I. The Overconfidence Problem
On 25 February 2026, the Administrative Court of Kassel issued two decisions concerning students at the University of Kassel who had used generative AI in examination work without disclosure. Both students had their work marked as failed and were excluded from resitting the examination on grounds of exceptionally serious deception. The court upheld the university's decisions. The rulings are not yet final; appeal has been admitted to the Hessian Administrative Court of Appeal.
Within days, the rulings had travelled through LinkedIn, legal commentary sites, and university policy circles with a confidence entirely out of proportion to what the court had actually decided. A representative post by a data protection lawyer with an LL.M. described the ruling as having "set a definitive boundary for the use of AI in universities." It was reshared thousands of times. The comments were enthusiastic. The post concluded with a gender angle — women use AI twenty-five percent less than men according to a Harvard study, possibly because they are more likely to view AI use as ethically problematic — from which the implied moral was clear enough: restraint is wisdom, and the court has now vindicated it.
None of this is what the court said. The ruling is narrow, contextually specific, disciplinarily situated, and explicitly open to being decided differently where institutional regulations permit AI use under specified conditions. What circulated was not a reading of the ruling. It was a moral resolution in legal dress, confident enough to share, thin enough to be wrong about almost everything that matters.
That this kind of commentary spreads so widely and is celebrated so warmly is not a minor annoyance. It is a diagnostic signal. It tells us that there is essentially no public infrastructure for careful thinking about what universities should actually do in the face of generative AI. The space that careful thinking should occupy is being filled by posts that offer resolution instead of analysis, and they spread because resolution is what people want, and careful thinking is not producing it fast enough.
This article attempts to produce some of that thinking. It begins with what the Kassel ruling actually holds, which is less than is claimed but still significant on its own terms. It then describes what universities are currently doing to police AI use and why that regime fails — not merely in its practical implementation but at a more fundamental conceptual level. It argues that existing rules are not abstract moral principles being violated by technology. They are practical compressions of a specific multimediated environment that no longer exists, and treating them as timeless standards is a category mistake with serious consequences. Finally, it proposes a different framework: one that begins not with prohibitions but with the question of what each educational practice is actually trying to cultivate, and designs assessment accordingly.
II. What the Kassel Ruling Actually Says, and What It Doesn't
Reading the court's own press release rather than the commentary about it produces a significantly different picture.
The two cases are importantly distinct. In one, the student admitted using AI. In the other, the university and the court relied on a convergence of circumstantial indicators: a substantial discrepancy between the written examination work and the student's oral demonstration of knowledge, difficulties understanding the task during the writing period, the sudden production of a near-complete draft, repetitive summarising passages, and stylistic features the court described as typical of AI-generated text. These are not equivalent evidentiary situations, and bundling them into a single headline claim about what "AI detection" established misrepresents the weaker of the two cases in particular.
The court's central legal finding is that undisclosed AI text generation can constitute the use of an impermissible aid, and in sufficiently serious cases, a form of deception that warrants exclusion from resitting. It explicitly distinguishes AI text generation from ordinary search engine use on the grounds that AI displaces the student's own selection, evaluation, and processing of materials, and is not itself cited as a source. That distinction is defensible as far as it goes, though it requires more unpacking than the ruling provides.
What the court does not do is establish a universal rule that any AI use in any academic context constitutes fraud. The press release is explicit that a different judgment would be possible where examination regulations expressly permit such use under specified conditions. This is not a minor caveat. It is a constitutional feature of the ruling, which means that the ruling is not about an eternal incompatibility between AI and scholarship. It is about a mismatch between a particular form of assistance and the assessment regime in force at the time.
The specific disciplinary context also matters enormously and has been almost entirely ignored in the commentary. The case that turned on hallucinated citations involved legal writing — a domain whose argumentative architecture depends structurally on the exact status of authoritative sources. A fabricated court decision in a legal footnote is not merely an inaccurate reference. It is a false invocation of authority in a system where authority is literally constitutive of argument. That damage is architecturally specific to legal epistemology. It does not translate straightforwardly into a general claim about hallucinations in academic work. A hallucinated reference in a theoretical anthropology essay is wrong and should not be there, but the epistemic structure of anthropological argument depends more diffusely on conceptual framing, ethnographic adequacy, and argumentative plausibility than on exact precedential chains. To pretend that the court's finding about legal citations settles the question across all disciplines is to erase the single most important disciplinary distinction the ruling was actually responding to.
The further legal claim circulating in commentary — that using AI implies "conditional intent" to bypass independent work — is a substantive legal concept being deployed with considerably more confidence than the ruling warrants. The court's reasoning on intent is specific to the cases before it. It has not established a new doctrine applicable to all AI use across all institutional contexts. Presenting it as such is not legal analysis. It is the lending of legal authority to a predetermined conclusion.
III. The Current Policing Regime: A Taxonomy
Universities are not responding to generative AI with a coherent policy. They are responding with an improvised assemblage of instruments that share a superficial family resemblance — they are all aimed at stopping something called "AI use" — but which rest on incompatible epistemological assumptions and pull in different practical directions. Understanding this assemblage clearly is a precondition for understanding why it fails.
The first and most prevalent instrument is the declaratory ban. Examination regulations across virtually every institution now contain language prohibiting the use of AI, typically in terms that vary between the vague and the circular. "All submitted work must be your own unaided product." "AI-generated content is not permitted unless otherwise specified." "Students must not use tools that perform intellectual tasks on their behalf." These formulations share a common structure: they presuppose that a clear boundary exists between independent and assisted work, and they instruct students not to cross it. What they do not do is specify where that boundary lies, what crossing it looks like, or what counts as an intellectual task as distinguished from a formatting or retrieval task. The phrase "your own unaided work" was already conceptually unstable before generative AI appeared. Under current conditions it is genuinely unintelligible as a guide to action. A student drafting an essay with word-processing software that auto-completes phrases, Grammarly suggesting structural revisions, a browser that summarises search results, and a reading interface that highlights key claims does not know, on the basis of this language, whether any of those tools constitutes crossing the line. The rule is present; its content has dissolved.
The second instrument is algorithmic detection. Tools such as Turnitin's AI detector, GPTZero, and their various competitors promise to identify AI-generated text with sufficient reliability for institutional use. Universities are licensing these tools at scale and, in many cases, using their outputs as primary evidence in misconduct proceedings. The epistemological problem with this is not merely that the tools are imperfect. It is that they systematically lack what would be required for their outputs to function as evidence in any serious sense. Forensic evidence, even in its real-world rather than ideal form, depends on established error rates, reproducibility across different analysts and conditions, validated inference chains between observed features and claimed conclusions, and peer-reviewed methodology. AI detection tools have essentially none of these. Their outputs vary across versions, across demographics of student writers, and across types of writing. They are known to flag second-language writers at substantially higher rates. They flag writers trained in formulaic institutional styles. They cannot, in principle, distinguish between a student who used AI and a student who has simply internalised the generic academic register that AI tends to reproduce. Using their outputs as primary evidence is not rigorous enforcement. It is the importation of a performance of scientific certainty into evidentiary contexts where no such certainty exists.
The third instrument is what might be called stylistic physiognomy. This operates partly through the detection tools but also through informal academic judgment: the identification of prose textures that "feel" machine-generated. Certain features have attained a quasi-official status as AI indicators — the em-dash used with uncharacteristic frequency, repetitive transitional phrases, excessive summary at the end of paragraphs, a generic enthusiasm applied uniformly across heterogeneous content, a smoothness of surface that sits oddly against thin substantive engagement. Some of these features do have some correlation with AI-generated text, particularly from earlier model generations. But the problems with treating them as evidence are multiple and serious. The correlations are not stable: models change rapidly, and what reads as "AI-typical" in 2024 may be ordinary by 2026. The features are not exclusive: they describe perfectly well the prose of weak human writers, of second-language students exercising deliberate caution, of students trained into formulaic academic registers by years of institutional feedback. And the inferential move from "this text has these features" to "this student used AI" to "this student deceived" involves two additional leaps, each requiring independent justification, that stylistic observation cannot supply. What is happening here is exactly what happened in the nineteenth century when Cesare Lombroso attempted to identify the "born criminal" through skull measurements, facial angles, and physiognomic features. The epistemic error is structurally identical: a set of surface correlations is treated as a reliable indicator of a hidden legal-moral status. The analogy is not rhetorical provocation. It names a specific failure mode in evidentiary reasoning that the current detection infrastructure is reproducing with considerable institutional confidence.
The fourth instrument — the one that is actually doing sound epistemic work, when it is used — is oral examination and process documentation. The Kassel court did not rely on detection software. It relied on a mismatch between written and oral performance, supplemented by other contextual indicators. That is a fundamentally different evidentiary standard, because it tests not the surface of the text but the student's capacity to inhabit and account for it. Where universities use oral defence, annotated bibliographies, staged drafting with tutor feedback, or in-class writing exercises, they are genuinely assessing something. The problem is that this instrument is deployed inconsistently, as an afterthought rather than as an architectural feature of assessment design, and without a principled account of what it is testing for.
The picture that emerges from this taxonomy is of a system that looks busy — lots of rules, lots of software, lots of stern language in handbooks — but that lacks both coherent epistemological foundations and any real theory of what it is trying to protect.
IV. Why the Current Regime Fails
The failures of the current regime can be understood at three levels, each of which reveals a different dimension of the problem. They are not independent. The third is the most fundamental and explains why solutions at the first two levels tend to reproduce the problem in a new form.
At the legal and procedural level, the regime fails because vague rules do not meet the basic standard of normative intelligibility. A rule is only capable of guiding action if those subject to it can, with reasonable effort, determine what compliance requires. Where this standard is not met, enforcement becomes arbitrary: two students who have engaged with AI in ways that are functionally similar may receive very different treatment depending on the particular features of their text, the particular sensibilities of their examiner, and the particular institutional culture of their department. This arbitrariness is not a minor operational problem. It is a structural injustice built into the design of the rule. The consequences are predictably asymmetric. Students who are more risk-averse, more deferential to institutional authority, or less confident in their judgment about where the line lies will over-comply, typically at a cost to the quality and efficiency of their work. Students who are more willing to test the boundaries, more confident that they can manage appearances, or simply more instrumentally oriented toward the degree will make strategic use of the ambiguity until the moment when catastrophic sanctions fall. The system is, in this sense, a machine for rewarding the behaviour it nominally prohibits.
At the epistemological level, the regime fails because its primary detection instruments do not meet the standards required of evidence in consequential decisions about students' academic futures. This has been addressed above in relation to algorithmic detection and stylistic physiognomy. The deeper point is that these instruments have been adopted not because their epistemological foundations were assessed and found adequate, but because they perform certainty in a context where institutions feel they need to be doing something. The licensing of detection software by universities is partly a legal risk management decision — it provides a documented process in case of challenge — rather than a genuine commitment to evidentiary rigour. The result is that the most serious decisions about students' careers are being made on the basis of outputs that could not survive scrutiny under any serious standard of forensic reasoning.
At the mediational level — and this is where the analysis must go deeper than either legal commentary or pedagogical handwringing has managed — the regime fails because it is grounded in a category mistake about what the existing rules actually were. The rules are being treated as if they express timeless moral truths about academic integrity, independent thought, personal intellectual effort, authentic authorship. But they are nothing of the kind. They are practical compressions of a specific multimediated environment. They encoded, in simplified symbolic form, a set of working expectations that were grounded in and stabilised by a particular configuration of how intellectual work was materially organised, socially coordinated, bodily enacted, and symbolically represented. That configuration has been transformed. The material dimension of academic work now includes interfaces, search algorithms, summarisation tools, recommendation systems, and generative platforms that actively transform the material before the student encounters it. Social coordination has been partially displaced into asynchronous, algorithmically shaped channels in which the distinction between a peer conversation and a platform interaction is genuinely unclear. The bodily rhythms of intellectual work have been reorganised around screen-based attention, rapid context-switching, and different temporal patterns of concentration and distraction. And the symbolic dimension — the production and circulation of text — has been transformed by a density and speed of available linguistic material that has no historical precedent.
In this environment, the existing rules are not merely outdated. They are ontologically misfitting. They presuppose a configuration of mediation that no longer exists, and they continue to claim authority over a domain whose actual structure they no longer track. When a regulation says that a student must not use "unauthorised assistance," it assumes that assistance can be clearly individuated, identified, and excluded. But in the current environment, assistance is ambient and layered. It is woven into the basic infrastructure of access itself. The question of where tool use ends and independent thought begins cannot be answered by the student with any confidence, because the distinction the question presupposes has become genuinely unclear.
There is a further theoretical point here that goes beyond this specific case. Symbolic systems — rules, regulations, norms, institutional declarations — have no intrinsic failure signal. A rule can persist, can be stated more loudly, can be enforced with greater severity, even as it ceases to coordinate anything. Unlike physical infrastructure, which tends to break down visibly when it no longer fits its purpose, symbolic infrastructure can remain formally intact while becoming practically uninhabitable. This is what is happening with academic integrity regulation. The rules are present, are regularly re-stated, are written into student handbooks and examination scripts, and are enforced through escalating procedural apparatus. What they are not doing is guiding students toward the forms of intellectual work they are supposed to protect, because the categories those rules depend on no longer carve the actual environment at its joints. The current explosion of institutional activity around AI detection is, in this analysis, not a response to the failure of the rules. It is the symptom of that failure. When a symbolic framework loses its grip on practice, the characteristic response is not to redesign the framework but to assert it more forcefully. That is what universities are doing. It does not address the problem.
V. The Moral Injury of Uninhabitable Rules
The framework of academic integrity is not supposed to be merely a compliance regime. It is supposed to sustain a set of norms — about honesty, about attribution, about the relation between intellectual effort and intellectual product — that have genuine value independent of their enforcement. When rules become uninhabitable, they do not simply fail to be followed. They do something more damaging: they sever the connection between the norms and the students' capacity to orient themselves by those norms. This is the sense in which the current situation produces moral injury.
The structural mechanism is what the philosophical literature calls a double bind. Students are simultaneously required to submit independent work and required to work within a mediational environment in which the concept of independence, as the rules define it, cannot be reliably enacted. If they use generative AI in ways that their actual working environment has made ambient, they risk sanction. If they avoid it entirely, they risk competitive disadvantage relative to peers who do not, and relative to the implicit efficiency expectations that contemporary study practice has built in. There is no option that is both fully compliant and costless. The system quietly rewards what it officially prohibits, and it punishes students asymmetrically depending on whether they chose compliance or strategy.
The predictable consequence — and it is a structural consequence, not a moral failing of individual students — is that ethical orientation gets replaced by strategic calibration. Students do not learn to think clearly about when and how AI use is appropriate. They learn to manage the surface of their work so as not to trigger suspicion. The unit of attention shifts from the intellectual quality of the work to the stylistic plausibility of its claimed authorship. This is the precise inverse of what assessment is supposed to achieve. It is produced, reliably and predictably, by rules that cannot be enacted.
The further consequence, which unfolds over time, is legal alienation in the full sense. When rules systematically fail to coordinate practice — when compliance is unintelligible, enforcement arbitrary, and the norm itself practically unreachable — people cease to experience the rule as a guide to action and begin to experience it as an external imposition to be managed. At that point the rule no longer performs its normative function regardless of how frequently it is restated or how severely it is enforced. The university that responds to AI by publishing stricter prohibitions, licensing more detection software, and running more misconduct proceedings is not defending academic integrity. It is performing it — for its own institutional protection — while the actual conditions of intellectual formation continue to deteriorate.
Universities have a genuine duty here, and it is not the duty to enforce more loudly. It is the duty to restore normative intelligibility to the rules, which means being willing to examine whether the rules, as currently formulated, remain capable of doing what they claim to do. That examination is uncomfortable because it requires admitting that the conceptual foundations of academic integrity regulation were always less secure than institutional rhetoric suggested. The Enlightenment model of the solitary original author was never fully enacted in academic work. Academic production has always been a coordination across inherited materials, institutional scaffolding, peer dialogue, textual models, and formal support structures. What universities called "independent work" always bundled together a wide range of forms of assistance, support, and mediation that were acceptable because they were familiar, and declined to examine them because examining them would have been inconvenient.
Generative AI has not created this problem. It has made the problem undeniable.
VI. What Originality Actually Means, and Why That Matters
The concept of originality that governs current academic integrity regulation is a historical artefact. It emerged in the Enlightenment, consolidated in the research university of the nineteenth century, and crystallised into the forms of examination, plagiarism detection, and personal intellectual property that now govern student assessment. It is not a timeless truth about what intellectual work is. It is a specific institutional settlement, tied to specific conditions of production, publication, and professional formation that have never been universal even within the academy, and that are now under terminal pressure.
The settlement was always internally heterogeneous in ways that institutional rhetoric concealed. Even within the Enlightenment tradition, what counted as original contribution varied enormously by discipline, genre, career stage, and purpose. A doctoral thesis in philosophy was expected to make an original argument. A professional legal memorandum was expected to demonstrate disciplined command of existing authority. A medical case report was expected to document clinical observation accurately within established frameworks. A literature review was expected to synthesise responsibly, not to innovate. These are different demands, and treating them all under the single heading of "original work" was always a bureaucratic convenience rather than a conceptual clarification.
In the context of student assessment, the difference between disciplines is even more significant. Students in law and medicine are not, in the main, expected to think originally in the way that students in the humanities and interpretive social sciences are. What law requires of students is disciplined, professionally responsible handling of authoritative materials — the capacity to identify the relevant authority, interpret it correctly, apply it accurately, and reason carefully within established doctrinal frameworks. That is a demanding standard, but it is a standard of reliability and precision rather than of novelty. What social anthropology requires, by contrast, is something closer to what might be called competent, accountable transformation: the capacity to bring inherited frameworks into productive relation with specific empirical or conceptual problems in ways that generate insight rather than merely illustration. These are different intellectual achievements, and they imply different relations between the student's own thinking and the materials that support it.
The current AI debate has collapsed this difference. The Kassel ruling, because it involved legal writing specifically, has been generalised into a universal principle about all student work. But the principle does not generalise. The architectural damage done by hallucinated legal citations is specific to a domain where the exact status of authoritative sources is constitutive of argument. That damage is not replicated in full in a domain where argument proceeds more diffusely through conceptual framing and interpretive adequacy. This is not a licence for sloppy referencing in any discipline. It is a recognition that the epistemic stakes of citation failure are disciplinarily specific, and that policy which ignores this specificity will produce both overenforcement in some areas and underenforcement in others.
What is needed — and what the current debate almost entirely lacks — is a more careful analysis of what each discipline is actually trying to cultivate when it assesses student work: what forms of intellectual competence, what kinds of professional responsibility, what relations between inherited knowledge and independent judgment. Until that analysis is done, per discipline and per assessment type, the concept of "originality" will continue to function as a flag waved over an empty conceptual space.
VII. A Better Framework: Five Principles
The argument so far is a diagnosis, and a diagnosis is not yet a remedy. What follows is a set of principles that a genuinely adequate institutional response would need to enact. They are not a policy template, because templates are part of the problem — they operate at the symbolic level and leave the underlying mediational questions unexamined. They are a framework for thinking, which institutions would need to adapt in detail to their own contexts, disciplines, and assessment cultures.
The first principle is to disaggregate the assessment. The single inflated notion of "independent work" must be replaced by a much more discriminating vocabulary of what is actually being assessed in each task. Source discovery, source evaluation, synthesis, argument design, formal presentation, disciplinary judgment, and professional responsibility are distinct competences. They call for different kinds of engagement, have different relations to tool use, and can defensibly be assisted in different ways and at different stages. A task whose purpose is to assess whether a student can locate and evaluate relevant literature has a different relation to AI assistance than a task whose purpose is to assess whether a student can construct an original argument. A task that is formative and developmental has a different relation to tool use than a task that certifies professional competence. These distinctions cannot be made at the level of a general policy across the institution. They have to be made by the people who design and teach each course, in relation to the specific competences that course is trying to develop.
This implies a significant shift in institutional culture. Currently, academic integrity policy is largely owned by registries, legal teams, and central governance structures. What it actually requires is detailed pedagogical thinking at the level of individual modules, conducted by staff who understand both what they are trying to teach and what the current mediational environment makes possible. Policy cannot substitute for that thinking. It can only provide the framework within which it happens.
The second principle is to shift the criterion of assessment from surface to inhabitation. The final text is increasingly the least reliable and least informative object for assessing student competence. When symbolic surfaces have become cheap — when a plausible-looking argument can be generated in seconds — using the surface as the primary evidence of intellectual formation is like using the smoothness of a cast to assess the strength of the bone underneath. What matters is whether the student can inhabit the work: explain how the argument was built, distinguish its strong claims from its weak ones, identify where the evidence is thin, defend the citation chain under questioning, reframe the central claim in response to a challenge, revise a paragraph in real time and justify the changes.
This does not mean abandoning written work. It means supplementing it — systematically, not as a suspicious afterthought — with forms of assessment that make the student's relation to the work visible. Oral defence, process documentation, staged drafting with annotated revision histories, in-class writing exercises that connect to submitted work, annotated bibliographies that require the student to characterise each source and justify its inclusion: these are not technically demanding innovations. They are adjustments to assessment architecture that follow logically from recognising that the surface has become an unreliable proxy.
The third principle is to build disclosure into the architecture rather than treating it as an afterthought or a threat. Where AI use is permitted, disclosure requirements must be specific and procedurally integrated. Not a catch-all statement at the bottom of a submission form — "I confirm that any AI use has been declared" — but a structured account of what tool was used, at what stage of the work, for what purpose, with what assessment on the student's part of the quality and reliability of its output, and what independent verification they applied. This level of specificity is only achievable if the first principle has been enacted: you cannot design meaningful disclosure requirements for "AI use" in the abstract, only for AI use at a specific stage of a specific kind of task.
The further benefit of this approach is that it makes AI use itself into an object of intellectual engagement rather than a hidden practice to be detected. A student who has used a generative tool to help structure a literature review, and who has documented that use carefully and critically — explaining where the tool's output was useful, where it was unreliable, what they had to verify independently, and what they decided to discard — is demonstrating intellectual competence that is directly relevant to the practice of contemporary scholarship. That is not a concession to laziness. It is a recognition that responsible engagement with new epistemic tools is itself an intellectual skill that higher education should be cultivating.
The fourth principle is to abandon pseudo-forensic detection as primary evidence. This does not mean abandoning attention to textual features altogether. Stylistic indicators can contribute to a contextual picture — as the Kassel court itself handled them, in combination with oral-written mismatch and other circumstantial evidence — but they cannot function as primary evidence in misconduct proceedings. The standard should be: would this evidence survive scrutiny in a serious forensic context? For algorithmic detection outputs, the answer is no. For stylistic physiognomy applied in isolation, the answer is no. For a documented mismatch between written and oral performance, combined with other contextual indicators, the answer can be yes. Institutions should be explicit about this distinction in their policies and should train the staff who make misconduct decisions accordingly.
The fifth principle is to treat the current situation as dynamic rather than as a problem to be solved once. The mediational environment is changing rapidly. A policy that is adequate to the tools available in 2025 will not be adequate to the tools available in 2027. An institution that publishes a comprehensive AI policy and considers the matter closed is not governing responsibly. It is pretending that a dynamic situation has been stabilised. The appropriate response is to build institutional capacity for regular reassessment: mechanisms by which departments and course teams can review what each assessment is actually testing, whether those competences remain relevant, whether the current mediational environment changes what tool use is plausibly ambient, and whether the balance between surface-based and process-based assessment needs to be adjusted.
This requires academic staff who understand what the tools do — not at a technical level, but at the level of what forms of intellectual work they mediate and what they displace. That understanding is currently rare. Many of the people designing and enforcing AI policy have not used generative AI extensively enough to have formed accurate intuitions about what it does well, what it does badly, and what it leaves untouched. Policy made in ignorance of the actual phenomenology of the tool is unlikely to be accurate about what the tool threatens.
VIII. The Fairness Objection, and Why It Cuts the Other Way
The most frequently raised objection to relaxing AI restrictions is that it disadvantages students without access to premium tools. If only students with paid subscriptions to the best generative platforms can use AI assistance, then permitting AI use entrenches existing inequalities. This is a real concern and should not be dismissed.
But the fairness argument cuts in a different direction from the one it is usually pointed. The current regime — vague prohibitions, algorithmically assisted enforcement, stylistic physiognomy — already produces systematic unfairness. Students from non-Anglophone backgrounds are flagged by detection tools at higher rates. Students trained in formulaic institutional styles produce prose that shares features with AI output and are thus disproportionately suspected. Risk-averse students who over-comply are disadvantaged relative to strategically confident peers. Students who cannot afford to retake examinations, or whose visa status makes an academic misconduct finding catastrophic, face qualitatively different stakes from those for whom a resit is a manageable inconvenience. The current regime is not a neutral baseline from which any relaxation would introduce unfairness. It is already deeply unfair, in ways that track existing social and economic inequalities.
A clearer framework — with specified permissions, meaningful disclosure requirements, and assessment that tests inhabitation rather than surface — is more equitable than a vague prohibition that is enforced inconsistently. It creates the same rules for everyone, makes those rules intelligible, and shifts the focus of assessment to competences that cannot be outsourced as easily as text generation. This does not dissolve the access problem — institutions should be actively working to ensure that where AI tools are permitted, all students have equivalent access to them — but it does mean that the fairness objection cannot be used as a reason to preserve a regime that is itself deeply unfair.
IX. The Age of Cheap Symbolic Surfaces
There is a risk, in an argument like this, of appearing to conclude that the situation is simply a technical policy problem awaiting a better solution. That would be a significant underreading of what is at stake. The arrival of cheap symbolic surfaces — the ability to generate plausible-looking text at scale without the processes through which that text would ordinarily have been produced — is a structural transformation of the conditions of intellectual formation. Its implications go beyond assessment design.
Academic work, when it is working well, is not the production of valid sentences. It is a coordination across multiple mediations: reading that builds familiarity with a field; notetaking that forces selective attention; the hesitation of not knowing how to say something; the reorientation that follows discovering you were wrong about something you thought you understood; the drafting that forces decisions about what actually matters; the checking that builds a sense of what can and cannot be relied upon; the discussing that reveals where your thinking is unclear; the revising that progressively tightens the relation between what you mean and what you say. Generative AI can produce, very quickly, a text that has the surface appearance of being the output of all these processes — the smooth transitions, the signalled structure, the appropriately hedged claims — while bypassing most of the inner sequence through which those surface features ordinarily arise.
This is why the tool feels so uncanny in educational contexts. It does not simply replace effort. It selectively replaces the most publicly visible trace of effort while leaving untouched the question of whether the underlying formation has occurred. The text arrives before the thought has ripened. In many cases, the ripening may then never occur.
This is not an argument for Luddism, and it is not an argument that all compression is illegitimate. Higher education has always depended on compression: textbooks replace the rediscovery of first principles, citation conventions replace the full recounting of archival journeys, lectures replace apprenticeship in every domain simultaneously. Much of education consists in learning to use inherited symbolic shortcuts responsibly. The question is not whether compression is permissible but where it remains answerable to understanding and where it drifts free of it. That is the line that good assessment is designed to police, and it cannot be drawn by detection software or by institutional declarations. It can only be drawn by educators who have a clear enough theory of what they are cultivating to recognise when the cultivation has occurred and when it has not.
Universities that respond to AI by asserting existing categories more loudly will look serious. They will issue policies, license software, run proceedings, and produce a stream of outcomes that can be documented and defended. What they will not be doing is thinking. The assertion of symbolic categories that have lost their grip on practice is not a form of institutional integrity. It is a form of institutional self-protection that happens to coincide with the appearance of vigorous action.
The alternative is more demanding and less comfortable. It requires admitting that the concepts "independent work," "original thought," and "authentic authorship" have never been as transparent as institutional rhetoric assumed. It requires examining, discipline by discipline and assessment by assessment, what is actually being cultivated and what assessment can reliably reveal about whether that cultivation has occurred. It requires building assessment architectures that make the student's relation to their work visible, rather than treating the polished surface of submitted text as a reliable window onto the mind that produced it. And it requires the intellectual honesty to acknowledge that some of this work will involve not restoring a system that was working, but designing for conditions that have never existed before.
Generative AI has not abolished the need for intellectual formation. It has made better thinking about intellectual formation unavoidable. That is, in the long run, not the worst thing that could have happened to universities. The question is whether they will recognise it as the opportunity it is, or spend the next decade enforcing a regulation that can no longer do the work it claims to do.
This essay is part of an ongoing series applying Living Value Theory (LVT) to questions of knowledge, formation, and institutional design. LVT is developing at livingvaluetheory.org.