What the published research actually says about AI text detection, and what to do when a score has real consequences for someone.
A practical, ordered guide for students and writers: what evidence to preserve first, what to say, what the research supports, and what not to do.
These tools measure register, not authorship. The published false-positive figures, and the four reasons the problem does not get engineered away.
The false-positive arithmetic at institutional scale, why the errors land on specific students, and assessment approaches that do not depend on detection.
Perplexity, burstiness, trained classifiers and watermarking — what each method measures and the specific way each one breaks.
Text-only AI detection does not work reliably, and the reason is structural rather than a temporary engineering gap. Almost every statistical marker anyone has found for machine writing turns out to be a marker of register — how formal, even and conventional the prose is — rather than of who produced it. That means the human writing most likely to be flagged is the most careful: scientific abstracts, formal essays, and work by people writing in a second language.
Three findings worth carrying around:
That includes the tool on this site. The analyzer on the home page has no validated accuracy rate and should not be used as evidence about anyone. It exists to make these statistics visible, not to settle questions.