The Teacher’s Checklist Before Filing an Academic Integrity Report Over AI

Before reporting AI cheating, run one checklist: confirm you have evidence beyond a detector score, check whether the student fits a known false-positive profile, review draft history, compare authenticated writing samples, talk to the student, verify the policy actually covers the conduct, and ask whether your file would convince a neutral panel. If the case survives all of that, file it with confidence. If it doesn’t, the checklist just saved a student from a wrongful accusation — and saved you from a hearing you’d lose.

An integrity report is the most consequential routine document a teacher files. It can shadow a transcript, a visa, a scholarship. The bar for filing should feel high, because for the student on the other end, the stakes are. Here’s the checklist, in the order that wastes the least of your time.

Key takeaways

• A detector score alone is never sufficient grounds to file — vendors themselves say so, and panels increasingly agree.
• The strongest evidence is process evidence: version history, writing comparisons, and the student’s ability to discuss their own work.
• ESL students, formal writers, and short texts trigger false positives at rates that should change your evidence bar.
• Talk to the student before filing unless policy forbids it. It’s the most informative single step, in both directions.
• File a documented timeline, not a verdict. You’ll defend the process, not the software.

The checklist before reporting AI cheating

1. Is the score your only evidence?

Start with the brutal question. If everything traces back to one percentage — or two percentages from two tools — you don’t yet have a case. Turnitin’s own guidance says its AI indicator shouldn’t be the sole basis for action against a student. OpenAI retired its own AI classifier in 2023 because it wasn’t accurate enough. A number can legitimately start an inquiry. It cannot end one, and a score alone can’t carry the burden of proof in any hearing that’s paying attention.

2. Does the student fit a false-positive profile?

Three profiles account for most wrongful flags: non-native English speakers (Stanford researchers found detectors flagged over half of ESL essays as AI), students with very formal or formulaic styles, and short submissions where detectors have too little text to judge. A flagged student in any of these groups doesn’t get a pass — but the flag itself becomes weak evidence, and the rest of your file has to do more work. If you want to feel this failure mode firsthand, see what a detector actually flags by running clean formal prose through one and watching the heatmap.

3. What does the draft history say?

Version history is the closest thing this domain has to direct evidence. Open the Google Doc or Word history and look at how the text came into being. Steady accumulation with edits, deletions, and rewrites over days reads one way. Eight hundred words arriving in a single paste at 11:40 p.m. reads another. No version history available? That’s a gap in your case, not a fact against the student — and a reason to require process evidence on future assignments.

4. How does it compare to authenticated writing?

Put the flagged essay next to work you know the student produced — in-class writing, exams, discussion posts. You’re looking for more than “this is better.” Students write better with time and revision; that’s the point of teaching. You’re looking for a different writer: new syntactic habits, vocabulary from another register, a voice that doesn’t connect to anything they’ve produced under observation.

5. Have you talked to the student?

Unless your institution requires going straight to the office, have the conversation first — private, calm, and framed around the work rather than the accusation. Ask them to walk you through a paragraph, explain a word choice, describe how they found a source. Students who wrote the essay can talk about it from the inside, even nervously. Students who didn’t tend to describe it from the outside. We’ve published scripts for talking with a student you suspect that keep this humane. Document what was said, whatever direction it points.

6. Does the policy actually prohibit what happened?

Reread the specific language before filing. Many syllabi still say nothing explicit about AI, or say something ambiguous (“outside assistance”). If the student used a grammar tool on a course with no stated AI rule, you may have a policy-clarity problem rather than a misconduct case. Panels notice this gap, and so do appeals.

7. Is your documentation dated and complete?

Every step above should exist in writing with dates: what prompted the concern, what you checked, what you found, what the student said. If your notes live only in memory, build the file now — our guide on how to document an AI concern step by step covers the format that holds up.

8. Would this convince a neutral stranger?

The final gate. Imagine a panelist who doesn’t know you, doesn’t trust software, and starts from the presumption the student is honest. Does your file move that person? If yes, file — clearly and without apology, because real cases deserve real consequences. If you’re honestly unsure, the file is telling you something.

After the checklist

Whichever way it resolves, close the loop. If you file, write the report as a timeline of documented steps with the detector positioned as a screening signal, not a star witness. If you don’t file, tell the student the matter is closed — an unresolved cloud is its own quiet punishment — and tighten the process evidence on future assignments so the next case is clearer in either direction. More case-handling guides live in the full library of educator guides.

The checklist isn’t about making reports harder to file. It’s about making the ones you do file impossible to dismiss.

Frequently asked questions

Is a detector score alone enough to file an integrity report?

No. Detector output is a probability estimate, and vendors themselves say it shouldn’t be the sole basis for action — Turnitin’s guidance is explicit about that, and OpenAI retired its own classifier over accuracy. Panels increasingly know this, so a score-only report is likely to fail while still putting a student through a damaging process. A score justifies a closer look at process evidence; only that evidence justifies a report.

What evidence should I gather before reporting suspected AI cheating?

Four things, in rough order of strength: version history or draft evidence showing how the text came into being; a comparison with the student’s authenticated writing, like in-class work; notes from a conversation where the student could or couldn’t discuss their own essay; and a dated record of each step you took. If after gathering all four you have only the detector score, that’s your answer about whether to file.

Do I have to talk to the student before filing a report?

Unless your institution forbids pre-report contact, yes — it’s the single most informative step. A short, private, non-accusatory conversation about the work itself resolves many cases in both directions: students who wrote the essay can discuss it from the inside, and students who didn’t often can’t explain their own vocabulary or structure. Skipping the conversation means filing on inference when direct evidence was one meeting away.

What if the student fits a known false-positive profile?

Raise your evidence bar substantially. Research led by Stanford’s Weixin Liang found detectors flagged over half of essays by non-native English speakers, and formal, formulaic, or very short writing misfires the same way. A flag on a student in these groups is weak evidence by definition, so the case must rest almost entirely on process evidence — and if that evidence isn’t there, the responsible move is not to file.

What should the report actually contain if I do file?

A timeline of documented steps, not a verdict: what prompted the concern, what the version history and writing comparison showed, what happened in the student conversation, and copies of the relevant policy language. Present any detector result as one screening signal among several, with its limitations stated. Reports built this way protect everyone — including you, because you’ll be asked to defend the process, not the software.