Why AI detectors are unreliable

Home / AI literacy guides / Integrity and AI

AI literacy guide · Integrity and AI

Why AI detectors are unreliable

The evidence on detection tools, and why they cannot anchor your integrity plan.

It is tempting to want a tool that flags AI writing. The research does not support trusting one. In testing of many commercial detectors, none reached high accuracy, and all produced both false positives, flagging human writing as AI, and false negatives, missing AI writing. Detectors can be defeated by light editing, and they offer no proof, only a probability.

The harm is not evenly distributed. Stanford researchers found that detectors misclassified a large share of essays by non-native English speakers as AI-generated, in some studies well over half, because simpler, more predictable wording reads as machine-like to the tool. A plan that leans on detection therefore risks falsely accusing exactly the students who are already most vulnerable. Several universities have declined to adopt these tools for this reason.

What the testing found
Accuracy In one review of 14 tools, none reached 80 percent accuracy.
Bias Detectors misclassified more than 60 percent of some non-native English essays as AI.
Conclusion A detector score is a guess, not evidence of misconduct.

Accessibility note

A detection-first plan falls hardest on multilingual and first-generation students, the opposite of a fair assessment.