AI detection false positives: what they are and what to do

Why they happen

Most detectors are classifiers scoring predictability. Text where each word follows naturally from the last scores as machine-like, because that is what fluent generation looks like.

Unfortunately that is also what good editing looks like. Clear, plain, heavily revised prose converges on the same statistical territory, so the writers most likely to be flagged include careful editors, technical writers working to a house style, and people writing in an additional language.

Base rates make it worse. Even a detector that is accurate most of the time, run across thousands of submissions of which few are actually AI-written, produces a large number of wrong flags relative to right ones. This is ordinary conditional probability, and it is the single most useful thing to explain to a panel.

What to do, in order

Ask which tool produced the figure and what its published false positive rate is. Ask what the institution’s policy says a detector score is sufficient to establish. Very often, on its own, nothing.

Assemble process evidence: drafts, version history, notes, search history, supervision correspondence. This is more persuasive than any detector output in either direction.

Run a different kind of test. A provenance-mark check asks whether a deliberate statistical signature is present, rather than whether the writing reads a certain way, so it is not vulnerable to the same failure mode.

Keep it procedural in tone. You are not accusing anyone of bad faith; you are asking what the evidence is and what the policy requires.

What not to do

Do not rewrite the document to score better. It concedes the premise, it destroys the version the accusation was about, and tools that promise to do it for you are evasion services with a marketing department.

Do not rely on a single clean result as proof. Absence of a mark is not proof of human authorship, and claiming otherwise hands the other side an easy correction.