There is no single accuracy number that travels across tools, languages and document types. A detector that scores well on long English essays can fail on short CVs, second-language writing or heavily edited drafts. When a vendor quotes "99% accuracy" without a false-positive rate and a described corpus, treat it as marketing.
The number that matters when a person is being accused is the false-positive rate: how often human writing is called AI. IA Checker publishes a dated regression run on the AI detector page, with corpus size and limits in the same block as the figures. It is small on purpose — a break-glass check between model versions, not a global accuracy claim.
Use any score as an indicator. Open sentence-level evidence, read the text, and keep a person in the decision.
