Universities: policy that survives contact with students

The student page opens from the student's side of the desk: no policy changes what happens at 23:40 the night before a deadline. This page is the other side of that desk. A policy that assumes students will not use generative AI is not a policy; it is a decision not to have one, plus paperwork.

The realistic goal is narrower and more useful: a policy students can actually follow, staff can actually enforce, and the institution can actually defend when a case reaches a hearing.

What the reference documents actually say

Read two positions in the original, because both are more permissive than the folk version circulating in staff rooms.

UNESCO's guidance asks institutions to "validate GenAI systems on their ethical and pedagogical appropriateness for education" — validate, not exclude. It is candid that "The immediate institutional strategy is to uphold academic integrity and reinforce accountability through rigorous detection by humans", and the qualifier immediate is doing real work in that sentence: detection is named as the stopgap, not the destination.

The UK Russell Group's principles, as the UNESCO guidance quotes them, state the destination: universities and schools adopting a progressive approach believe that "rather than seek to prohibit their use, students and staff need to be supported in using GenAI tools effectively, ethically and transparently".

Neither body is telling universities to relax. UNESCO's assessment section is blunter than most integrity offices: "The implications of GenAI for assessment go far beyond the immediate concerns about learners cheating on written assignments" — the call is to "rethink what exactly should be learned and to what ends, and how learning is to be assessed and validated". That is a harder demand than a ban, which may be why bans are more popular.

The ingredients of a policy that survives

A line drawn on ability, not on tools. The enforceable question is the one from the student page: does this use build the ability being certified, or replace it? A tool list goes stale with every model release; the ability test does not.

An authorship boundary stated as fact. Submitting generated work as one's own asserts something false independent of any tool rule, the same accountability logic research publishing applies. Putting this in the policy as misrepresentation rather than unauthorised tool use makes cases about something provable.

A disclosure route that is cheap to use. If declaring AI assistance costs a student more than concealing it, the policy has priced honesty out. One sentence in a submission footer is enough, and the disclosure data tells the institution what is actually happening; a ban produces silence instead.

Per-assessment clarity, not per-institution vibes. "Permitted unless the assessment says otherwise" or the reverse, stated on every brief. Most integrity cases grow in the gap between a vague institutional statement and a lecturer's unstated expectation.

A version date. The policy describes a moving target. One written for the models of 2023 misdescribes what current models can do, in both directions.

The detection trap

A policy that rests on catching generated text rests on an unreliable instrument: fluent human writing gets flagged, lightly edited generated text does not, and the error rates fall on students unevenly. UNESCO's framing — detection by humans, as the immediate strategy — is not an endorsement of detector scores as evidence. An accusation needs the shape it has always needed: a process, a human judgement, and a chance to respond. A percentage from a classifier is none of those.

The deeper problem is the arms race is unwinnable at exactly the assessments that matter most. The durable answer is the one UNESCO points at: assessment formats where the ability is demonstrated live — the viva, the in-room exam, the supervised practical — for the outcomes with real stakes, and honest tolerance elsewhere.

What goes wrong

The confident ban. Use continues, disclosure stops, and the first contested case collapses on the question "how do you know?"

Detector scores as verdicts. An instrument with a meaningful false-positive rate, used on thousands of students, manufactures wrongful accusations at scale.

Policy by lecturer folklore. Ten modules, ten unwritten rules. Students comply with the strictest imagined version or none.

Forgetting staff. Marking with AI assistance raises the same disclosure questions, plus student data questions the students' own use never touches.

Writing it once. A policy with no owner and no review date is a snapshot of 2023 wearing the authority of the institution.

How to check it worked

Give the policy to five students and five reasonable use cases — outline generation, grammar cleanup, practice problems, full drafting, reference finding. If they cannot sort the five cases the same way in ten minutes, the policy fails at the point of use, which is the only point there is. Then check the disclosure numbers a term later: near-zero disclosure under a permissive policy means concealment still pays, and the price is set by the institution.