Most integrity policies in force today were written for a world where the main question was “did this student copy someone else's words?” Generative AI changed the question. A student can now produce fluent, original-looking text they did not write, without copying anyone — and the same tools are legitimately embedded in grammar checkers, search engines and the software your institution itself licenses. A policy that doesn't name this reality leaves students guessing and staff improvising, case by case, in ways that will not survive an appeal.
The good news: institutions that have rewritten their policies successfully tend to converge on the same handful of moves. None of them require predicting where the technology goes next. Here is the framework.
1. Define AI use per assessment, not per institution
The single biggest shift is granularity. A campus-wide rule — whether “AI is banned” or “AI is fine” — cannot be right for both a first-year reflective essay and a capstone coding project. The policies that work push the decision down to the assessment level: every brief carries a short, explicit AI-use statement, chosen by the instructor from a small set of institution-approved tiers.
| Tier | What it means for the student | Typical fit |
|---|---|---|
| Permitted | AI tools may be used freely; the graded skill is the outcome, not the drafting. | Ideation exercises, code scaffolding, translation prep |
| Permitted with disclosure | AI may assist, but the student states which tools were used and for what, in a short declaration. | Essays, reports, literature reviews |
| Prohibited | The submission must be the student's own unassisted writing; the graded skill is the writing itself. | Language assessment, reflective work, exams |
Three tiers are usually enough. More than five and instructors stop reading the definitions. The point is not taxonomy — it is that a student should never have to infer the rules from the vibe of the module.
Make disclosure cheap and honest
If disclosure takes a form, a login and a paragraph of justification, students will skip it. A single line at the end of the submission — “I used an AI assistant to suggest structure for sections 2–3 and to check grammar” — is enough. Disclosure that is easy to make honestly is also easy to weigh fairly when something later looks wrong.
2. Set an evidence standard before you need one
The most common failure we see is procedural: an AI-writing percentage appears on a report, and someone treats it as a verdict. No detection signal — ours included — can prove authorship on its own. Detection models estimate how much of a text is statistically consistent with machine generation; they are calibrated aids for triage, with known failure modes on short texts, heavily edited text and some non-native writing styles. That is why iOriginally's AI-writing analysis reports sentence-level signals separately from similarity and prints a human-judgment policy note on every screen and PDF: the tool is built to open a conversation, not to end one.
Write the standard into the policy itself. A defensible finding of misconduct should rest on a convergence of evidence, for example:
- The detection signal — flagged segments and their distribution, read by a trained person, not just the headline number.
- Writing history — drafts, version history, notes, and how the submission compares with the student's prior work in the same module.
- A conversation or viva — can the student explain their argument, define their terms, reproduce a passage of reasoning on request?
3. Build due process students can see
Nothing damages an integrity office faster than the perception that an algorithm accuses and a committee rubber-stamps. Due process for AI-era cases looks much like due process always has — it just needs restating so that everyone knows it applies here too:
- The student is told what was flagged, sees the same report the assessor saw, and is told which tier applied to the assessment.
- They respond before any finding is recorded — in writing, in a meeting, or through a short viva on the submitted work.
- Decisions are made by people with training in reading the evidence, documented with reasons, and appealable to someone who was not involved in the original decision.
- First instances in grey areas default to education — a rewritten submission, a workshop — rather than sanction, unless deception was deliberate.
4. Tell students the rules before the rules are used on them
A policy that lives in a PDF on the governance site is not a policy; it is a liability. Communication is part of the framework, and templates make it cheap:
- A one-paragraph AI-use statement template instructors paste into every brief, with the tier named in the first sentence.
- A plain-language student guide with worked examples of each tier — including examples of what disclosure looks like when done well.
- A standard first-contact letter for suspected cases that states what was flagged, what happens next, and what the student's rights are — in the same calm register you would want applied to yourself.
5. Train faculty on the boundary, not the basics
Faculty do not need a lecture on what a language model is. They need practice on the boundary cases: the strong student whose polished prose trips a detector, the disclosed-use submission where the disclosure seems too thin, the assessment where the tier was never stated. Short, scenario-based sessions once a term — reading real (anonymised) reports together — do more than any handbook chapter. Train the people who will sit on panels first.
6. Review the policy every term — and say so in the policy
Any AI policy written today contains assumptions that will be wrong within a year. Build the revision cycle into the document: a named owner, a standing review each term, and a short change log so staff and students can see what moved. Feed the review with your own operational data — which tiers instructors actually choose, where flags cluster, how cases resolved. Institution-level analytics and configurable policy bands in a platform like ours make that evidence easy to pull, but the discipline matters more than the tooling: a policy nobody revisits is a policy nobody trusts.
Two pitfalls that keep recurring
The blanket ban
Total prohibition feels decisive and fails quietly. It is unenforceable against determined misuse, it criminalises legitimate tools students are taught to use elsewhere in the same institution, and it pushes AI use underground — which destroys exactly the disclosure habit you need for fair adjudication. Worst of all, it makes every case an all-or-nothing fight, because there is no middle tier to resolve into.
The shrug
The opposite failure is treating the question as unanswerable and leaving it to individual instructors with no shared tiers, no evidence standard and no process. That abandons your most honest students — the ones who want to know where the line is — and produces wildly inconsistent outcomes that fall apart on appeal. Inconsistency, not strictness, is what students experience as injustice.
Start smaller than you think
You do not need a perfect document. You need the tiers, the evidence standard, the process, and a date on which you will revise all three. Draft it with students in the room, pilot it in one faculty for a term, and publish the change log. The institutions handling this era well are not the ones with the cleverest technology position — they are the ones whose students can state the rules from memory.
