The Detector That Accuses Non-Native Speakers

61.22%

Average false-positive rate of seven AI detectors on TOEFL essays by non-native speakers — all 91 essays were written by humans (Liang et al. 2023).

A tool built to catch deception becomes the case file. Meridian — the lab’s research collective — audited AI text detectors against their own spec sheets. The finding fits on a coaster: the strongest guarantee holds where the stakes are lowest — and inverts where they’re highest.

Promised (GPTZero spec sheet, false-positive rate)

0.24%

Measured (Ibrahim et al. 2023, real student submissions)

18%

Vendor claim vs. independent measurement (GPTZero, document-level FPR). The 75× gap is the finding — and not an outlier: across seven detectors, the average false-positive rate on TOEFL essays by non-native speakers was 61.22% (Liang et al. 2023).

Calibrated on clean corpora. Deployed against people who write English as a second language. The difference has a name: calibration gap.

89 of 91 TOEFL essays were flagged as AI by at least one detector. All 91 were written by humans.

Australian Catholic University: roughly 6,000 misconduct cases in 2024, one in four dismissed, AI detection scrapped in March 2025. Evidence you have to scrap was never evidence.

// Sources

Liang et al. 2023 — GPT detectors are biased against non-native English writers (Cell Patterns) accessed 2026-07-11
Dugan et al. 2024 — RAID Benchmark (ACL) accessed 2026-07-11
Pratama 2025 — AI detector over-detection of non-native writing (PeerJ CS) accessed 2026-07-11
Meridian — Instrument 001 "Calibration Certificate" (source work, incl. correction history) accessed 2026-07-11