Log In

AI Detection False Positives

False positives are the most consequential failure mode of AI detection: original human work flagged as AI-generated, with real career-affecting consequences for students. The teachers on this forum have spent months documenting how often this happens, who it happens to most, and how to respond when a detection score points at a student who did the work themselves. Threads here cover ESL false positives, formulaic-writing false positives, the academic policy implications, and what a defensible response process looks like.

18 discussions · 98 replies · 17 participants

Community Discussions

Sorted by engagement. Most-voted and most-replied first.

Community Wisdom

Questions teachers actually ask, with answers distilled from real discussions.

How often do AI detectors produce false positives?
Vendors typically advertise false-positive rates between 1 and 3 percent. Community testing here, especially on ESL student work and formulaic academic writing, documents practical false-positive rates of 5 to 15 percent. On test-prep writing that follows a templated structure (intro-thesis-three-points-conclusion), false-positive rates climb higher still.
Why does GPTZero say everything is AI when it is not?
Two common causes show up in threads here. First, AI-likelihood scoring penalizes predictable phrasing, which is what students are taught to produce in structured writing instruction. Second, students writing in English as an additional language often produce text with limited vocabulary variation that the detector reads as low-perplexity AI output. Neither is evidence of AI use.
What should I do if an AI detector is wrong about my student?
The defensible path documented here is process-based. Ask the student to walk you through how they wrote the piece. Look at draft history, version timestamps, and any process artifacts. Compare voice and reasoning style against work the student produced in class. A high detection score plus a strong process trail almost always points to a false positive.
Do AI detectors flag ESL students more often?
Yes, and this is the most-documented bias in our community testing. Formulaic phrasing, limited vocabulary, and stilted syntax all trigger AI-likelihood signals that have nothing to do with AI use. Several threads here track this longitudinally with practical false-positive rates of 5 to 15 percent on original ESL student work.
Can I appeal an AI detection result?
Most Canadian institutions allow some form of appeal, though the formal process varies sharply. The strongest appeals documented here combine three elements: documented drafts and version history, a brief written reflection on the student's process, and a request for an oral defense of the work. Detection scores alone almost never survive a properly documented appeal.

Related Discussion Hubs

AI Detection AI for ESL Students Turnitin AI Detection GPTZero