Writing an Academic Integrity Policy for AI: A Working Template
Most schools now have a sentence about artificial intelligence in their conduct code, and most were written in a hurry: they prohibit what cannot be policed and leave a teacher with a suspicion and nothing beyond a detector score and an uncomfortable meeting. An academic integrity policy written for the present has a harder job: it has to say what learning each assessment certifies, what help leaves that learning intact, how a student shows their authorship, and what the institution owes a student it wrongly suspects.
This article is a working template: the principles a policy should rest on, then each section a department head needs, with model wording. Its premise is the one this publication has held since machine text arrived: the finished product is a weak witness to authorship.
The short answer. A workable academic integrity policy for AI states what each assessment certifies, permits by default any assistance that does not undermine that, requires disclosure, treats detector output as a prompt for inquiry rather than proof, places the burden of proof on the institution, and builds process evidence into the submission itself.
What an academic integrity policy for AI has to settle first
Every assessment certifies something specific: constructing an argument, recalling knowledge, analysing a source, executing a method. The policy's first job is to name that learning, because "permitted assistance" has no meaning until the assessment's purpose is fixed.
Second, permit by default whatever does not undermine the certified learning, since blanket bans teach concealment. The UNESCO guidance on generative AI in education and research, published in 2023, takes the same line, asking institutions to regulate use through human-centred rules rather than attempt exclusion.
Third, require disclosure: a student who used a language model and says so has done nothing wrong, whatever the tier. Fourth, never act on a detector score alone. Fifth, the burden of proof sits with the institution, since a student cannot prove a negative. Sixth, make process evidence part of the submission, so that authorship is established as the work happens. This site's essay on proof of work is the foundation for all six, and the fundamental values published by the International Center for Academic Integrity, honesty, trust, fairness, respect, responsibility and courage, belong in the preamble; a policy that convicts on statistics fails the second and third of them.
Scope and definitions
"AI" covers everything from a spelling checker to a system that writes the whole essay, and a policy that does not distinguish them invites both over-enforcement and evasion.
This policy applies to all assessed work submitted by students of [Department]. "Generative assistance" means software that produces text, code, images, analysis or argument in response to a prompt, edited or not. "Editorial assistance" means software that corrects spelling, grammar or formatting without altering the substance of the student's own text. "Certified learning" means the knowledge or skill an assessment is designed to evidence. Every assessment brief will state its tier, its certified learning, and the process evidence required with submission. Where a brief is silent, Tier 2 applies.
Permitted and prohibited assistance by assessment type
A single rule for all work is wrong in both directions; the workable alternative is a tiered scale, stated once and applied per assessment. The UK Department for Education's position on generative AI in education is compatible: the technology can support learning, and formal assessment must still evidence the student's own work.
| Tier | Typical assessment | Generative assistance | Disclosure | Process evidence with submission |
|---|---|---|---|---|
| 0: Closed | Timed examination, in-class writing, oral examination | Not permitted | None needed | None; the supervised setting is the evidence |
| 1: Restricted | Certifying coursework, extended essay, dissertation chapter | Planning and feedback only; no generated text in the submission | Required, itemised | Draft trail or version history; sampled oral defence |
| 2: Open with disclosure | Formative essays, problem sets, lab write-ups, homework | Permitted, including generated text, if disclosed and the student can account for it | Required, itemised | Disclosure statement; version history on request |
| 3: Integrated | Tasks that assess use of the tool itself, such as critiquing a generated draft | Required | Required, with prompts and outputs | Prompts, outputs and the student's commentary |
Each assessment is assigned a tier from 0 to 3. Editorial assistance is permitted at Tiers 1 to 3. Undisclosed generative assistance at any tier is a breach regardless of extent.
The assessments that carry the certifying weight of a course belong at Tier 0 or 1; most homework belongs at Tier 2.
The disclosure requirement
Disclosure converts an integrity problem into a pedagogical one. A student who writes that they asked a language model for three counterarguments and kept one has handed the teacher a window onto their judgment.
Every submission at Tiers 1 to 3 must include a disclosure statement in the form provided, listing each generative tool used, what it was asked to do, what was kept, altered or rejected, and where in the submission its influence appears. A statement of "none" is treated as a declaration. Disclosure is not an admission of wrongdoing and will not, by itself, affect the mark.
The companion piece on writing an AI disclosure statement for coursework gives students the form.
Evidence, investigation and the falsely flagged student
This section decides whether the policy is just. Detection software produces probabilities, and its failures fall on identifiable groups: Liang and colleagues, 2023, found that widely used detectors flagged a majority of essays by non-native English writers as machine generated while passing native speakers' work.
A detector score is not evidence of a breach. It may prompt a request for process evidence or a conversation. Where a breach is suspected, the teacher will first ask the student to account for the work using the evidence submitted with it and a short oral discussion. The burden of establishing a breach rests with [Department], on the balance of probabilities. The student will be told in writing what is suspected and on what evidence, and may respond with a person of their choosing present.
The policy owes a positive duty to the student who is falsely flagged, and two provisions do most of the work: a suspicion resolved in the student's favour leaves no trace, and the school tells the student what evidence would have ended the inquiry sooner. The article on what a falsely accused student can show lists that evidence; a policy that requires version history at Tier 1 has given the student the means in advance.
Asking a student to walk through their own work is the apprenticeship principle applied to assessment, the argument for making thinking visible that Collins, Brown and Holum made in 1991; a short oral defence that is routine for everyone accuses no one.
Sanctions, appeals and the review date
Sanctions should be proportionate to intent and to the tier. Disclosed use that exceeds the tier is a teaching moment; concealment is the breach.
Where a breach is found, the response will be proportionate to the tier, the extent of undisclosed use, whether the student disclosed on inquiry, and any previous finding. The range runs from resubmission with process evidence, through a mark reduction limited to the affected portion, to referral under the conduct procedure for repeated breaches or breaches in certifying assessment. Use disclosed at submission that exceeded the tier is a matter for feedback and is not recorded as a breach.
A student may appeal any finding to [named role], who was not involved in the original decision, within [ten] school days. A finding overturned on appeal is expunged from the record.
This policy will be reviewed by [date, within twelve months of adoption], and after any change to the tiers or to the tools generally available to students.
The review date is load-bearing: the tools change quarterly, and a policy without one outlives its accuracy. Examination boards and inspection frameworks are taken up in the Briefings pillar, beginning with the seven questions for the Select Committee.
Frequently asked questions
Should an academic integrity policy ban AI outright?
No. A blanket ban cannot be enforced, teaches students to conceal rather than disclose, and misstates the purpose of assessment, which is to certify specific learning. The defensible position is tiered: prohibit assistance where the assessment certifies the skill the tool performs, and permit it elsewhere.
Can a school discipline a student on the basis of an AI detector score?
A policy should say it cannot. Detector output is a probability with documented error rates that fall unevenly on non-native writers. A score can prompt a request for process evidence; it cannot ground a finding, since the burden of proof rests with the institution.
What should a student AI use policy require students to submit?
At any tier above supervised work, a disclosure statement listing tools used, what they were asked, and what was kept or rejected. For certifying coursework, a draft trail or version history alongside the final text. This is ordinary evidence of authorship, gathered before suspicion.
How often should a generative AI policy for schools be reviewed?
At least annually, with a fixed date written into the policy, and sooner if the tiers change or new tools become generally available. The review should ask whether each assessment still certifies what it claims, whether disclosure is honest, and whether any group of students is flagged disproportionately.
Where this leaves a school
An academic integrity policy for AI is a statement of what a school believes its assessments prove and how it intends to know. Written on the six principles above, it protects the honest student, gives the teacher a procedure that does not depend on an unreliable instrument, and turns most integrity cases into conversations about work.
The Academic Integrity hub collects the rest of the lane, and the process portfolio article describes the submission format that makes most of this policy self-enforcing.
A policy that requires disclosure needs to tell students how to make one: the essays on how to cite AI assistance and on contract cheating after generated text supply the wording and the reasoning a department can lift into its own document.