Why the detection approach fails

Over half of non-native English writing samples are misclassified as AI-generated by GPT-based detectors, while accuracy for native English samples stays near-perfect.

This finding, from Liang, Yuksekgonul, Mao, Wu, and Zou, published in Patterns in 2023,11 Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. doi:10.1016/j.patter.2023.100779 is not a footnote. It is the central failure mode of a multi-million dollar industry that educational institutions have adopted without evaluating it first.

The mechanism is specific. Detectors penalize text with low lexical diversity and high syntactic regularity. Non-native writers tend toward simpler, more constrained word choices. Detectors read that regularity as the output of a language model.

The bias is not incidental. It is structural: the same feature that makes a detector confident is the feature non-native writing produces by default.

TWFF takes a different approach. It records what happened during writing, edits, pastes, AI interactions, instead of classifying the finished text. A process log doesn't need to guess who wrote something; it already knows.