must decide what to disclose and whether a learning activity is safe to attempt; teachers must make defensible judgments without necessarily having time, training or shared examples.

I therefore need to distinguish uncertainty that research cannot yet remove from ambiguity that a policy leaves its readers to negotiate. The first may require modest claims, revisable rules and explicit limits. The second can sometimes be reduced by specifying the assessed capacity, giving contrasting examples and explaining who decides difficult cases. More wording alone would not necessarily help. A long document can remain unclear if it repeatedly uses “own work” without explaining which contribution must be the student’s and why. Equally, a concise rule can be useful where the conditions and purpose are well specified, as in the distinction between preparation and a controlled interview.

Existing design proposals also prevent me from treating the problem as untouched. The AI Assessment Scale developed by Perkins and colleagues links permitted levels of AI use to the learning outcomes of an assessment (Perkins et al., 2024). Its practical contribution is to make expectations more explicit than a single universal permission or ban. A scale does not, however, validate every boundary it contains or tell an assessor what a student has learned. The research question persists within more differentiated arrangements: how should the assigned level be justified, and how should a genuinely iterative case be classified when its activities cross the categories?

The 2024 version of the scale makes the comparison unusually concrete. Its second level permits AI-assisted brainstorming, structuring and idea generation while requiring the submitted text to be human-authored. A later level explicitly assesses human evaluation of AI-produced material (Perkins et al., 2024, table 1 and discussion of levels 2–4). This is a different allocation of assessed work from the admissions example, rather than proof that either allocation is universally correct. It also preserves a difficult boundary: the table’s prohibition on AI “content” in a level-two submission is clarified in the prose as content directly generated by AI. Ideas, wording and final responsibility still need distinguishing. A more nuanced scale makes choices visible; it does not remove the need to interpret them.

Why I want to keep this question within theology

Much of the argument applies across higher education. That was one reason I kept asking how to make it a substantive question in theological pedagogy rather than simply placing a general assessment problem in a theology faculty. I had considered theology broadly and wondered whether a narrower field would help. Science and Religion became a possible setting because it could specify the kind of understanding at stake. The example about methodological and metaphysical naturalism suggests one manageable problem: recognising which claims a scientific explanation supports, where additional philosophical commitments enter, and how a theological argument uses or challenges those commitments. This remains a possible focus, rather than a research design already chosen or validated.

An earlier discussion of adapting my existing framework had exposed a related distinction. The assistant warned that replacing computing components with theological ones might offer limited novelty, particularly given an existing biology adaptation. I objected that adequacy in theology had not been demonstrated by success in another field. Novelty and validity were separate questions. The objection did not prove that a straightforward adaptation would make a strong dissertation; it clarified why disciplinary transfer could itself require substantive examination. I wanted to know whether the framework’s assumptions about evidence, judgment and responsibility remained defensible in theology, and which of those assumptions mattered most urgently for teaching and assessment.

In such a setting, grammatical fluency and successful paraphrase would provide incomplete evidence. I would want to know whether the student can identify the premises, explain why the distinction matters and assess an objection from another position. The relevant standards would need to come from the actual theological and philosophical inquiry. They could not be obtained simply by relabelling a general measure of textual similarity as a measure of theological understanding.

Theology also makes the relationship between receiving and judging especially visible to me. Studying it involves sustained engagement with texts, traditions, teachers and arguments whose claims I did not originate. My contribution can include learning to interpret those claims responsibly and to understand the grounds on which they should be accepted, qualified or rejected. Recognising that dependence does not make all sources equivalent. It makes the justification of their authority, and my responsibility in responding to them, part of the educational task.

Recent theological scholarship confirms that this is already an active conversation. Wang argues for integrating AI literacy, institutional policy and classroom experimentation in theological education, developing a proposal for partnership between human learners and AI (Wang, 2026). I take that as a relevant position to engage, rather than evidence that the learning question has been settled. My narrower concern is what would warrant a judgment that such a partnership has developed theological understanding. Learning to elicit a persuasive theological answer and learning to assess the answer’s premises, sources and implications may overlap, but the relationship needs to be demonstrated.

Wang’s article combines a proposal for theological education with reports of practice, including a survey receiving 34 responses from faculty and staff across fifteen Hong Kong seminaries. It also describes his development of a sermon-preparation GPT and possible classroom uses, including simulated conversations with theological figures (Wang, 2026, pp. 65–67). These details provide concrete practices to examine, while stopping short of a controlled demonstration of improved theological understanding. That distinction lets me take the proposal seriously without turning reported enthusiasm, institutional activity or the author’s account of learning through tool design into evidence for every educational claim.

The simulated interlocutor raises a specifically interpretive problem. A model speaking in a theologian’s voice may help me formulate a question, yet its reply is not historical evidence that the theologian held that position. I would need to compare it with relevant texts and distinguish quotation, reconstruction and imaginative extension. The same concern applies when a tutor presents a synthesis across traditions: which disagreements has it preserved, and which has it smoothed away? These are questions for investigation, rather than findings about every system. They identify where technical design and theological judgment meet—through source selection, instructions, generated explanations and the student’s ability to recognise the status of what is being offered.

For example, the distinction between methodological and metaphysical naturalism could be recited accurately while being applied badly. A learner might explain that a scientific method restricts certain kinds of explanation, then mistakenly treat that methodological restriction as a proof about everything that exists. Conversely, a learner could challenge an argument’s metaphysical step without rejecting the scientific finding on which it draws. These are proposed contrasts for examining understanding, not observations from a completed study. They show why the disciplinary content matters: a generic measure of semantic resemblance could miss the precise inference whose legitimacy is in dispute.

I would also need to state which educational purpose is at issue. An academic study of theological arguments, preparation for ministry and formation within a particular faith community need not assess identical achievements. My immediate question concerns the student’s interpretive and argumentative judgment; it does not presume that agreement with a theological conclusion demonstrates understanding, or that an assessor can infer a person’s faith from polished prose. A useful framework should make such differences explicit rather than allowing “theological formation” to become a name for several unexamined outcomes.

This keeps the technical inquiry relevant. Understanding how models acquire and use representations could help explain what kinds of distinctions their responses preserve, reconstruct or obscure. HCI could examine how a learner engages with those responses. Value Sensitive Design could help investigate how assessment arrangements distribute responsibility and how they balance transparency, fairness and educational opportunity. None of these perspectives can establish the whole case alone, and a model’s internal behaviour would remain different evidence from a student’s learning.

Value Sensitive Design offers a way to keep those inquiries connected. Friedman, Kahn and Borning describe an approach combining conceptual, empirical and technical investigation (Friedman, Kahn and Borning, 2008). Applied to this problem, the conceptual work asks what understanding and responsible contribution mean; the empirical work examines how students and teachers actually interpret and use assistance; the technical work examines what a system enables, records or obscures. These investigations should constrain one another. A demand for complete interaction logs might improve visibility while creating burdens of privacy, access and interpretation. Those tensions belong inside the design of assessment evidence, rather than being treated as administrative details after a technical system has been built.

The supplied VSD text made me reconsider how confidently I could import the framework. It is a 2013 reprint of a 2006 chapter, with an addendum by Alina Huldtgren, rather than simply the 2008 handbook version cited earlier. The addendum discusses criticism of predetermined value lists and distinguishes values explicitly supported by a system, values held by stakeholders and values held by designers (Huldtgren, 2013, pp. 85–88). This is directly relevant to adaptation in theology. Agreement that “understanding,” “fairness” or “formation” matters does not establish agreement about what each requires. My existing framework supplies resources for inquiry; its categories must also remain open to challenge from the people and practices it is intended to serve.

A case in the original chapter makes the iterative method tangible. While redesigning browser consent, the researchers found that excessive interruption could undermine informed consent itself, leading them to add minimal distraction to their account of the value being supported (Friedman, Kahn and Borning, 2006/2013, pp. 62–63). My analogous concern is a demand for exhaustive AI disclosure: documentation intended to support responsible judgment might become burdensome ritual. That is a design hypothesis, not an established effect in theological assessment. It would need investigation with students and teachers, including those indirectly affected by how records are interpreted.

This also suggests a more fundamental role for programming than measuring how many words AI changed. A research system could make competing assumptions about contribution and evidence explicit, then examine how different policy rules classify the same documented sequence of suggestion, criticism, revision and later explanation. Visualisation could reveal where a judgment depends on information the record does not contain. This is a possible extension emerging from the review, rather than a validated method for inferring authorship from a transcript. Its value would depend on whether it exposes consequential assumptions that ordinary discussion leaves hidden, while allowing the theological judgments themselves to remain contestable.

Kane’s caution that validation is systematic without being automatic helps define the computational boundary (Kane, 2013, p. 64). A program could expose which conclusions follow only after a disputed assumption is added, or show how the same record receives different interpretations under different assessment purposes. It could not establish that the assumption is educationally justified merely by executing it consistently. For an MSt, a manageable contribution might be a tightly bounded analysis of one such inference in Science and Religion, with a small computational demonstration if it genuinely assists that analysis. A possible DPhil could investigate the resulting claims empirically across interactions and assessment settings. Those are possible scales of inquiry, not a settled project or a promise of validated automation.

For me, this extends the question behind my previous assessment work. A framework can make permissions, disclosure and evidence more explicit. It still depends on assumptions about what the evidence establishes. The next inquiry may need to examine those assumptions themselves: when does an interaction support the learner’s intellectual development, when does it substitute for a capacity being assessed, and how should uncertainty affect an institutional judgment?

The conversation belongs inside the problem

The exchange through which I reached these questions is itself relevant evidence of a process, although it cannot validate that process simply by describing it favourably. I brought the concern about the policy, the comparison with friends and professors, and the example of returning to AI-assisted writing as a learner. The AI articulated distinctions, retrieved documents and offered possible research directions. I rejected several formulations because they narrowed the technical question too quickly or answered a different case from the one I had posed. Those interventions were intellectual actions even when they consisted of a short question or a request to reconsider an answer. The archive records more polished formulations supplied by AI alongside the objections that made those formulations change. Counting the words each participant produced would miss part of that causal history.

The assistant’s hypothetical example about naturalism helped give the final difficulty a concrete form. Its repeated qualifications also became part of what I had to examine. I could agree that understanding differs from origination, or that feedback differs from someone writing a complete proposal, while still finding that those correct statements did not answer the comparison. Some progress consisted in identifying why an apparently careful reply was insufficient.

There is a further caution in reconstructing the dialogue. My persistence does not prove that my criticism was correct, and the model’s eventual agreement is not independent confirmation. The visible exchange shows that its replies changed after my objections; it does not establish the internal mechanism responsible for that change. I cannot infer a particular training objective or an institutional motive from the transcript. The criticism has to be assessed through the comparison, the documentary evidence and the reasons offered, including the possibility that a relevant difference between cases has been overlooked.

Technical research supplies a reason to take that caution seriously without diagnosing this particular exchange. Sharma and colleagues found sycophantic behaviour in the assistants they studied and evidence that optimisation against human preference models can favour agreement with a user’s views over truthfulness (Sharma et al., 2024). This creates a question about educational alignment: an interaction that satisfies my immediate preference for reassurance may fail to help me recognise a weak argument. Yet disagreement would not automatically be evidence of better tutoring either. The relevant aim would be responsiveness to reasons, including the possibility that my objection is sound. Establishing that responsiveness requires evidence beyond the model’s agreeable tone or my satisfaction with its answer.

The study’s methods also discourage a simplified story about alignment inevitably producing agreement. Sharma and colleagues separately varied users’ stated preferences, examined changes to answers after challenge, and investigated optimisation against preference models. The effects of optimisation differed across methods and measures; some forms of sycophancy were already present before reinforcement learning (Sharma et al., 2024, pp. 3, 7–9). The pedagogical issue therefore becomes more specific: which design choices favour reassurance, correction or reasoned reconsideration, and under what conditions? In theology, where legitimate disagreement may concern premises as well as conclusions, responsiveness cannot mean enforcing one answer indiscriminately. A useful tutor would need to help expose the disagreement and its grounds without treating my preference as evidence that my position is sound.

Reconstructing the exchange now adds another layer. The conversation record helps me distinguish what I initially asked from what became articulable later. It also leaves room for a polished account to make the development look more coherent than it felt while it was happening. I do not want to erase the repetitions or imply that I had already formulated the eventual criticism at the beginning. AI participated in the formulation of that criticism, including its expression here. That participation is part of the authorship question the article examines.

The point at which the AI agreed that the policy was weak was also the point at which I asked to leave the question open. That decision matters to the history of the inquiry. Agreement had given the concern a clearer formulation, but it had not produced a rule that classified every case. Saying that an applicant must retain substantive intellectual judgment points toward something important, while leaving further questions about what counts as substantive, how development should be demonstrated and what burdens of documentation are reasonable. The same is true of treating comparable human and AI assistance comparably: the relevant similarities and differences must still be investigated in particular educational settings.

The further literature has made that openness more demanding. It challenges an ideal of thinking without intellectual dependence, provides evidence that assistance can either support or undermine the acquisition of skills, and documents institutional uncertainty. It also gives me reasons to distrust an easy defence of my own practice: felt ownership, fluent explanation and a preserved dialogue each provide incomplete evidence. The intellectual task is to connect those kinds of evidence to a specific educational claim without allowing one to stand in for all the others.

Closer reading also changes how I would state a positive finding about my own practice. A claim that I learned from an AI-assisted text, a claim that I can now exercise a capacity independently, and a claim that I deserve credit for originating its argument require different histories and evidence. They can become warranted at different times. Later understanding does not erase earlier assistance; earlier assistance does not settle what I can subsequently learn. This temporal distinction is perhaps the most useful development of the original threefold account. It allows the work to remain a record of contribution while also becoming material through which its maker changes.

What changed was the question I wanted to pursue. Initially, I hoped a clearer technical account of AI might help explain whether certain uses should be allowed. I now also want to know whether the categories used to permit and prohibit those uses adequately describe the learning they regulate. Feedback can help bring an idea into existence. An externally suggested idea can become something a learner understands, develops and takes responsibility for. The extent of that development cannot be settled by its origin alone.

I want to leave the comparison open at precisely that point. If I critically develop the same suggestion through the same intellectual work, what educationally relevant difference would justify accepting the assistance from a friend and rejecting it from AI? If there is such a difference, how should we establish it? If there is not, what would an adequate policy have to recognise?

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122.

Bender, E. M., and Koller, A. (2020). “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data.” Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463.

Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. arXiv:2005.14165. https://arxiv.org/abs/2005.14165.

Carless, D., and Boud, D. (2018). “The development of student feedback literacy: enabling uptake of feedback.” Assessment & Evaluation in Higher Education, 43(8), 1315–1325. https://doi.org/10.1080/02602938.2018.1463354.

Chi, M. T. H., and Wylie, R. (2014). “The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes.” Educational Psychologist, 49(4), 219–243. https://doi.org/10.1080/00461520.2014.965823.

Draxler, F., Werner, A., Lehmann, F., Hoppe, M., Schmidt, A., Buschek, D., and Welsch, R. (2024). “The AI Ghostwriter Effect: When Users Do Not Perceive Ownership of AI-Generated Text But Self-Declare as Authors.” ACM Transactions on Computer-Human Interaction, 31(2), article 25, 1–40. https://doi.org/10.1145/3637875.

Flower, L., and Hayes, J. R. (1981). “A Cognitive Process Theory of Writing.” College Composition and Communication, 32(4), 365–387. https://doi.org/10.58680/ccc198115885.

Friedman, B., Kahn, P. H., Jr., and Borning, A. (2008). “Value Sensitive Design and Information Systems.” In K. E. Himma and H. T. Tavani (eds.), The Handbook of Information and Computer Ethics, 69–101. Wiley. https://doi.org/10.1002/9780470281819.ch4.

Friedman, B., Kahn, P. H., Jr., and Borning, A. (2006/2013). “Value Sensitive Design and Information Systems.” In N. Doorn et al. (eds.), Early Engagement and New Technologies: Opening Up the Laboratory, 55–95. Springer. Reprint of the 2006 chapter, with an addendum by Alina Huldtgren. https://doi.org/10.1007/978-94-007-7844-3_4.

Geva, M., Schuster, R., Berant, J., and Levy, O. (2021). “Transformer Feed-Forward Layers Are Key-Value Memories.” Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 5484–5495. https://doi.org/10.18653/v1/2021.emnlp-main.446.

Graduate admissions guidance (n.d.-a). “Using Artificial Intelligence in your graduate application.” Institutional example; the university is unnamed in this discussion. Accessed 31 October 2026. Official guidance.

Graduate admissions guidance (n.d.-b). “Research proposal.” Institutional example; the university is unnamed in this discussion. Accessed 31 October 2026. Official guidance.

Hardwig, J. (1985). “Epistemic Dependence.” The Journal of Philosophy, 82(7), 335–349. https://doi.org/10.2307/2026523.

Huldtgren, A. (2013). “Addendum: Practical Considerations of Value Sensitive Design.” Section 4.8 of B. Friedman, P. H. Kahn Jr., and A. Borning, “Value Sensitive Design and Information Systems,” in N. Doorn et al. (eds.), Early Engagement and New Technologies: Opening Up the Laboratory, 85–95. Springer. https://doi.org/10.1007/978-94-007-7844-3_4.

Jiang, Y., Xie, L., and Cao, X. (2025). “Exploring the Effectiveness of Institutional Policies and Regulations for Generative AI Usage in Higher Education.” Higher Education Quarterly, 79(4), e70054. https://doi.org/10.1111/hequ.70054.

Kane, M. T. (2013). “Validating the Interpretations and Uses of Test Scores.” Journal of Educational Measurement, 50(1), 1–73. https://doi.org/10.1111/jedm.12000.

Kestin, G., Miller, K., Klales, A., Milbourne, T., and Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6.

Khosrowi, D., Finn, F., and Clark, E. (2025). “Engaging the many-hands problem of generative-AI outputs: a framework for attributing credit.” AI and Ethics, 5, 4495–4513. First published online in 2024. https://doi.org/10.1007/s43681-024-00440-7.

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). “GPT detectors are biased against non-native English writers.” Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779.

Messick, S. (1994). Validity of Psychological Assessment: Validation of Inferences from Persons’ Responses and Performances as Scientific Inquiry into Score Meaning. ETS Research Report RR-94-45. https://doi.org/10.1002/j.2333-8504.1994.tb01618.x.

Perkins, M., Furze, L., Roe, J., and MacVaugh, J. (2024). “The Artificial Intelligence Assessment Scale (AIAS): A Framework for Ethical Integration of Generative AI in Educational Assessment.” Journal of University Teaching and Learning Practice, 21(6). https://doi.org/10.53761/q3azde36.

Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., and Feizi, S. (2025). Can AI-Generated Text be Reliably Detected? arXiv:2303.11156, version 4, revised 17 January 2025; initially submitted in 2023. https://arxiv.org/abs/2303.11156v4.

Sharma, M., et al. (2024). “Towards Understanding Sycophancy in Language Models.” International Conference on Learning Representations. Conference paper.

Soderstrom, N. C., and Bjork, R. A. (2015). “Learning Versus Performance: An Integrative Review.” Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000.

Tsao, J. (2025). “Trajectories of AI policy in higher education: Interpretations, discourses, and enactments of students and teachers.” Computers and Education: Artificial Intelligence, 9, 100496. https://doi.org/10.1016/j.caeai.2025.100496.

Wang, J. Z. (2026). “Faith Seeking Prompting: Reimagining Theological Education in the Era of Generative AI.” Teaching Theology & Religion, 29(1), 61–68. https://doi.org/10.1111/teth.70033.

Yin, R. (2026a). “What Counts as Progress in AI-Assisted Theology.” 15 September. https://www.yin.roma.it/2026/09/what-counts-as-progress-in-ai-assisted-theology/.

Yin, R. (2026b). “Who Creates Whom? Art, AI, and the Transformation of the Maker.” 18 August. https://www.yin.roma.it/2026/08/who-creates-whom-art-ai-and-the-transformation-of-the-maker/.

@misc{YIN2026_4990,
  author = {{YIN}, {Renlong}},
  title = {{Whose Thinking Is Being Assessed? The Unsettled Boundary Between Assistance and Authorship in AI-Assisted Education}},
  year = {2026},
  month = {10},
  day = {31},
  date = {2026-10-31},
  howpublished = {The personal webpage of YIN Renlong (blog)},
  url = {https://www.yin.roma.it/2026/10/whose-thinking-is-being-assessed-the-unsettled-boundary-between-assistance-and-authorship-in-ai-assisted-education/}
}