Falsely accused of using AI, when you did write it

What to do in the first hour

Preserve the evidence before you reply to anyone, because the most persuasive evidence is the one most easily destroyed by editing the file.

• Stop editing the document. Every further save can merge or push out the revision history that proves you wrote it.
• Open the version history and export or screenshot it, with timestamps visible, before anything changes.
• Collect the untidy material — outlines, notes, half-finished paragraphs, the sources you actually read. Untidy helps; a clean single draft is what an AI-generated file looks like.
• Ask, in writing, which tool produced the flag, what score it returned, and what your institution's policy says a score triggers. You are entitled to know what the allegation rests on.
• Do not rewrite or run the text through any rewriting tool to "fix" the score. It changes the file, destroys the timeline, and looks exactly like a cover-up.

Version history is your best evidence, and it has limits worth knowing

Both Google and Microsoft document what their version history keeps, and neither keeps everything. Knowing the limits before you cite it stops the other side using them against you.

• Google Docs — history is automatic, and Google's own support page states that revisions for a file "may occasionally be merged", which is why a long writing session can collapse into fewer entries than you remember. You can pin up to 40 named versions per document, so name the state you are in now before it merges.
• Microsoft 365 — version history only works for files stored in OneDrive or SharePoint. A document written in Word on your own machine and emailed in has no version history at all. On a personal Microsoft account you can retrieve the last 25 versions.
• Either way, what convinces is the shape of the record: text appearing in fragments, sentences being rewritten, paragraphs moved. A single paste of 900 finished words is the pattern that reads badly, whoever wrote them.
• If your work was written somewhere without history — Notes, a local text file, on paper — say so plainly and lead with the other evidence rather than being caught short later.

What the published research says about false positives

The strongest peer-reviewed finding is that these tools misclassify non-native English writing at a rate that would be unacceptable in any other kind of test.

Liang, Yuksekgonul, Mao, Wu and Zou (Stanford, published in Patterns; read here from arXiv:2304.02819v3, 10 July 2023) ran seven widely-used GPT detectors over 91 human-written TOEFL essays and 88 US eighth-grade essays. On the eighth-grade essays the detectors were near-perfect, averaging 5.19% false positives. On the TOEFL essays written by non-native English speakers, the average false positive rate was 61.22%: all seven detectors unanimously flagged 18 of the 91 essays, and 89 of the 91 — 97.80% — were flagged as AI-generated by at least one detector. The authors' explanation is mechanical rather than moral. Perplexity-based detectors score predictable, low-variety wording as machine-written, and a writer with a smaller working vocabulary in English produces exactly that. The same paper shows the effect in reverse: rewording the American eighth-graders' essays with the prompt "Simplify word choices as if written by a non-native speaker" pushed their misclassification rate from 5.19% to 56.65%. Two limits are the authors' own — the samples are small, and most detectors tested were built on GPT-2 — but the direction of the bias has been replicated widely enough to be worth citing in an appeal.

Institutions and vendors that stopped trusting detection

You are not making a fringe argument. A major university switched its detector off, and the company that built ChatGPT withdrew its own detector for being inaccurate.

• Vanderbilt University disabled Turnitin's AI detection tool in August 2023 and published its reasoning. Its own arithmetic is the useful part: Vanderbilt submitted 75,000 papers to Turnitin in 2022, so at the 1% false positive rate Turnitin claimed at launch, "around 750 student papers could have been incorrectly labeled". The post also objects that Turnitin gives no detailed account of how the determination is made. The live page now sits behind credentials; we read it from the Internet Archive snapshot of 22 December 2024.
• OpenAI took its own AI text classifier offline on 20 July 2023, and the note still on the launch page says it was withdrawn "due to its low rate of accuracy". Its published figures at launch were 26% of AI text correctly identified and 9% of human text incorrectly flagged, with the instruction that it "should not be used as a primary decision-making tool".
• Turnitin's own FAQ, read 17 August 2026, states that the percentage should not be used as the sole basis for action, hides scores between 1% and 19% behind an asterisk because they are unreliable, requires 300 words of prose before it will score at all, and claims its under-1% false positive figure only for documents scored above 20% AI writing.
• Our own test of two free detectors on twenty texts of known origin found they gave opposite verdicts on half the samples, and flagged Frederick Douglass's 1845 narrative at 99.9% AI.

How to write the reply

Short, factual, and organised around your evidence rather than around the tool's unreliability. Lead with the record, use the research as support, and ask for the process your institution actually documents.

A reply that opens by attacking the detector reads as a defence lawyer's move. A reply that opens with "here is the version history of the document, exported today, showing 40 minutes of drafting on the 12th and revisions on the 14th" changes what is being discussed. Attach the export. Name the sources you read. Offer to discuss the argument of the piece in person, which is the test no detector can fake for you and the one most instructors accept quickly. Then, and only then, note that the tool's own vendor says the score is not a determination, cite the false positive research if it applies to you, and ask what the formal appeal route is. Keep it to one page. If your institution has a student advocate, union representative or ombudsperson, involve them before a meeting rather than after.

If your poem, or anything short and patterned, was flagged

Short, regular, compressed writing is what these models score as machine-made. Poetry is the worst case, and a high score on a poem carries almost no information.

In our own test, Emily Dickinson's published poems scored 99.8% AI on one detector. Detectors work on statistical predictability, and verse is deliberately patterned — regular line length, repeated structures, compressed phrasing — so it scores the way models score. The same applies to lab reports, structured abstracts, exam-style answers and anything written to a template. Turnitin declines to score anything under 300 words for related reasons, and OpenAI's classifier was documented as "very unreliable on short texts (below 1,000 characters)". If the flagged work is short, say so and ask what the tool's own documented minimum is.

If English is not your first language

Say so, in the appeal, and cite the Stanford finding by name. It is the single most powerful piece of external evidence available to you.

A 61.22% average false positive rate on non-native English essays, against 5.19% on native-speaker essays from the same study, is not a rounding error — it is the tool failing at the exact task it is being trusted with. Ask whether your institution has considered that disparity in its policy, and ask for the assessment to be judged on evidence that does not have a documented bias against writers like you. Nothing about this argument requires you to concede anything about your own work.

Making yourself provable next time

The habits that produce an evidence trail cost nothing and are worth adopting whether or not you are ever accused again.

• Draft in Google Docs, or in Word with the file in OneDrive, so history exists without you thinking about it.
• Type or dictate into the working document rather than pasting finished blocks from elsewhere. Where you must paste your own earlier text, do it in stages.
• Keep the outline and the notes in the same document, above or below the draft, instead of deleting them when you are done.
• Name a version at the end of each session. Google allows 40 per document, which is more sessions than most assignments need.
• Save the reading list — browser history, library loans, PDFs with your highlights.

Disclosure

We make an app that scores text for AI-likeness and rewrites it, so read this page knowing that. It argues that scores of this kind, including the one our own app reports, are weak evidence about a person.

Two things follow from that, and we would rather state them than have them found. First, our app cannot clear you: no third-party score, ours included, is evidence that you wrote something, and a low score from us would not be accepted by anyone assessing you. Second, if you are currently under investigation, running the disputed text through any rewriting tool — ours or a competitor's — is the wrong move, because it alters the file whose history is your defence and may itself breach your institution's policy. The honest use for a tool like ours is on your own drafts, in your own time, with disclosure where your institution requires it.