Guides

Should teachers use AI detectors? What our test data says

Why one detector is risky and several are worse

A single detector makes mistakes in both directions. Desklib, the most careful one we tested, flagged 8 of 126 human texts. In a class of thirty honest essays that rate would still flag about two.

Running an essay through several free tools does not fix this. Each detector flags different human texts, so every extra tool adds new false alarms. In our test, 117 of the 126 human texts were flagged by at least one of the seven detectors. A teacher who treats any “AI” result as a warning would have doubted almost every honest student.

Some writing is flagged far more often

Detectors react to style, and some kinds of writing look formulaic whoever writes them. Fast-DetectGPT flagged 40.0% of human academic abstracts. Fakespot flagged 70.8% of human news reports. Formal school writing, lab reports and summaries share much of that formulaic style, so they are a likely place for false flags to pile up. Our pages on the best detector for academic writing and on false positives break this down.

Students can get past detectors easily

A light rewrite with a humanizer cut Desklib’s catch rate on Claude Opus 5.5 texts from 62.7% to 32.5%. Some models are also much harder to catch than others: a majority of detectors flagged 38.9% of texts from Claude Fable 5.1. A clean result from a detector does not show that a student wrote the text.

A fairer way to check

Ask for drafts or version history, which most writing tools keep. Talk to the student about the choices in the text. Set some writing in class. If you run a detector, run one, read its false positive rate first, and treat its answer as a reason to ask questions.