I. The Missing Principle
Something remarkable has happened in the space of about three years. Without any coordinating body, without a founding conference, without so much as a shared bibliography, virtually every institution that has published a position on AI in education and research has converged on the same four sentences. AI should not replace human judgement. AI should not replace critical thinking. AI should not replace supervision. AI should assist rather than substitute. Some version of this claim now appears in a Russell Group principles document, in a funder’s grant-application policy, in a UNESCO competency framework, in a vice-chancellor’s welcome email to incoming doctoral students. The convergence is not trivial: a great many institutions that agree on almost nothing else concerning AI governance agree on this.
And yet nobody can explain why. Ask a policy committee what "judgement" means as distinct from the many things AI is explicitly permitted to do, and the answer is invariably a list. Good: proofreading, transcription, formatting, reference management, basic translation. Bad: writing essays, writing peer reviews, conducting the substantive parts of a literature review, drafting the discussion section of a paper. The lists are not wrong so much as they are ungrounded. They are compiled by consensus and intuition rather than derived from any stated principle, which means they can only grow by accretion: a new tool appears, a new controversy erupts, a new item gets added to whichever column feels right at the time. Six months later the list is longer, no more coherent, and no closer to explaining itself than it was the day it was written.
This article proposes that there is, in fact, an underlying principle, and that the reason it has not surfaced is that the entire debate has been conducted at the wrong level of description. The question "did AI produce this labour, or did AI usurp this judgement?" sounds like a serious ontological distinction. It is not one. It is a psychological redescription of a distinction that has to be made somewhere else entirely: in the structure of the process being delegated, not in the faculty that would otherwise have performed it.
II. Existing Frameworks Almost Get There
Two recent interventions come close enough to the right answer to be worth taking seriously on their own terms, and both are worth reviewing before proposing an alternative, because each captures something true that a purely ontological account has to preserve.
Surface policing. The regime described at length in a recent piece is one in which universities try to determine whether AI produced a given piece of submitted text.¹ Detection software, stylistic profiling, declaratory bans on "unaided work": all of this treats the presence of AI in the symbolic surface of a text as the thing that matters. That earlier article argued that this mistakes a symptom for a diagnosis: what is actually being protected by academic integrity regulation was never "text nobody else touched," but some form of intellectual formation that assisted text can either support or hollow out depending on how it was produced. Surface policing fails not because detection tools are unreliable, although they are, but because it asks the wrong question of the artefact. The text is the wrong object of scrutiny. What should be scrutinised is the process that produced it, and the text alone underdetermines that process almost completely.
Labour versus judgement. A recent Instats policy report puts a more promising distinction at the centre of its framework for responsible AI in doctoral research training.² The report’s central move deserves real credit: it displaces the plagiarism framing that still dominates university policy (its own survey of thirty-eight top-tier doctoral universities across fifteen countries finds that the great majority of institutional AI policy does not extend past student-conduct language, with only a small minority reaching what the report calls AI-literacy and valid-research-practice territory) and replaces it with a research-integrity framing built around exactly the kind of demarcation this piece has been pressing for: some tasks are labour, where AI augmentation is acceptable and mistakes are recoverable, and some tasks are judgement, where AI substitution silently degrades the product. This is real progress. It moves the conversation from "who wrote this sentence" to "what kind of cognitive work is this task."
But the framework cannot explain itself, because labour and judgement are psychological categories, not ontological ones. Why is proofreading labour and hypothesis formation judgement? The honest answer, buried in every worked example the report offers, is that judgement tasks feel like the kind of thing a competent researcher does with effort and care, and labour tasks feel like the kind of thing a research assistant, a photocopier, or a well-trained secretary could once have done instead. That is a serviceable intuition. It is not a principle, and its instability shows up exactly where it matters most: at the boundary. Is deciding which papers belong in a literature review labour, because it resembles retrieval, or judgement, because it requires weighing? Is running a validated statistical test labour, because a calculator could once have done the arithmetic, or judgement, because choosing the test at all is a methodological commitment? The report’s own task-by-task taxonomy has to hedge constantly, adding notes and exceptions almost everywhere the boundary gets interesting. That in itself is a symptom that the underlying category is doing less work than it appears to.
III. Living Value Theory Starts Somewhere Else
Living Value Theory begins from a different starting point entirely. It does not ask what a person does with a task, psychologically speaking. It asks what kind of thing the task is, ontologically speaking, independent of who or what is performing it.
LVT holds that reality, insofar as living coordination is concerned, divides into three domains. Nonrecursive phenomena are processes in which a stable input is transformed into a stable output by a procedure that does not feed back into the identity of whatever performs it. selfrecursive phenomena are processes in which the activity loops back onto the agent performing it, such that the agent is changed by having performed it: not merely informed, but constituted differently afterward. Interrecursive phenomena are processes that exist only in the loop between two or more parties, such that neither party’s contribution is separable from the relationship itself, and the coordination is the phenomenon rather than a means to it.
These are not categories of cognition, and this is the point that both surface policing and the labour-versus-judgement framework miss by starting from the wrong side of the relation. They are categories of reality itself: of what kind of process is actually occurring in the world, whether or not a human, a machine, or a committee happens to be the one occurring it. A person can perform a nonrecursive task with tremendous psychological effort, and the task remains nonrecursive. A person can perform a selfrecursive task carelessly and half-attentively, and the task remains selfrecursive; only its outcome, not its ontological status, is degraded by the carelessness. What kind of process a task is does not depend on how it feels to do it. It depends on the structure of what is actually happening.
Education, on this account, is fundamentally about participation in selfrecursive and interrecursive domains: not about the accumulation of correct outputs, which is what nonrecursive processes produce, but about the ongoing formation of a person’s capacities and the ongoing coordination between people that those capacities require and sustain.
IV. The Ontological Criterion
This yields a principle in place of a list. Appropriate AI delegation depends entirely on whether the process being delegated is itself recursive: not on how difficult the process is, not on how it is conventionally described, not on which faculty of mind it seems to call upon, and not on whether the output looks intelligent. The rule that follows is correspondingly simple to state, even where it is not always simple to apply.
Nonrecursive processes may be delegated. selfrecursive and interrecursive processes may not. The reasoning is not a moral injunction bolted onto an otherwise neutral technology. It follows from what each kind of process is. A nonrecursive process does not constitute living coordination in the first place; delegating it changes nothing about who the delegator is or what relationships they sustain, any more than delegating long division to a calculator changed anything about the mathematician who used the calculator to check an already-understood proof. A selfrecursive or interrecursive process, by contrast, is the very thing that formation (of a researcher, a student, a clinician, a citizen) consists in. Delegating it does not offload a burden from the person while leaving the person intact. It substitutes an artefact for the change the process was supposed to produce, and the artefact, however fluent, does not carry the change with it. The person who received AI’s output in place of their own recognition of confusion is not a person who has become less confused. They are a person who has acquired a text that looks like the product of having become less confused.
This is why the labour-versus-judgement framework gets so close and still misses: it is gesturing at exactly this distinction using the vocabulary of psychological effort, when the distinction that actually does the explanatory work is a distinction in the structure of the process, available independently of anyone’s effort or intuition about it.
V. What Counts as Nonrecursive?
The nonrecursive category turns out to be large, and its boundaries are considerably clearer than the boundaries of "labour" ever were, because they do not depend on how demanding a task feels. Nonrecursive processes are material or informational transformations: stable computation, stable retrieval, stable conversion from one format or register into another, where the transformation itself does not depend on, and does not alter, the transforming agent’s own capacities.
Optical character recognition. Bibliography formatting. Transcription against a source that can be independently verified. Translation of a text whose content the translator did not generate. Syntax checking. Statistical calculation once the analytic plan has been fixed. Database searching once the search terms have been chosen. Format conversion. Sorting and indexing. Code execution against an already-specified test. Image resizing.
What is elegant about this list (and this is worth dwelling on, because it is the strongest evidence that the ontological criterion is tracking something real rather than something invented to fit the conclusion) is that every item on it predates generative AI by decades or centuries. People delegated these tasks long before large language models existed: to calculators, to research assistants performing purely mechanical steps, to professional indexers, to photocopiers, to typists, to spreadsheets. Nobody ever thought that hiring a typist to produce a clean copy of a handwritten manuscript compromised the intellectual formation of the author. Nobody thought that a research assistant alphabetising a bibliography was doing something that should have been reserved for the professor’s own developing judgement. AI does not introduce a new category of legitimate delegation here. It expands the range and lowers the cost of a category that has always existed and has never been controversial, precisely because these tasks were never where formation happened.
This classification presupposes a performer for whom the relevant capacities have already consolidated. Section VIII returns to what changes, and what does not, once that assumption is dropped.
VI. Self-Recursive Processes
At the opposite pole sit processes that are not transformations of information at all, however much they may involve information as raw material. They are transformations of the person performing them.
Recognising that one is confused, as distinct from being told that one’s argument contains an error. Changing one’s mind in response to evidence, as distinct from producing a paragraph that states a changed position. Deciding what matters in a body of material: which finding is the interesting one, which objection actually threatens the argument and which is merely noisy. Generating a research question that comes from a genuine gap in one’s own understanding rather than from a plausible-sounding prompt. Weighing evidence against competing considerations that cannot be reduced to a formula. Synthesising disparate material into a genuinely new conceptual configuration, as distinct from producing a fluent paragraph that summarises a configuration other people have already reached. Developing expertise, which is not the accumulation of correct facts but the reorganisation of one’s whole way of engaging a domain. Revising one’s own assumptions in light of a result that does not fit them. Exercising judgement, in the ordinary sense the labour-versus-judgement framework correctly gestures toward but cannot ground.
None of these can be delegated, for a reason that has nothing to do with AI’s current technical limitations and everything to do with what the processes are. They are not tasks that happen to be difficult, such that a sufficiently capable tool might one day perform them adequately in the researcher’s place. They are transformations of the researcher, full stop, and a transformation of the researcher cannot be performed by anything other than the researcher, in the same sense that being trained for a marathon cannot be performed by anything other than the body that will run it. An AI system can produce a text that states a changed position. It cannot change the mind of the person who receives that text without the recipient doing the selfrecursive work of actually reading, weighing, and being altered by it; at which point the selfrecursive process has occurred in the human after all, and the AI’s contribution has functioned, at best, as one more piece of nonrecursive material for that process to work on.
VII. Interrecursive Processes
A further class of processes exists only between people, and education turns out to consist very substantially of exactly this class.
Doctoral supervision. The viva voce examination. Ethnographic interviewing. Peer review, properly conducted rather than reduced to a disclosure checklist. Classroom discussion. Mentoring. Collaborative research in which the collaboration is not merely a division of separable labour but a genuine interdependence of perspective. Negotiation. Leadership, in any sense that involves more than issuing instructions.
None of these can be outsourced to AI, and the reason is structurally sharper than the reason selfrecursive processes cannot be delegated. A selfrecursive process fails when it is delegated because the person is bypassed and does not undergo the transformation. An interrecursive process fails when it is delegated because there is, quite literally, no third term available to receive the delegation: the phenomenon is the relationship, not a product that the relationship happens to generate. A viva examination conducted with an AI system standing in for the examiner is not a degraded viva. It is not a viva at all, in the same sense that a conversation conducted with one party absent is not a degraded conversation but simply not a conversation. This is the sharpest form the argument takes, and it is why the interrecursive category deserves to be treated as analytically distinct from the selfrecursive category rather than folded into it, even though both are equally non-delegable.
VIII. Task Labels Are Not Ontological Kinds
The lists in Sections V through VII invite a natural objection. They classify tasks by their nominal description: bibliography compilation is nonrecursive, gap recognition is selfrecursive, as though "compiling a bibliography" named a single fixed process rather than a label that different people fill with different work. That objection is correct, and the framework needs to say so explicitly rather than leaving the reader to notice the gap on her own.
A task label is not a process. The same operation, described the same way on a syllabus or a methods section, can be a different ontological kind depending on who is performing it and where they stand in relation to the capacity the task calls on. Consider a student compiling her first bibliography. What looks like clerical work is not: she has to learn what makes a source relevant rather than merely adjacent, has to notice when a citation is doing less argumentative work than the text around it implies, has to build, through repeated trial and correction, a sense of a field’s landscape that she does not yet have. That is genuine selfrecursive formation wearing the costume of a mechanical task. Her supervisor performing the physically identical operation, sorting citations, checking formats, deciding what belongs, is not doing anything of the kind. That judgement consolidated decades ago. What remains for her is pure execution, indistinguishable in its outward form from what the student is doing and entirely different in what it is.
This does not reopen the door Section III closed on psychologism. The determining fact is not how the task feels to the person doing it; a bored expert alphabetising references is not thereby forming anything, however tedious the boredom, and an anxious student is not exempt from formation just because the task presents itself as grunt work. The determining fact is a checkable one: whether the capacity in question is still being shaped by this performance or was shaped by performances that finished long ago. That is an empirical question about a person’s developmental history, not a report of how the task feels, and it is exactly the kind of question the framework was always equipped to ask, since Section V already defined nonrecursive transformation as transformation that does not alter the transforming agent’s own capacities. That clause was agent-relative from the start. It simply had not been followed to its conclusion.
There is an obvious worry this raises, and it needs answering directly or the whole framework threatens to unravel: if expertise can turn a selfrecursive task nonrecursive for the person who has already mastered it, why does the same argument not eventually turn everything nonrecursive for a sufficiently senior scholar, including the very judgement the framework was built to protect? It does not, because capacities come in two kinds with respect to how they relate to repetition. Some consolidate: citation mechanics, transcription technique, the execution of a statistical test once the design is fixed. These are closed skills with a learning curve that terminates, and once it has terminated the operation really is procedural, and stays procedural indefinitely, no matter how many more times it is performed. Others do not consolidate, because each occasion presents materially new content that has to be freshly judged rather than merely executed: recognising the gap in a literature that has itself changed since the last time anyone looked, weighing a body of evidence nobody has weighed in precisely this configuration before, deciding whether an unexpected result is signal or noise. Growing expertise makes a scholar faster and more reliable at this second kind of work. It does not make the work stop happening, because the content each occasion presents is genuinely new, not a repetition of content already mastered. This is why nobody examines a senior scholar’s bibliography formatting and everybody, at every career stage, still examines her judgement. The asymmetry is not a convention. It tracks a real difference between capacities that terminate and capacities that do not.
A second and entirely independent axis needs separating out from the first, because the two are often run together and they answer different questions. Even granting that some trace of formation survives in a task for a given performer, a further question remains: is this a capacity the discipline’s own assessment apparatus is actually organised around certifying? Nobody’s publication record is evaluated on whether she compiled her own bibliography. Nobody’s tenure case turns on whether she personally transcribed her interviews. Building bibliographies and transcribing interviews have been routinely delegated to research assistants for as long as research assistants have existed, and this fact does not depend on the ontological argument at all: it would justify the delegation even if a residue of selfrecursive value were still, in principle, available to be extracted from the task, because that residue was never what the discipline was protecting when it built its assessment instruments. What a field certifies is a separate empirical question from what a task might, in principle, still be teaching someone, and institutional policy has to track the former even where the latter persists.
A third axis is different again, and it should not be folded into the first two even though it points toward the same practical conclusion. Something can remain selfrecursive, in the strict ontological sense, and still not be worth insisting on, because selfrecursive engagement has a real cost in time and energy that is itself scarce and that could go toward other selfrecursive work with a far higher return. Transcribing a one-hour interview by hand can take the better part of a working day. For a student, that day may buy a meaningful, if modest, increase in a skill she genuinely still needs. For a seasoned researcher, the same day buys almost nothing she has not already learned, while the analysis, the synthesis, the argument-building that the freed time could go toward instead is exactly the high-stakes recursive work the whole framework exists to protect. On balance, delegating the transcription is not a compromise of the framework’s principles. It is what the framework recommends once opportunity cost is taken seriously as a distinct consideration from the ontological status of the task itself, because the question was never only whether a trace of recursive value exists in an activity, but whether pursuing that trace is the best use of a scarce recursive capacity that has better places to go.
Put together, classifying a nominally identical task requires three questions rather than one: whether the relevant capacity is still forming in this particular performer or has already consolidated; whether it is a capacity the discipline’s own certification practices are actually organised around; and, even where some recursive residue survives both tests, whether engaging it is worth its cost against competing demands on the same scarce capacity. This is why "it depends" survives as the framework’s honest answer to a great many concrete cases without becoming an excuse for abandoning the ontological criterion altogether. It does not depend on mood, on disciplinary custom, or on how demanding a task feels. It depends on a performer’s actual developmental relation to a capacity, on what a field’s assessment architecture is actually built to certify, and on an economising judgement about where scarce recursive effort is best spent: three questions the framework can now ask explicitly rather than leaving buried in an unexamined intuition about labour and judgement. It also means, concretely, that AI policy calibrated for a doctoral student in her first year cannot be the policy that governs a professor with decades in the same task, not as a pedagogical courtesy extended to the less experienced, but because the nominally identical task is, for the two of them, a different ontological kind.
IX. Interlocution Without Delegation
The account so far might seem to imply that selfrecursive and interrecursive processes must occur in isolation from AI altogether: that the moment a researcher opens a chat window to think something through, the thinking has already been compromised. That implication would be a mistake, and correcting it requires a distinction the framework has not yet made explicit: the distinction between delegating a recursive process and actively engaging AI within one.
Delegation has a characteristic structure regardless of which domain it operates in: a task is handed over, a product comes back, and the product is accepted as though the process that would ordinarily have generated it had actually occurred. This is what makes delegating a selfrecursive process illegitimate, not that AI touched the work, but that the human’s own transformation was supposed to happen and did not, because the artefact that returned in its place merely resembles what that transformation would have produced.
Using AI as a sparring partner has a different structure entirely, and the difference is visible in the shape of the exchange rather than in any feature of the tool. A researcher who poses a half-formed argument and asks a model to attack it, to generate the strongest available counterexample, to defend the position she does not hold, has not handed the argument over. She has entered a loop. The model’s resistance is calibrated to whatever she has just said, and what she says next is calibrated to that resistance; the position she ends up holding, if she ends up holding a revised one, is not a position the model gave her but a position she reached by working against what the model gave her. That reaching (the recognition that the counterexample actually lands, the discomfort of realising the argument does not survive it, the reformulation that follows) is exactly the selfrecursive work described in Section VI, not a bypassing of it. The model has not performed her judgement. It has given her judgement something to do.
This is worth calling by its proper name rather than treating it merely as a permissible contamination of selfrecursive work. It is an active interrecursive engagement, even though only one party to it is actually formed by the encounter. What makes an exchange interrecursive was never, on this account, a requirement that both parties be symmetrically transformed by it: that requirement would rule out a great deal of what LVT elsewhere treats as genuinely interrecursive coupling between poles that are not equally reflexive. What makes an exchange interrecursive is that it is a loop in which no party’s contribution is separable from the relationship itself, and genuine sparring produces exactly that: neither the researcher’s revised position nor the model’s next counterexample is separable from the exchange that produced it, even though the model itself carries nothing forward from one exchange to the next. What makes it a loop rather than a delegation is that the human is never permitted to step outside it and simply receive an answer; she has to keep reacting, in real time, to a target that keeps moving in response to her.
The distinction from delegation is therefore not a matter of degree. It is a matter of whether the human’s own recursive judgement is what settles the matter. This is also where the exception is most easily counterfeited, and the danger deserves to be named precisely because the exception is legitimate. A researcher who poses an argument, receives a model’s agreement, and treats that agreement as validation has not sparred with anything; she has delegated the adversarial-review step to a system whose tendency to agree with the framing it is given is now well documented, and mistaken the appearance of a loop for the reality of one.² The same failure occurs in the opposite direction when a researcher treats a single round of AI pushback as though it had settled the question the way a genuine interlocutor’s sustained disagreement would, closing the loop after one exchange rather than continuing to do the recursive work the counterexample was supposed to provoke. Interlocution stays legitimate exactly as long as the researcher remains the party whose judgement the loop is testing, rather than the party whose judgement the loop has replaced.
Nor does this collapse the distinction drawn in Section VII between interlocution and the interrecursive processes proper (supervision, the viva, peer review) that cannot involve AI at all. A sparring partner is not a party to those processes; it cannot examine, cannot certify, cannot be the second term in a supervisory relationship, and nothing about the legitimacy of using AI to stress-test an argument beforehand extends to using AI to stand in for the examiner who tests it afterward. The two cases differ in exactly the way the framework predicts. In the viva, the coordination itself (between examiner and candidate, each answerable to the other in ways that count institutionally) is the phenomenon being assessed, and AI cannot be either term in it. In sparring, the coordination is a private rehearsal that produces one thing: a researcher whose own recursive capacity has been exercised against real resistance before she has to exercise it against a human interlocutor for whom the outcome actually counts.
X. The Hidden Structure of Research
The obvious objection at this point is that research, on this account, becomes wholly off-limits to AI delegation, since research is paradigmatically the kind of activity through which a researcher’s judgement is formed and exercised. That objection moves too fast, and addressing it carefully is what allows the ontological criterion to avoid collapsing into the unhelpful claim that "everything is recursive, so nothing may be delegated."
Research is recursive at the level of the whole activity. It is not recursive uniformly at the level of its components. Take literature review, the task both the Instats report and the earlier discussion of surface policing return to repeatedly, because it is where the boundary is most visible. Finding papers that match a set of search terms is nonrecursive: it is retrieval against a specification, and a well-configured retrieval-augmented tool can do it competently, provided its citations are verified against primary sources rather than trusted on the strength of a plausible-looking reference list. Formatting references is nonrecursive. Extracting a quotation that has already been identified as relevant is nonrecursive.
Recognising that a body of literature has a gap (that nobody has actually asked the question the researcher is about to ask) is selfrecursive: it requires the researcher’s own accumulated sense of a field, which is precisely the kind of transformation of the person that cannot be handed to a tool without the recognition simply failing to occur in anyone. Evaluating whether an argument holds up, as distinct from summarising what the argument claims, is selfrecursive. Integrating genuinely contradictory findings into a coherent account, rather than listing them side by side, is selfrecursive. Reformulating the research question in light of what the literature turns out to actually contain (which almost never matches the question the researcher started with) is selfrecursive in the fullest sense: it is the moment at which the researcher’s own understanding is remade by contact with the material.
The classification just given, like the classification of nonrecursive tasks generally, presupposes a researcher for whom the relevant retrieval and formatting skills have already consolidated, in the sense developed in Section VIII. A first-year doctoral student compiling her first serious literature search is doing something closer to selfrecursive work: learning what a productive search term looks like, learning to recognise when a source is adjacent rather than relevant, learning the shape of a field she does not yet know. Supervision has to calibrate for this difference rather than applying a single rule for "literature review" across a candidature. What is legitimately delegable in the third year of a doctorate may still be worth doing by hand in the first.
The upshot is that AI may legitimately perform the nonrecursive components of research, may never legitimately substitute for the selfrecursive and interrecursive components, and may legitimately serve as an interlocutor that helps drive those components without ever completing them on the researcher’s behalf. This holds task by task, not activity by activity, and person by person. A blanket policy that treats "research" as a single unit to be either permitted or restricted misses the internal structure entirely, in exactly the way that a blanket policy treating "writing" as a single unit misses it. The demarcation has to run inside the activity, not around it, and it has to run differently for different researchers doing what looks, on the surface, like the identical task.
XI. The Hidden Structure of Teaching
Exactly the same analysis applies to teaching, and it is worth stating separately because the two hidden structures are frequently conflated in institutional policy, as though "academic work" were a single undifferentiated mass rather than a composite of nonrecursive supports and recursive formation held together by a shared institutional label.
Teaching consists of nonrecursive supports plus recursive formation. Designing slides is nonrecursive; scheduling is nonrecursive; assembling a reading list from an already-settled set of topics is nonrecursive; generating administrative feedback of the "you are missing a citation on page four" variety is nonrecursive. AI can perform all of this competently, and a great deal of the anxiety currently attached to AI in teaching contexts is misdirected anxiety about precisely these supports, which were never where the formation happened and were never particularly well protected by human labour in the first place. Nobody thought that a badly formatted syllabus constituted a pedagogical achievement.
Understanding a particular student’s confusion (not confusion in general, but this student’s confusion, which is never quite the confusion the textbook anticipates) is interrecursive, because it exists only in the relation between the teacher’s attention and the student’s actual state, and it cannot be inferred from a transcript of the student’s error without the teacher’s own developed sense of how that particular student thinks. Responding to a misunderstanding in a way that actually reaches the student, rather than in a way that is merely correct, is interrecursive for the same reason. Motivating a student, building their confidence, diagnosing a conceptual error that is masked by a superficially correct answer: these remain irreducibly interrecursive, and delegating them to AI does not produce a worse version of teaching. It removes the teaching and leaves behind an administrative simulation of it, however well the simulation is executed.
XII. Competence Finally Becomes Clear
This returns the argument to a question raised at length in the earlier discussion of surface policing: what should universities actually be assessing, once the presence or absence of AI in a submitted text is recognised as the wrong question to ask?³ That essay argued that universities should assess competence rather than symbolic surfaces, and left the term "competence" somewhat underspecified, a gap this framework can now close.
Competence is the capacity to participate successfully in recursive processes. It is not knowledge, in the sense of a store of correct propositions, because a store of correct propositions is exactly what a nonrecursive retrieval system can now supply on demand at negligible cost, and if that store were what competence consisted in, competence itself would have become worthless the moment retrieval became cheap. It is not information, for the same reason. It is not the production of correct outputs, because correct outputs are the signature of successful nonrecursive delegation and can now be generated by systems that have undergone no formation whatsoever.
Competence is the demonstrated capacity to do the selfrecursive and interrecursive work that a domain requires: to recognise one’s own confusion and resolve it rather than paper over it, to weigh genuinely competing evidence rather than perform the weighing, to inhabit an argument well enough to defend it under a genuinely difficult question, to participate in a supervision relationship, a viva, a peer review, in a way that changes and is changed by the encounter. It also includes, on the account developed in Section IX, the capacity to use AI as a sparring partner productively (to pose a position, take a counterexample seriously, and revise), which is itself a recursive skill that can be assessed rather than merely permitted. Assessment, on this understanding, stops asking "did AI write this?" (a question that is both unanswerable with confidence and, more importantly, not the question that actually matters) and starts asking "has this student’s recursive capacity increased?" That is a demanding question to design assessment around. It is also, unlike the question it replaces, actually answerable, because it can be tested directly through exactly the process-based instruments (oral defence, staged drafting, annotated revision, supervision records) that the earlier essay recommended on independent grounds and that this framework now grounds ontologically rather than merely pragmatically.
XIII. A New Taxonomy of AI Use
The practical upshot for institutional policy is a considerable simplification of what has become an unmanageably complicated regulatory apparatus. In place of an ever-lengthening list of permitted and prohibited tools, updated every time a new product launches, every assessment task can be evaluated by asking three questions.
Which components of this task are nonrecursive? These may be delegated freely, with the ordinary disclosure and verification standards that already apply to any tool-assisted work: the standards the Instats report’s taxonomy of AI-use modes describes reasonably well for exactly this category.
Which components are selfrecursive? These must remain human, not as a matter of institutional preference but as a matter of what the task actually is; an AI system cannot perform them on the student’s behalf without the formation the task exists to produce simply failing to occur. This does not prohibit engaging AI as an interlocutor while doing the work, per Section IX; it prohibits accepting AI’s output as though it were the work.
Which components are interrecursive? These must equally remain human insofar as they are processes proper (supervision, examination, peer review) where AI cannot be a party to the coordination at all. Where the interrecursive work in question is a private rehearsal rather than an institutional process, sparring with AI is a legitimate way to prepare for it, provided the preparation is understood as preparation and not substituted for the process itself.
These three questions have to be asked relative to a specific performer, not to a task label in the abstract, for the reasons developed in Section VIII. An answer that is correct for a professor with twenty years of transcription behind her is not automatically the answer for a first-year student who has never transcribed an interview, and an institution that writes a single AI policy covering "research assistants" or "doctoral students" without distinguishing where each candidate actually stands in relation to a given capacity has re-imported the very ambiguity the ontological criterion was meant to remove. Career stage is not a separate consideration bolted onto the framework after the fact. It is one of the facts the framework already requires an answer to before any of the three questions above can be answered at all.
Everything else follows from correctly answering these three questions at the level of the specific task and the specific performer, rather than at the level of the general activity. No list of prohibited tools has to be maintained and continually revised as new products appear, because the criterion tracks the structure of the task rather than the branding of the tool that might be used on it. No impossible detection regime has to be built, because the question being asked no longer depends on identifying AI’s fingerprints in a finished text. No arbitrary disclosure rule has to be enforced case by case, because disclosure obligations attach naturally to the nonrecursive components where AI assistance is legitimate and simply do not arise for the selfrecursive and interrecursive components, where AI assistance was never legitimate as substitution regardless of disclosure. What remains is ontological fit: does the delegation match the structure of the process being delegated, for this performer, or does it not; and, where it is not delegation at all but interlocution, does the human’s own judgement remain the thing that settles the matter.
XIV. Beyond Universities
The criterion generalises immediately beyond the university, because nothing in its derivation is specific to doctoral training, or to education at all.
In medicine, AI may analyse scans, flag anomalies, and pre-screen for known patterns, all of which are nonrecursive transformations of imaging data against learned parameters. It cannot replace the interrecursive coordination between clinician and patient through which a diagnosis is actually delivered, negotiated, and acted upon, nor the selfrecursive clinical judgement by which an experienced physician recognises that a textbook pattern does not, in this particular case, fit. It can serve, as in research, as a differential-diagnosis sparring partner the clinician tests her own reasoning against before committing to it.
In law, AI may retrieve precedent and draft boilerplate, both nonrecursive. It cannot replace the selfrecursive legal reasoning by which a lawyer weighs competing lines of authority and commits to an argument, nor the interrecursive work of advocacy, negotiation, and courtroom examination.
In anthropology, AI may organise fieldnotes, transcribe interviews, and search a corpus, all nonrecursive. It cannot replace ethnographic participation itself, which is interrecursive by definition (the ethnographer’s understanding is constituted in the relationship with interlocutors, not extractable from it), nor the selfrecursive work of theoretical synthesis by which fieldwork becomes an argument.
In psychotherapy, AI may schedule appointments and manage administrative records, nonrecursive tasks with no bearing on the therapeutic relationship. It cannot become the therapy, because therapy just is an interrecursive process; an AI system that produces therapeutic-sounding text is not conducting a degraded therapy session, for the same reason a viva conducted without an examiner is not a degraded viva.
In scientific research more broadly, AI performs information transformation at genuinely extraordinary scale and speed: literature retrieval, code generation, statistical computation, structural prediction. Scientists perform the recursive discernment without which none of that transformed information becomes a finding: recognising which result is surprising, which anomaly is a signal rather than noise, which finding actually answers the question that was asked. That is discernment AI can sharpen by contesting, but never supply by concluding.
XV. The First General Theory of AI Delegation
The arrival of generative AI has exposed something considerably deeper than a plagiarism problem, and it is worth being direct about the scale of the claim this final section is making. For as long as human institutions have existed, work has been divided between what can safely be handed to a tool, an assistant, or a procedure, and what cannot be handed to anything because it constitutes the very capacity, relationship, or coordination the institution exists to sustain. That division has never before required an explicit ontology, because the available technologies mostly delegated obviously nonrecursive work (arithmetic, transcription, retrieval, formatting), and the boundary took care of itself without anyone having to state it. Generative AI is disruptive not because it is unusually powerful in some generic sense but because it is the first widely available technology whose outputs are fluent enough to be mistaken for the products of selfrecursive and interrecursive processes, even though it performs none of them by itself. It produces the surface of judgement without the structure of judgement when it is used as a substitute, and institutions built around assuming that fluent output reliably indicated underlying process have discovered, all at once, that the assumption was never actually load-bearing. It was a convenience of an earlier technological environment, not a principle.
The response to that discovery cannot be a longer list of prohibited tools, because lists cannot outrun the pace at which new tools appear, and it cannot be better detection software, because detection targets the surface, which is precisely the level at which the disruption has occurred. The response has to be a shift in what is being asked at all: not "did a machine produce this," but whether the underlying process is nonrecursive, selfrecursive, or interrecursive for this performer, and, if it is recursive, whether the human’s own judgement is what is actually settling it. That is a question that is stable across tool generations, across career stages, and across disciplines, because it concerns the structure of activity and the standing of the performer rather than the provenance of a text.
Delegate nonrecursive operations. Preserve recursive participation. That is the whole principle, and its simplicity is not a weakness but the appropriate reward for having found the level of description at which the phenomenon actually sits. Preservation does not mean insulation: it means participation remains the site where the recursive work actually happens, whether AI is a source of raw material, an adversary to think against, or absent from the room altogether, and whether the performer is a first-year student for whom a task is still formative or a veteran for whom it no longer is. It is not a recommendation confined to universities, though universities are where the current controversy happens to be loudest. It is a general ontology of AI governance, applicable wherever the question of appropriate delegation arises, because it explains (by reference to the nature of the process being delegated, and to who is delegating it, rather than to the technology doing the delegating) why some uses of AI deepen the capacities institutions exist to cultivate while other uses, however capable the underlying model, quietly erode them. Labour and judgement gestured at this distinction and could not ground it. Surface policing protected the wrong object entirely. What was missing in both cases was not more policy. It was the recognition that recursivity, always relative to who is doing the recursive work, is where the line has always actually run.
Notes
1. Ecks, Stefan. 2026. "From Policing Surfaces to Assessing Competence: Why Universities Are Getting AI Wrong and What They Should Do Instead." Living Value Theory, livingvaluetheory.org.
2. Zyphur, Michael J. 2026. Responsible AI in Academic Research: A Competency Framework for Research Training. Instats Policy Series. DOI: 10.61700/t31oy23grr. See especially the report’s discussion of sycophancy detection and human-as-verifier discipline as an AI-era research competency.
3. Ecks, "From Policing Surfaces to Assessing Competence," §VII.