Study · 2023
GPT detectors are biased against non-native English writers
Liang, Yuksekgonul, Mao, Wu and Zou, Stanford
Seven commercial detectors run over ninety-one TOEFL essays written under exam conditions by non-native English speakers, and eighty-eight essays by American eighth-graders. More than sixty-one percent of the TOEFL essays were classified as machine-written. The eighth-grade essays were identified as human with near-perfect accuracy.
What it changed
Settled that detection was the wrong thing to build against. Break the Test trains the judgment a short structured prompt is designed to surface, rather than trying to beat or evade a classifier that fails hardest on careful writing.
Break the Test →