Stanford's Detector-Bias Audit — GPT Detectors Flag Non-Native English Essays as AI-Written

Researchers at Stanford University (James Zou's group) tested seven widely used GPT detectors on 91 human-written TOEFL essays by non-native English speakers and on US 8th-grade essays by native speakers. The detectors classified the native essays almost perfectly but wrongly labelled more than half of the TOEFL essays as AI-generated; one detector flagged about 98%. About 19% of the TOEFL essays were unanimously misclassified by all seven detectors, and 97% were flagged by at least one.
The authors trace the bias to perplexity-based scoring: simpler vocabulary and syntax, common among non-native writers, looks 'machine-like'. They also showed that prompting ChatGPT to enrich the wording of essays let AI text evade the detectors, so the tools are both biased and easy to fool.
Published in Patterns (Cell Press, July 2023), the study advises extreme caution against using detectors in classrooms, admissions or hiring. Limits: a modest TOEFL sample and detectors available in 2023; it does not test newer tools or measure learning outcomes. It has since underpinned institutional decisions to disable AI-detection scores and redesign assessment instead.

Read the full analysis: https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7