Here we go again, trotting out policies that sound like sticking one’s chin out and denying the future. We are in the age of AI now, like it or not. Policies like Edinburgh sound like the old days in the industrial revolution where lumberjacks like Paul Bunyan were against the invention of the chainsaw for cutting down trees – they wanted to insist it keep being done with an axe. Guess who went out of business (it wasn’t the chainsaw companies). Bunyan was in the age of the chainsaw, and Edinburgh University in 2026 is in the age of AI; the aim should be a reasonably AI integration policy not hamfisted denialism.

Hear hear!

I suppose the justification works as a kind of cumulative argument, for many of the reasons do not, on their own, justify such a restrictive policy. For mainly two reasons: some of the listed claims are contested (e.g., environmental costs, cognitive harm); and some overgeneralize (at least half of those claims would justify prohibiting using many resources which, while suboptimal for a variety of reasons, we consider acceptable). But I take it that, together, these claims do justify eschewing the use of AI by students. Which prompts two questions:

1. How will the policy be enforced? Presumably not using tools that fall prey to many of the objections listed under the justification. If not then, is the policy enforceable in its wide scope?

2. Is there discussion of extending it to apply to faculty too? If not, why not? If yes, why?

I don’t think we yet have a formal departmental policy on AI based cheating, but this is what I tell my 200-300-level students. 

DO YOU NOT USE ChaTGPT OR ANY SIMILAR SYSTEMS TO WRITE YOUR ESSAYS OR IN PREPARING YOU PRESENTATIONS 

You should write your essays and write them in your own words. No ChatGPT, nor any of its competitors, including Gemini, Claude, Llama, Ernie, and Grok. Not even Grammarly. So far as I am concerned, AI generated essays are instances of plagiarism and that includes essays that are partially generated by an AI. If detected they will be treated a such (that is as instances of Level 3 Misconduct) See below for our plagiarism policy. However, you will probably be doing yourself a disservice if you submit an AI-generated essay even if I don’t manage to catch you out. For a start, you won’t have done the brain-stimulating work that you would have done if you had written the essay yourself. Essays are both a summative and a formative mode of assessment. You get a mark – that’s the summative part – but you also learn by writing them – that’s the formative part. If you don’t write your own essays you may get the summative mark – though of course it will be fraudulent – but without the formative experience. There is also some evidence that undue reliance on AIs can be bad for your brain. The Flynn effect (Look it up!) may be going into reverse. 
Furthermore, cheaters are unlikely to prosper when it comes to the final grade, since they are likely to come a cropper with respect to the final exam, worth a potential 50% of your total mark. This will be three hours long with three questions to answer. AI cheats who have not written their essays are likely to to make a hash of things when answering the corresponding exam questions. You may think that it is possible to get around this by memorising an AI-generated essay, corresponding to an essay prompt, and then regurgitating it in the exam. Not so. Though there will be exam questions corresponding to most of the essay prompts, they will be questions with a twist.  They will ask you to come at the topics from a slightly different angle (so to speak) thus requiring a modicum of intellectual agility and a level of understanding that transcends rote-learning. You will have to answer the specific questions that I ask, not the similar questions that we considered in class. This means that if you have not thought through the issues for yourself – specifically, by writing your own answers to (two of) the class questions – you are likely to regurgitate exam answers that respond to the questions that you hoped I would ask rather than the questions that I shall actually ask. And since I will be marking hard for relevance, there is a good chance that you will fail. 
———————-
Most of my colleagues have instituted or refurbished similar assessment regimes designed to minimise the benefits of AI-based cheating. 

What’s the rationale?

1) As noted on earlier threads LLM’s can facilitate the further education of the already educated. (The ‘already educated’ being those who already know ‘enough’ about a subject to learn more with relative ease.) Furthermore AIs can sometimes facilitate research. We have had plenty of testimonials to that effect.  

2) However, LLM’s pose a major threat to higher education and to education generally since they make it very easy to fake understanding and to fake mastery of some skill, subject or discipline. Unless steps are taken to prevent (or mitigate) AI-based cheating, students can wind up without ENOUGH knowledge in their heads either to deserve their credentials or to use AIs (and other resources) intelligently to extend their knowledge.

The ‘enough’ is important. Obviously you can have mastered some subject even though a lot of the relevant information is NOT stored in your own brain but in books and online resources. But a substantial dollop of knowledge must be there in your head or you won’t have really mastered the subject. Suppose that economic necessity compelled me to put on a course on ‘Stuart Britain’ for the benefit of High School students, a U3A group or even a 100-level history class. I could do a decent job with a week or two’s preparation because fifty years ago we did Stuart Britain as a GCE subject and I have retained an interest in the subject ever since. Consequently a lot of information about Stuart Britain is indeed stored in my brain, and this knowledge is structured. This means a) that I would be in a position to evaluate texts and online resources and b) that having a structured understanding of the period I would be able fill in any gaps in my knowledge with relative ease. I would also know what I needed to know but didn’t. (In this area there are, for me, many known unknowns but not many unknown unknowns.) I would probably be in a good position to detect AI ‘hallucinations’. Because I have internalised ENOUGH information about Stuart Britain I would soon be in a position to lead discussions and answer questions off-the-cuff. Now suppose I was asked to put on a similar course on high-school physics. Could I do this? Perhaps (because I am a competent teacher) but not without months of preparation. The O’level physics that I did all those years ago means that I would (just about) know where to begin but I would have to put myself through an intensive course of study and it would be much harder me to evaluate texts and resources and to distinguish good information from bad. Furthermore, to begin with, I would not know much about what more I needed to know. LLMs alone would not enable me to do a decent job. The reason is that when I was a teenager I did not take as much pains to internalise what I was taught about basic physics as I did to internalise what I was taught about Stuart Britain. And in the interim a lot of that knowledge has gone to rust. 

The point generalises. The things you ‘know’ as an extended mind depend upon your knowing ENOUGH as an unextended mind. In order to be a competent doctor, lawyer or engineer, you have to have internalised enough structured information to use external resources effectively, and that includes AIs. And unless the education system takes serious steps to prevent or mitigate AI-based cheating, many graduates will end up not knowing ENOUGH about anything.Furthermore they will not have learned to learn or and are unlikely to have developed the skills and capacities that we profess to teach. 

Why not? Because people tend to acquire knowledge, a love of learning and the associated skills and capacities, if they are incentivised to do so. (Though we should bear in mind that some of the incentives can be, so to speak, internal, a) because learning things, acquiring skills and doing something difficult can be fun, and b) because many of our students are to some degree morally motivated and are therefore inclined to despise AI-based cheating).) However, LLMs create an environment where students are incentivised not to learn but to fake learning (since by doing so they tend to get higher grades). The robots help create an environment in which cheaters tend to prosper and in which, accordingly, many people cheat. 

Thus to re-incentivise learning we have to disincentivise cheating by reconfiguring our assessment regimes, so that cheaters tend NOT to prosper. This can be done, for instance by putting a large proportion of the mark in a cheat-proof final exam. Is it otherwise undesirable to do this? Yes, because final exams are, in the odious jargon which we have to put up with, ‘summative’ rather than ‘formative’: you don’t learn by doing them but only manifest the learning that you have already acquired. In the absence of cheat-facilitating LLMs an assessment regime based on long-form take-home essays would definitely be preferable (at least in philosophy) since students learn about the topic in the course of writing their essays. So in minimising the share of the grade that goes on take-home essays (for example) and maximising the share that goes to cheat-proof mechanisms such as final exams, there is undoubtedly some loss. Thus the LLMs have done us all a damage by forcing us to revert to to assessment and teaching regimes which are otherwise suboptimal. That does not mean that we don’t have to do it. 
. 
A final point. I think that the graduates of universities which make a point of disincentivising AI-based cheating – for instance by making a song and dance about the policy that at least 50% of the grade in every course must be due to cheat-proof methods of assessment – are likely to be at a premium in the job market, since employers can be reasonably confident that they have earned their credentials. As for the objection that they may be slightly less adept at the use of AIs for legitimate purposes than the graduates of other universities – well, I think they will be able to make up their deficiencies in fairly short order since most AIs are designed to be user-friendly. It isn’t that hard to learn how to use AIs intelligently, especially if you know how to get by without them. 

I used Claude with a set of argument analysis skills developed by Marco Meyer to evaluate the policy. I paste here verbatim the result:

I’ll analyse the policy with your argument tools, but let me pin down scope and format first.

## Analysis of the Edinburgh “No AI” Policy

A structural point first, because it organises everything below. The nine justifications fall into two families that do very different argumentative work:

• Internal/pedagogical (1–5): reasons about what happens to the student’s learning and assessment when AI is used.
• External/ethical-political (6–9): reasons to object to the AI industry — bias, copyright, energy, corporate power.

Only the first family bears on the policy’s actual conclusion, which is a rule about assessed work in a philosophy course. The second family, even where sound, supports something else (procurement policy, divestment, regulation), not a course-assessment prohibition. Bundling them inflates the apparent weight of the case. I’ll critique each, then build the strongest version that survives.

A second thread runs through 1–4: each treats “using AI” as AI does the thinking and writing instead of the student (the substitution assumption). None engages the spectrum between substitution and assistance (brainstorming, generating an objection to rebut, checking a chain of reasoning). Once that distinction is drawn, the strong arguments support banning substitution — which integrity rules already cover — not the blanket scope of “writing, revising, and editing.”

1. Educational significanceReconstructed: writing/organising one’s thought constitutes philosophy (P1); AI use avoids that activity (P2); so it hinders learning (C).
The inference is valid only under the substitution assumption. P2 overgeneralises: not all AI use “avoids the activity” — using AI to surface a counterexample one then evaluates is engagement, not avoidance. The gym analogy is a good intuition pump but cuts against the scope: we don’t ban machines, trainers, or spotters at the gym; we ban having someone else lift for you. So the analogy licenses a ban on cognitive outsourcing, not a ban on all AI assistance. Moderate as stated; strong only for the substitution case.

2. Academic integrityThe strongest of the nine — but it proves too little for a blanket ban. Dishonesty requires misrepresentation. If a student discloses AI use, there is no passing-off and hence no integrity violation, yet the policy bans disclosed use too. So this argument supports a disclosure requirement, not prohibition. P2 (“represents something not one’s own”) again presupposes substitution. And the “manifestation of one’s own learning” principle, applied consistently, indicts permitted aids too — human proofreaders, writing centres, supervisor feedback, citation managers all mediate the final text. The carve-out is in tension with the stated principle.

3. Intellectual autonomyLargely a re-description of (1) in the register of integrity — redundant. The distinctive claim, “failure to think for oneself,” faces a direct counterexample: a student who drafts alone, then has AI generate the strongest objection and answers it, has arguably thought more for herself. The argument also needs a principled reason why mediation by a tutor or peer counts as good pedagogy while mediation by AI counts as a failure of autonomy. None is given.

4. Cognitive harmThe empirically load-bearing argument, and the one that could ground a genuine prohibition if the evidence supported the breadth of the claim. It doesn’t yet. The cited work (a Microsoft self-report survey, an SSRN preprint, a BJET study) largely shows correlational or self-reported effects in substitutive usage; it doesn’t establish that scaffolded or disclosed use causes lasting harm. More deeply, “cognitive offloading” is not per se harmful — writing itself, books, and calculators are offloading (Plato’s Phaedrus worry). The premise needs offloading that is pedagogically damaging in a way distinctive to AI, and the studies don’t deliver that at the needed generality. Conclusion that follows: a case for howAI is taught and sequenced, not for prohibition. Moderate, with an evidential-quality caveat.

5. QualityNon sequitur as a ground for a ban. If AI output is shallow or false, normal marking penalises it — that is a reason it is self-defeating for students, hence a reason the ban is less necessary, not more. It also pulls against (2) and (4): if the work is poor, the worry about students passing it off for high marks weakens. And it is a claim about current output, so it cannot stably ground a standing policy. Weak.

6. BiasA reason to read AI output critically, not a reason it undermines assessment. The assigned canon and the open web are also biased; we respond by teaching critical reading, not by banning the sources. Applied consistently, the principle proves too much. At most supports “use critically” — which is itself a philosophical exercise, mildly in tension with (3). Weak as support for prohibition.

7. PiracyTwo problems. First, “piracy” begs a genuinely unsettled legal/moral question (the fair-use / learning-exception debate over training data) by labelling it. Second, the inference from “trained on unconsented text” to “using the product is collusion” needs a complicity principle that isn’t supplied — we don’t generally treat downstream use of ethically-tainted infrastructure as collusion. The “we contribute prompts” point is weak (user prompts are not training data for most assessed-work tools). There’s also an institutional tension: the university hosts ELM tools while the policy bans them on piracy grounds. Weak, and scope-mismatched: an industry objection, not an assessment one.

8. Environmental impactSame scope mismatch, plus an equivocation between training-scale energy and the marginal footprint of a student query (small). Applied consistently, the principle would target conference travel, cloud storage, and GPU use across the university — selective application looks like special pleading. At most a reason for moderation. Weak here.

9. Political significanceRhetorically strongest, argumentatively weakest. The corporate-profit premise flirts with the genetic fallacy: the producer’s motives don’t bear on whether the tool harms learning. The anti-“inevitability” point is fair but only establishes that we have a choice — it does not establish that prohibition is the right choice (non sequitur). It rests on a false dilemma: prohibit, or capitulate to “corporate takeover,” omitting the middle of critical, regulated integration. And there’s a performative tension — the policy elsewhere prizes autonomous critical thinking, yet here recommends shielding students from the technology rather than equipping them to think critically with and about it.

Cross-cutting weaknessesProves-too-much / consistency. The permitted-tools carve-out (spell check, citation managers, human proofreaders, “grammar checkers conforming to the Proofreading Policy”) is conceptually unstable against the banned “automated paraphrasing” and Grammarly — these sit on a continuum, and the principles in (2) and (4) would indict the permitted side too.

Strength–force mismatch. The justifications assert categorical wrongs (dishonesty, piracy, harm), yet the policy is a defeasible default any Course Organiser can simply switch off. If AI use were really dishonest and piratical, individual opt-out would be incoherent. The premises overshoot the conclusion’s actual deontic force.

Hidden assumption. The policy presupposes AI use is reliably detectable and attributable. Current detection is unreliable, which raises fairness and false-accusation worries the justifications never address.

Steelman: the strongest defensible versionDrop the external arguments (6–9) and the quality argument (5) as decorative or scope-mismatched. Reframe the conclusion to match the policy’s actual defeasible structure, and add the one strong consideration the document gestures at but never states — administrability and fairness under a collective-action problem.

P1. The aim of assessed philosophy work is to develop and demonstrate the student’s own capacities for critical reading, reasoning, and argumentative prose. P2. In philosophy especially, these capacities are constitutively exercised in writing and revising one’s own work — the writing is the thinking, not merely a means to it. P3. Generative AI can perform large parts of this activity to a standard hard to distinguish from student work, and, undisclosed, misrepresents whose capacities are on display. P4. Given P2, even assisteduse risks displacing the constitutive activity; given P3, case-by-case permission is hard to administer fairly and creates temptation that disadvantages abstaining students — a collective-action problem. P5. A bright-line default (“don’t use it”) is more administrable, more equitable across students, and more protective of the constitutive activity than nuanced permissions that are hard to police. C. Therefore a defeasible defaultprohibition — overridable by Course Organisers who design AI-integrated assessments, and by approved adjustments — best serves philosophical education.

This is stronger because it (i) rests only on internal, relevant reasons; (ii) replaces “categorical wrong” with risk management under detection-uncertainty plus fairness, which actually matches a defeasible default; and (iii) makes explicit the genuinely good bright-line argument — that line-drawing and enforcement costs can justify a blanket default even when some nuanced uses are individually fine (as with speed limits and other bright-line rules).

Even this version remains vulnerable on three points: it must still defend banning disclosed, minor assistance (it leans entirely on administrability and the constitutive-activity worry there); it depends on the contested empirical claim that even scaffolded use displaces skill development; and the unreliability of AI detection undercuts the fairness that P4–P5 are meant to secure.

Want me to develop any one of these into a full reply-and-rejoinder, or turn the steelman into a counter-argument forpermitting disclosed use?