How accurate are AI detectors?
By Sourabh Singh18 September 20266 min read
The headline figure and the number that actually matters are not the same number.
Every detector vendor advertises an accuracy figure above 95%. Those figures are usually real, usually measured on unedited model output, and usually irrelevant to the situation you are in.
Why the headline number misleads
Accuracy on clean, unedited GPT output is the easy case. Detectors do well there because that text is maximally low-perplexity and low-burstiness. The hard cases are human writing that looks machine-written, and machine writing that has been edited by a person, and almost every real submission is one of those.
The number that matters
False positive rate on human writing. A detector running at a 1% false positive rate across a 300-student cohort will wrongly flag three students per assignment. Across a semester, that is a policy problem, not a rounding error.
And the false positives are not randomly distributed. They concentrate on non-native English writers, on writers with plain, even prose, and on technical writing where vocabulary is necessarily narrow.
Detectors disagree with each other
The simplest test anyone can run: put one passage through four detectors. Wide disagreement is routine. Since the text has one true origin, at least three of the four are wrong about something, which tells you how much weight a single reading deserves.
What to do with a score
- Treat it as a writing note, never as evidence of origin.
- Look at which sentences drove it rather than the aggregate.
- Cross-check with a second tool, and expect disagreement.
- Keep drafts and version history, process evidence beats any detector.
See the sentences behind a score, not just the number.
Run an explained checkCommon questions
Vendors advertise figures above 95%, but those are usually measured on unedited model output. The hard cases are human writing that looks machine-written and AI writing a person has edited, and almost every real submission is one of those.
Yes, and the false positive rate on human writing is the number that matters. At 1%, a detector would wrongly flag about three students in a 300-student cohort on each assignment.
Non-native English writers, writers with plain, even prose, and technical writers whose vocabulary is necessarily narrow. These false positives are not randomly distributed.
Treat the score as a writing note, not evidence of origin, and look at which sentences drove it. Cross-check with a second tool, expect disagreement, and keep drafts and version history, since process evidence beats any detector.
