Somewhere right now, a student who wrote every word of their essay is sitting outside an office, waiting to explain a number they've never heard of: "your paper came back 82% AI."
Search data says this scene is everywhere. "AI detector false positive" and "accused of using AI didn't" are among the most searched integrity phrases by students and parents, and the volume isn't going down. It's worth being precise about why.
A detector outputs a probability, not a fact
AI detectors don't catch anything. They estimate how statistically predictable a piece of writing is, because AI text tends to be smooth and probable. That's the whole trick. There's no watermark, no fingerprint, no receipt. Just a model guessing about another model.
Which means clear, simple, well-structured human writing gets flagged. The students who write the way we spent a decade teaching them to write, plainly and cleanly, produce exactly the texture detectors are suspicious of.
The false positives don't land randomly
Stanford researchers tested detectors on essays by non-native English speakers and found them flagged as AI at rates above half in some configurations, while the same detectors waved through native speakers' work. Predictable vocabulary and conventional sentence structure look "machine-like" to a probability model, and that describes a lot of perfectly honest writing by multilingual students.
It's worth remembering that OpenAI built its own AI text classifier and shut it down in 2023 for low accuracy. The company with the best possible insight into how its models write concluded it couldn't reliably tell. Detection vendors quote low false positive rates per document, and even taking those numbers at face value, run one percent across a university's essay submissions for a term and you've manufactured hundreds of false accusations a year, concentrated on students least equipped to fight them.
An accusation with no floor under it
Picture the meeting. The score says 82%. The student says "I wrote it." There are no drafts because they wrote it in one sitting, like students do at 1am. The detector vendor's own documentation says scores shouldn't be used as sole evidence. Now what?
Most schools have no answer. The score can't be cross-examined. The process falls back on gut feel, which is exactly the thing the detector was supposed to replace. Meanwhile one real lawsuit, one wrongly failed scholarship student, is an expensive way to learn that "the software said so" isn't a case.
Judge the process, not the vibes of the prose
The way out is to stop interrogating the finished text and start being able to see the work. Text is evidence of nothing. Process is evidence of everything.
That's the principle Themisto EDU is built on. On school devices, it sees actual AI activity as it happens: the session, the tool, whether it was tutoring or answer-seeking, in the moment. When a question comes up, the conversation starts from a record, not from a probability score and a hunch. Honest students get cleared by facts instead of convicted by statistics, and the students who did outsource the essay aren't saved by "the detector must be wrong."
If your integrity process currently rests on a detection score, run our free classroom snapshot to see the scale of what detectors are guessing at. Then ask what your appeals process says when the guess is wrong. Because at scale, it will be.