Falsely accused of using AI

What the score behind the accusation actually is, how to get an independent measurement, and where the record of your writing process already exists without you having planned for it.

If you wrote it yourself and a detector says otherwise, the first useful thing to know is that this happens at a rate somebody has measured and published. It is not rare, it is not mysterious, and the number is not secret.

The facts below come from the tool vendors themselves, and they hold whichever detector produced your score.

The number is a probability, not a match

The two things on a Turnitin report look similar and are not.

A similarity score matches your text against a corpus. When it flags something there is a source, and you can go and read it.

An AI writing indicator is a prediction. Nothing was matched, there is no source, and there is nothing to click through to. A classifier looked at your words and estimated how likely it is that a model produced them.

Classifiers are wrong at a measurable rate. Turnitin publishes theirs: under 1% of documents, around 4% of sentences. They also say documents scoring below 20% have a higher incidence of false positives, and that the indicator should not be the sole basis for action.

Wrong flags cluster, which explains a pattern people misread

Turnitin's own analysis found 54% of falsely flagged human sentences sit directly next to a sentence scored as AI written. Another 26% sit two sentences away.

Four in five wrong flags land right beside a real detection rather than scattering randomly. If your document has any generated or heavily assisted passage in it, even a short one, the human writing around it becomes markedly more likely to be flagged too.

Some writing is flagged more, and none of it is about you

Detectors measure how regular text is. Several kinds of perfectly ordinary writing are regular:

Non-native English. Learned English gets applied more consistently and reaches for the

common construction more often. Research keeps finding elevated false positive rates here, covered in why non-native English writing gets flagged

Technical and scientific prose. Methods sections are formulaic by design
Anything written to a template. Structured abstracts, lab reports, prescribed formats
Heavily edited work. Editing smooths prose, and smoothing removes exactly the variation

detectors read as human. The more careful you were, the more machine-like it can look

The full mechanism is in why AI detectors flag human writing.

Get a second measurement

Detectors disagree with each other for structural reasons: different training data, different thresholds, different ways of chopping the document up. A second score tells you whether the first reflects something about your text or something about that tool's calibration.

Ours is free and unlimited with no account. Two things make it useful here. It scores each paragraph separately, so you can see whether the signal sits in one section or runs through everything. And we publish our false positive rate: 0.8% at the public setting, 0.4% stricter, measured on pre-LLM academic writing.

If two tools disagree sharply, that usually means the text is genuinely borderline. See why AI detectors disagree.

The record you probably already have

Finished text carries no information about how it was written. That is a limit of the method, not a gap in any one tool, which means the record has to come from somewhere else. Most people have more of it than they realise:

Version history. Google Docs keeps it automatically. Word tracks changes when enabled.

Both show a document growing over time

Earlier files. Outlines, notes, partial drafts, anything with a timestamp on it
What you read. Browser history, reference manager, downloaded PDFs
Correspondence. Supervision notes, feedback, messages about the work while it was in

progress

Your other writing. Earlier pieces establish how your prose normally reads

What a score cannot establish

It describes how text reads. It cannot recover how the text was made.

A score describes how writing reads, which is exactly what an institution's tool is scoring too. That is the argument for knowing what yours says before somebody else runs theirs.

What follows from a score is governed by your institution's own procedures, which vary a great deal and which we are not the right people to describe.