Safe AI help · Updated 1 September 2026

Falsely Accused of Using AI? What to Do

A detector score is a probability, not proof — and the companies that sell detectors say so themselves. Here is how to answer the email, what evidence actually clears you, and how to protect yourself for the rest of the semester.

If you wrote the work yourself and a detector flagged it, do four things: reply calmly and ask in writing what specific evidence the allegation rests on, do not admit to anything you did not do, gather your process evidence before you touch the file again — version history, drafts, notes, sources — and ask for your institution's academic misconduct procedure so you know your deadlines and your right to appeal. The burden of proof is on the institution, not on you, and a detector percentage on its own does not meet it.

Why honest work gets flagged

AI writing detectors do not find evidence of AI use. They score how statistically predictable your text looks, and predictable prose is something plenty of humans write — especially students writing in a second language, students trained to write formally, and anyone writing to a rigid rubric.

The scale of that problem is documented. A Stanford study published in Patterns, GPT detectors are biased against non-native English writers, ran seven detectors over 91 TOEFL essays written by non-native English speakers and recorded an average false positive rate of 61.22%. All seven detectors unanimously flagged 18 of those 91 essays as AI-authored. On US eighth-grade essays by native speakers, the same tools averaged about 5.19% false positives. The researchers put the gap down to lower perplexity — less linguistic variety — in non-native writing. Note the limit of that study: Turnitin was not one of the seven tools tested.

Turnitin has published figures of its own. Its chief product officer told Inside Higher Ed in June 2023 that the sentence-level false positive rate is around 4%, having claimed under 1% at the document level when the tool launched that April. Turnitin now flags any score between 0 and 20% with an asterisk, because results in that band are less reliable.

Small percentages stop being small at scale. When Vanderbilt disabled Turnitin's AI detector in August 2023, it did the arithmetic publicly: at its submission volume, a 1% false positive rate meant "around 750 student papers could have been incorrectly labeled as having some of it written by AI" in a single year. Vanderbilt also objected that "Turnitin gives no detailed information as to how it determines if a piece of writing is AI-generated or not."

What the detector vendor and universities actually say

This is the part most students do not know, and it is the strongest card in your hand.

You are not arguing that detectors are bad and should be ignored. You are arguing something narrower and much harder to refuse: a score is a flag for investigation, and an investigation needs evidence. Ask what the evidence is.

The first 48 hours

1. Do not edit or delete anything. Not the document, not your notes, not your browser history. Your version history is the evidence; a tidy-up looks like a cover-up.
2. Reply promptly and neutrally. Confirm you received the email, say you wrote the work yourself, ask for the specific allegation, the evidence behind it, and a copy of the procedure. Do not argue the case by email.
3. Do not admit to "a bit of AI" to make it go away. Students do this constantly under pressure, and it converts an unproven flag into a confession. If you did use AI in a permitted way — grammar check, brainstorming — describe exactly what you did, accurately, and no more.
4. Export your version history now. In Google Docs: File → Version history → See version history. In Word, check File → Info → Version History or your OneDrive/SharePoint file history. Screenshot the timeline and download the earliest drafts.
5. Find out who is on your side. Students' union or student advocacy service, a disability or international student adviser if relevant, your personal tutor. Many institutions let you bring someone to the meeting. Ask whether you can.

The evidence that actually works

Arguments about style rarely land. Process evidence does, because it is dated and it is boring.

• Version history. The single strongest item. A human document grows over multiple sessions with deletions, reorderings and false starts. Show the timeline, not just the final file.
• Earlier drafts and outlines, ideally emailed to yourself or saved with timestamps.
• Handwritten notes, annotated readings, whiteboard photos. Photograph or scan them.
• Source trail. Library loan records, database access logs, browser history, the PDFs in your downloads folder. If your citations are real and you can show where you found them, you are already ahead of most genuine AI submissions, which tend to cite things that do not exist.
• Prior graded work from the same course. Melbourne's guidance explicitly points staff toward comparing a student's other work — so use it. If you always wrote like this and nobody flagged it before, that matters.
• Your ability to explain it. Expect to be asked how you developed the argument, which sources you used, and why you concluded what you did. This is the test most false accusations fall apart on, because you can answer and a cheat cannot.

Prompts to prepare — used honestly

There is an obvious irony in using AI here, so be precise about what it is for: organising facts you already have and rehearsing questions you will be asked. Never ask it to invent a writing process, produce drafts to pass off as earlier versions, or write your appeal statement for you. Fabricated evidence turns a survivable misunderstanding into an unsurvivable one — and your statement needs to sound like the person who wrote the essay.

1. Build a timeline from what you actually have. Paste in your real version-history timestamps and note titles.

Below are the real timestamps from my document version history, plus a list of
notes and sources I have.

Organise them into a plain chronological table: date, what I did, what evidence
proves it. Use only what I have given you - do not add, infer or embellish any
step. Where an entry is thin, say "weak evidence" instead of filling the gap.

Then list what a reasonable reviewer would still want to see, so I can check
whether I have it.

2. Rehearse the meeting. This is the highest-value one, because the meeting is where it is usually decided.

Act as an academic integrity panel member. I wrote this essay myself and have
been flagged by an AI detector.

Ask me one question at a time about my argument, my sources and how I wrote it -
the kind of questions used to test whether a student really understands their own
work. Wait for my answer before the next question.

After ten questions, tell me which of my answers were vague or unconvincing and
exactly why, so I can prepare better. Do not write my answers for me.

3. Pressure-test your written statement. Write it yourself first, in your own words.

Here is a statement I have written for my academic misconduct meeting.

Do not rewrite it. Tell me:
- anything that reads as an admission I did not intend
- any claim I have not backed with evidence
- anywhere I sound defensive or sarcastic rather than factual
- anything missing that the procedure I have pasted below asks for

Quote the exact phrases you are flagging.

If you want the checking habit that stops most of these problems earlier, our guide to verifying AI answers before you study them and the citation audit checklist cover the same discipline applied to sources.

Three things not to do

• Don't run your work through consumer AI detectors. They are not the tool your institution used, so a clean score proves nothing to anyone, and a bad score will panic you into rewriting honest sentences.
• Don't use a "humanizer" or bypass tool. Deliberately disguising how a piece was produced is treated as misconduct in its own right under many policies. It also destroys the version-history story you would otherwise have.
• Don't let the deadline pass. Appeal windows are short and strictly enforced, and "I was waiting to hear back" is not usually accepted as grounds. Diarise every date the moment you get the procedure.

Protecting yourself for the rest of the semester

Ten minutes of habit makes this whole scenario answerable in advance.

1. Write in one cloud document per assignment, from the outline onwards, in Google Docs or Word online. Never draft somewhere else and paste the finished thing in — a single large paste with no history is the pattern that looks worst.
2. Keep the outline and the notes. Don't delete the messy version once the clean one exists.
3. Log your sources as you read, not at the end.
4. Know your course's policy in writing before week three. Our class AI policy checklist and the email template for asking a professor what's allowed take about five minutes each, and an answer in writing is worth a great deal later. If you are also applying to grad, law or med school this year, the rules there are stricter and separate — see what applications actually allow on a personal statement, where the binding text is the certification you tick to submit.
5. Disclose permitted AI use as you go. If your course allows AI for brainstorming or proofreading, say so in a short note with the submission. A student who documented their use is a much harder person to accuse — the AI disclosure statement guide has the wording.

FAQ

Can an AI detector score alone prove I cheated?

No, and the vendor agrees. Turnitin's own guidance says its AI writing indicator does not make a determination of misconduct, that the percentage should not be used as the sole basis for action, and that the final decision rests with the instructor. Several universities have written that into policy: the University of Melbourne's academic integrity guidance states an AI writing report alone is not a sufficient basis for an allegation and that a second piece of evidence is required. If your institution is treating a number as the whole case, that's the first thing to challenge.

How accurate are AI detectors really?

It depends heavily on the tool and the writer. The Stanford Patterns study ran seven GPT detectors over 91 TOEFL essays written by non-native English speakers and found an average false positive rate of 61.22%, with all seven unanimously flagging 18 of the 91. The same detectors were near-perfect on US eighth-grade essays, averaging about 5.19% false positives. Turnitin was not among the seven tools tested. Turnitin has separately said its sentence-level false positive rate is around 4%, and it marks scores between 0 and 20% with an asterisk because those results are less reliable.

What evidence proves I wrote my own essay?

Process evidence, not stylistic argument. The strongest items are document version history showing the piece being built over multiple sessions, earlier drafts and outlines with timestamps, handwritten or annotated notes, your browser and library history for the sources you cited, and previous graded work from the same course. You should also be ready to explain your argument out loud: how you built it, which sources you used, and why you reached that conclusion.

Are non-native English speakers more likely to be flagged?

The published evidence points that way, and it's a legitimate thing to raise. The Stanford study found detectors misclassified non-native English writing as AI-generated far more often than native writing. Vanderbilt cited that bias among its reasons for disabling Turnitin's AI detector. In the UK, the Office of the Independent Adjudicator upheld part of a student complaint partly because the provider had not considered whether Turnitin's AI detection might be less reliable for non-native English speakers, which was relevant given the student's international status.

Should I run my work through an AI detector before submitting it?

Usually a waste of time and sometimes actively harmful. Consumer detectors aren't the model your institution uses, so a clean result proves nothing, and chasing a lower score pushes you to rewrite honest sentences into worse ones. Avoid humanizer or bypass tools entirely: many policies treat deliberately disguising authorship as misconduct in its own right. Keep your drafts instead — version history is worth more than any pre-check.

Bottom line

A flag is the start of a conversation, not the end of one. Answer it calmly, in writing, with dated evidence of how the work came into existence, and make the institution show what the allegation rests on beyond a percentage — which its own detector vendor says was never meant to carry that weight. Then spend ten minutes setting up the habits above, so that if it happens again you can answer it in an afternoon.