Your work was flagged as AI and it wasn't. What now?

You are not an unlucky outlier

False positives are a known, measured property of every AI detector in use. Published evaluations have repeatedly found that non-native English writers are flagged at dramatically higher rates than native speakers — in one Stanford study, more than half of TOEFL essays were classified as AI-generated, against almost none of the native-speaker samples. Other groups get caught the same way: people who write in a plain, structured register; people in technical fields where phrasing is conventional; anyone who uses Grammarly or a similar tool, which smooths precisely the variation detectors read as human. This matters for how you respond. You are not arguing that a reliable system made a rare mistake. You are pointing out that the system has a documented error rate concentrated in your situation.

Do this first, before replying

Resist the urge to send an immediate defensive message. Spend twenty minutes gathering, because the evidence gets harder to assemble the longer you wait.

1. 1Open the document's version history — Google Docs, Word online and most editors keep it, often for months
2. 2Export or screenshot that history showing incremental edits with timestamps
3. 3Collect your notes, outlines, sources, drafts and any research tabs or annotations
4. 4Find earlier pieces you wrote on similar material, to establish your normal voice
5. 5Write down what you remember about composing it: where, when, in what order, what you changed

Version history is the strongest evidence you have

AI-generated text appears in a document as a paste: a large block arriving intact, at once. Human writing accumulates — sentences added and reordered, paragraphs rewritten, a section moved, typos introduced and fixed. Google Docs version history shows that trajectory directly, and it is very hard to argue with. Microsoft Word's AutoSave history and most modern editors do the same. If you drafted in something that keeps no history, look for what does exist: email drafts, message threads where you discussed the piece, notes-app captures with timestamps. This single artefact resolves most disputes. It is also the reason the next section exists.

How to frame the conversation

Tone determines outcome more than evidence does. Two framings to avoid: an emotional protest, and an aggressive attack on the institution's tools. Both make the other person defensive, and a defensive reviewer looks for reasons to hold their position. What works is calm, specific and forward-looking. Acknowledge the flag without accepting the conclusion. Offer your process evidence unprompted. Cite the known limitation neutrally and once — that detectors have documented false-positive rates that fall disproportionately on writers in your situation — without turning it into the centrepiece. Then offer something concrete: walk them through your argument, discuss your sources, or write a comparable piece under supervision. Someone genuinely trying to establish the truth will take one of those. Someone who refuses every verification route is telling you the score was a pretext, which is useful to know.

Ask what the number actually was

Ask which tool was used, what score it returned, and what that institution's threshold is. You are entitled to know what you are answering. Often the answer is softer than the accusation implied — a 'mixed signals' mid-range result presented as a finding. Sometimes the tool flagged only a section, or the submission was short enough that the vendor's own documentation says the result is unreliable. Vendor documentation is public, and quoting a tool's own stated limitations back is more persuasive than anything you can say about yourself.

If the first conversation fails

Escalate along the documented path rather than relitigating with the same person. Most institutions have an academic-integrity appeals process with defined steps, and most have a policy document specifying how AI-detection evidence may be used — frequently that it cannot be the sole basis for a finding. Find that document. An appeal that cites the institution's own policy is a different conversation from one that asserts innocence. Keep everything in writing from this point, and bring your process evidence, the tool's stated limitations, and the policy language to the panel.

Protecting yourself going forward

The uncomfortable truth is that defending against this is mostly preparation, not argument. Draft in an editor with version history, always. Keep your outlines and research notes rather than deleting them when you are done. Where a piece matters, do a quick self-check before you submit — if your own draft reads as high likelihood, that is worth knowing in advance, and it lets you decide whether to revise or simply to have your version history ready. If you use AI assistance at any stage, disclose it according to whatever rule applies to you. Disclosed assistance is a policy question with a clear answer. Undisclosed assistance discovered later is a credibility problem you cannot argue your way out of.

A note on the wider situation

Detection is unlikely to get decisively better, because it is an adversarial problem: every improvement in detection is training signal for the next generation of models. Institutions are slowly recognising this and shifting toward assessment designs that make the question less central — supervised writing, oral defence, process portfolios. Until that shift finishes, the practical advice is unglamorous and it works: keep your drafts, keep your notes, and stay calm when a number says something false about you.