Flagged as AI When You Wrote Every Word

You wrote it. You remember writing it. You remember the paragraph you rewrote four times and the source you nearly gave up on finding. And now there is an email in your inbox asking you to explain a percentage.

This post is about that situation. Not about how to argue with a number, but about what the number actually is, why writing you genuinely produced can trigger it, and what kind of evidence carries weight when it does.

A score is a probability, not a finding

No detector has access to a record of what a language model has written. It never sees the model, the prompt, or your screen. It measures statistical properties of the words on the page - how predictable they are, how much the sentence rhythm varies, how often certain constructions appear - and converts those measurements into a probability that text with these properties was machine-generated.

That is a real signal. It is also, unavoidably, a statement about populations rather than about you. A detector can tell you that writing like this is usually machine-generated. It cannot tell you that yours was. If you want the mechanics, we wrote them up in detail in How AI Detectors Work.

The practical consequence is the thing worth internalising: a percentage is not evidence of what you did. It is evidence of what your text looks like.

The error rates are worse than most people assume

This is not a fringe complaint. It is documented in the published literature and in the behaviour of the companies who build these tools.

In 2023, Stanford researchers ran seven commercial detectors over 91 TOEFL essays written by non-native English speakers under supervised exam conditions - essays that were, beyond question, human-written. The detectors misclassified them as AI-generated at an average rate of 61.3%. Every one of the seven flagged the same 19.8% of those essays, and 97.8% were flagged by at least one. The same detectors handled essays by US eighth-graders near-perfectly. When the researchers asked a language model to enrich the vocabulary of the TOEFL essays, the misclassification rate fell from 61.3% to 11.6% - meaning the tools were largely measuring linguistic range, and calling the lower end of it artificial. (Liang et al., Patterns, 2023)

OpenAI - a company with more insight into AI-generated text than anyone - launched its own detector in January 2023 and withdrew it that July, citing a low rate of accuracy. It had correctly identified 26% of AI-written text while wrongly flagging 9% of human writing.

Institutions have reached the same conclusion from the other direction. Vanderbilt University disabled Turnitin's AI detector in August 2023, pointing out that it had submitted 75,000 papers the previous year, so even the vendor's claimed 1% false-positive rate implied around 750 papers wrongly flagged. Australian Catholic University logged close to 6,000 misconduct referrals in 2024, roughly 90% of them AI-related; about a quarter were dismissed after investigation, and the university dropped the detector in March 2025, treating any case that rested on the tool alone as unsustainable.

None of that means detectors are worthless. It means a flag is a starting point for a conversation, and you are entitled to have that conversation.

What makes honest writing look artificial

If you have been flagged, it is genuinely useful to know which property of your writing did it - not to change who you are as a writer, but because it usually explains itself once you see it.

Writing in a second language. This is the largest and best-documented fairness problem in the field. Composing in a language you learned tends to produce a narrower vocabulary, more regular sentence structure, and heavier reliance on connectives you were explicitly taught - however, moreover, in addition. Statistically, that is close to a description of the AI signal. The Stanford result above is what this looks like at scale.

The essay structure you were taught. Topic sentence, three developed points, signposted transitions, restated conclusion. That template exists because it makes arguments easy to follow, and following it well produces exactly the uniform, predictable prose a detector scores highly. Doing the assignment correctly can raise your score.

Formulaic genres. Methods sections, lab reports, clinical write-ups, case summaries. These are supposed to be uniform and impersonal. The genre working as intended reads as machine-like.

Short submissions. Every signal a detector uses is a statistic, and statistics need data. Six sentences do not establish a rhythm. Short human writing and short AI writing genuinely resemble each other, which is why Turnitin now requires 300 words before it will score a document at all, and why our own detector applies a much higher bar to short text.

Heavy self-editing. Polishing a draft smooths exactly the irregularities that mark writing as human. Careful revision moves your text toward the machine end of the scale. So does writing with a grammar checker on, and so does translating your own work.

Notice what these have in common: every one of them is either good practice, a taught convention, or a fact about who you are. None is evidence of anything.

Evidence that actually helps you

An assertion that you wrote it is not evidence, and a detector score is not evidence, so a dispute between the two goes nowhere. What resolves these cases is a record of the writing happening.

Version history is the strongest thing you have. If you drafted in Google Docs, open File → Version history → See version history: it shows named revisions with timestamps and, importantly, lets you watch the document grow. Microsoft Word with AutoSave to OneDrive or SharePoint keeps a comparable history under File → Info → Version History. Both show something a finished file cannot: a document that accumulated over days, with false starts, restructured paragraphs, and deletions.

Everything upstream of the draft. Reading notes, annotated PDFs, the outline you abandoned, photos of handwritten planning, library loan records, browser history for the sources. Individually thin, collectively convincing.

Your sources. You can describe an argument you actually read - what it claims, where it is weak, why you chose it over the one next to it. Fabricated citations are one of the more reliable signs of generated work, and the inverse holds: sources you can discuss fluently are hard to fake.

Your earlier work. If the department has your previous submissions, they are a stylistic baseline. Consistency with your own past writing is precisely the comparison Vanderbilt recommends its instructors make.

You cannot generate any of this after the fact, which is the entire reason it is persuasive - and the reason it is worth setting up before you need it.

If it happens to you

Ask what the evidence is, in writing. If the answer is a percentage and nothing else, that is worth establishing early and politely, because most institutional policies - and most detector vendors, in their own documentation - state that a score is not sufficient on its own.

Ask which passages were flagged and read them. Sometimes the answer is immediately obvious: the flagged span is your methods paragraph, or the section you rewrote most.

Ask for the policy and the procedure. You are entitled to know what process you are in, what the timeline is, and what happens next.

Bring your version history to the meeting rather than describing it. Offer to walk through the document's growth.

Do not confess to something you did not do to make the process end. It feels like the fast exit and it is not. If you did use AI in a way the assignment allowed - a grammar checker, translation help, a search tool - say so plainly and specifically. Partial, accurate disclosure is much stronger than a blanket denial that later needs qualifying.

Be concrete about your process. "I wrote it myself" is unfalsifiable and lands as such. "I drafted the second section on the Tuesday after the seminar, and the version history shows me moving it ahead of the case study on the Thursday" is a different kind of statement.

The thing not to do

Do not start rewording your work to push the number down.

It does not address the accusation - the question on the table is who wrote the document, and shuffling the vocabulary of a finished text says nothing about that. It leaves you with worse writing, because the changes that most reliably move a score are the ones that make prose less clear. And if you genuinely wrote the thing, you have just spent your evening degrading the strongest evidence you have: a document whose history matches how you actually work.

We build a detector and we will not help anyone do this. The useful reading of a score has never been am I under the line. It is where do the highlights fall, and why - which is a question about your writing, and occasionally a fair one.

Before the next assignment

Draft in something that keeps version history and leave it on. Keep your notes in the same folder as the essay rather than throwing them away when you are done. If your first language is not English, know in advance that you are in the group these tools treat worst, and keep your process visible accordingly - not because you owe anyone proof, but because being able to produce it costs you nothing and having it takes the argument off the table.

If you want to see what a detector sees in your own writing before anyone else does, ours is at AI Detector. It is free, it needs no account, it shows you which sentences drove the score and which measurements drove each sentence, and it prints its own error rates on every report - because a tool that tells you it is never wrong is telling you something false.