AI detectors using perplexity disproportionately misclassify non-native English writers, and the mechanism plausibly extends to other low-perplexity writers¶

Claim · All claims
Evidence · 1 study · theoretical r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Liang et al. (2023) provide direct empirical evidence that non-native English writers tend to produce lower-perplexity text, the same statistical feature detectors use to identify AI output. → Craig Wright 2026 q1 i? The paper argues, as an inference from the same mechanism, that writers whose style systematically produces lower perplexity would be expected to be misclassified at elevated rates; this is flagged as a hypothesis requiring empirical testing. → Craig Wright 2026

Evidence¶

Craig Wright 2026¶

Craig Wright. (2026). Transcription is not generation: distinguishing non-generative AI tool use from academic misconduct in higher education assessment. International Journal for Educational Integrity. https://doi.org/10.1007/s40979-026-00234-w

q2 · i? · theoretical · r3

The paper reports, citing Liang et al. (2023), that detectors use perplexity as a primary signal and that non-native English writers 'tend, in the data studied by those authors, to produce text with lower perplexity'. The paper's extension of the mechanism to other populations is presented as a plausible inference, not an established finding.

"Non-native English writers tend, in the data studied by those authors, to produce text with lower perplexity (more predictable word choices, simpler sentence constructions), which is the same statistical feature that detection tools use to identify AI output."

Discussion¶

Related Claims¶

• GPT detectors frequently misclassify non-native English writing as AI-generated, raising fairness concerns for AI policy — possibly the same claim (merge candidate)
• AI-text detectors misclassify authentic TOEFL essays by non-native English speakers at a mean false-positive rate of 61.3% across seven detectors, far more often than native-speaker essays — possibly the same claim (merge candidate)
• AI-detection tools produce inequitable outcomes, disadvantaging non-native English speakers through false positives — a broader claim this one bears on
• The review reports that AI detection tools disproportionately flag multilingual writers, which the authors frame as digital structural violence — a broader claim this one bears on
• Detection-led responses to AI in assessment face validity, fairness and trust limits — a broader claim this one bears on
• Detection-centered academic integrity regimes are unreliable and potentially biased, making sole reliance on AI detectors unsuitable for high-stakes decisions — related
• AI-content detection tools achieve only marginally higher success than human reviewers, around 80% versus 78.4% — related
• Blinded review identified 68% of ChatGPT-generated abstracts as chatbot-produced while 14% of authentic abstracts were misclassified as chatbot-generated — related
• A detector can achieve high aggregate accuracy while producing materially different over-detection rates across author groups — related