AI Detection Policy for Schools: A Practical Policy Skeleton

A workable AI detection policy for schools does five things: it says when students must disclose AI use, sets an evidence standard that a detector score alone cannot meet, gives students a real appeal route with named reviewers and timelines, trains staff on what detectors actually measure, and commits to a review cycle. Everything below is a skeleton you can adapt.

What should students actually disclose?

Disclosure is the part most policies fumble, because they either ban AI outright (unenforceable) or say nothing (unfair to the honest student who used it for brainstorming). A disclosure rule works when it is tied to the task, not to the tool.

Write it as a scale, and put the scale on every assignment brief:

1. No AI use. Used for closed-book exams, in-class writing, and tasks where the process is the assessment.
2. AI for preparation only. Brainstorming, outlining, generating practice questions, checking grammar. No AI text appears in the submitted work. A one-line declaration covers it.
3. AI for drafting, with disclosure. A short note at the end of the piece: which tool, what it did, what the student changed, and what the student checked. This is the tier that needs the most clarity, because it is where most real use happens.
4. AI use expected. Permitted or required, as long as the student can explain the output and verify every claim, citation and figure.

Two details matter more than the wording. First, the declaration should be a sentence, not a form; a bureaucratic form teaches students to tick boxes. Second, ask for the process, not just the product. A student who used a chatbot to outline and then rewrote every paragraph in their own words should be able to say so without fear. That is exactly the behaviour you want to reward.

If your staff are still arguing about what “misrepresentation” means, what counts as misrepresentation is the companion piece to hand them before the first policy meeting.

What counts as evidence, and why a detector score is not enough

Detector output is a probability estimate, not proof of authorship. Every policy should say this in plain language, in the policy itself, not in a footnote. Two tools looking at the same essay routinely disagree, and the reasons are structural: they weigh signals like sentence-length variation and word predictability, and they differ in how they aggregate those signals into a document-level number.

We measured this on our own instrument rather than trusting a vendor claim. On 2026-09-13 we scored 23 hand-written DeAIze guides with the DeAIze detector at a 30% flag threshold: the mean AI score was 21%, the median 22%, and the range ran from 0% to 46%. Five of the 23 human-written documents were flagged. That is our own measurement, on our own local ensemble, and it is the honest picture: human writing produces a spread, and some of it lands above a threshold.

So what does count as evidence? Build a hierarchy into the policy:

Evidence type	Weight in a finding	Notes for staff
Detector score (single tool)	Never sufficient alone	Treat as a prompt to look, not a conclusion
Two or more detectors agreeing	Corroborating only	Still not proof; note which tools and what they reported
Draft history (version history, timestamps)	Strong	Shows process, or its absence
Student explanation and oral defence	Strong	Ask them to explain a specific paragraph’s reasoning
Source verification (citations, quotes, data)	Strong	Fabricated references are a separate, clearer failure
Writing-style mismatch with prior work	Weak to moderate	Non-native writers and students who write formally are flagged disproportionately

Two safeguards belong in this section. First, never let a score be the sole basis for an accusation — say it in the policy so a panel cannot quietly ignore it. Second, require the person raising the concern to record what else they noticed. “The detector said a high score” is not a case file; “the detector flagged the middle section, the student could not explain the argument in paragraph three, and the citations do not exist” is.

If you want to see what sentence-level detection actually reports before you write the evidence standard, run a sample of your own past marking through the AI detector — it highlights which sentences carry the signal and which do not, which is far more useful to a panel than one document-level number.

How should the appeal process work?

The appeal is where policies earn or lose the trust of students and staff. Keep it short, time-bound, and separate from the person who raised the concern.

1. Notification within a set window. The student is told in writing what is alleged, which evidence is being relied on, and which section of the policy applies. Give the detector report to the student, not just the conclusion.
2. A chance to respond before any finding. The student submits a written response and, if they wish, a short oral explanation of their process. This is the step that catches most false positives.
3. A reviewer who is not the accuser. One named role — head of department, academic integrity lead, or a small panel — reviews the file. For high-stakes cases, what panels actually assess is worth reading before you draft the terms of reference.
4. A decision with reasons, in writing. Not just “upheld” or “dismissed”, but which evidence carried weight and why the detector score did or did not contribute.
5. A second route. One further review, with a different reviewer, and a stated deadline for requesting it.

Timelines should be in working days and short enough to be real: a notification window, a response window, and a decision window. If your policy cannot name the reviewer role, it is not finished.

Should staff be trained on detector limits?

Yes, and it should be mandatory rather than optional, because inconsistent application is the fastest way to make a policy look arbitrary. The training does not need to be long. It needs to cover four things:

• What a score means. A probability estimate about patterns in text, not a statement about who wrote it.
• Why some writers get flagged more. Non-native English writers, students who write in a formal register, and students working from heavily templated genres all sit closer to the boundary. Why an AI detector flags non-native English writers is a good pre-read for a staff meeting.
• What to do when a score looks wrong. Document, ask, do not accuse. A first conversation is not an allegation.
• Consistency. Two teachers should reach the same decision on the same file. Run a short calibration exercise each term: give staff the same anonymised case and compare conclusions.

There is a counter-intuitive finding from our own measurement worth putting in front of staff. On the same 2026-09-13 run, we counted AI-tell frequency — phrases and constructions commonly associated with machine output — at 4 per 100 words in the human-written sample, versus 1 per 100 words in the unedited model output. Human writers used more of the supposed tells than the model did. Any staff instinct that says “this phrase sounds like AI” is therefore weaker than it feels.

How often should the policy be reviewed?

Annually at minimum, and out of cycle whenever a major tool changes. Assign an owner — a named role, not a committee — and keep a short changelog at the top of the document so staff can see what moved and when.

A practical review checklist:

• Are the disclosure tiers still matching how students actually work?
• Do the detector tools named in the evidence standard still exist in the form you described?
• Did any appeal in the past year reveal a gap in the process?
• Have staff had training this year, and did calibration produce consistent results?
• Are the timelines still realistic given your actual caseload?

Publish the policy where students can find it, and give them the same disclosure scale their teachers have. A rule students cannot read is a rule they cannot follow.

One last thing worth writing into the preamble: the ideas, data, arguments and conclusions in a piece of work must be the student’s own. No policy, and no tool, changes that. What a good policy changes is how fairly you establish whether they are.

---

Measurement note: figures in this article come from our own detector run on 2026-09-13 — n = 23 hand-written guides versus n = 12 unedited model outputs, median AI score 22% and 34% respectively. Method: /methodology/.