AI Detection Accuracy by Language: Why Detectors Fail Outside English (2026 Data)

1The English-vs-non-English accuracy gap

Every detector you have heard of was trained and tuned on English first, and the numbers show it: accuracy falls 10 to 20 points the moment the text is in another language.

Copyleaks — the tool with the widest language support — scores about 91% on English with a 7.2% false-positive rate, then averages roughly 74% across non-English content. An independent 300-sample test tells the same story by language, and the ranking is consistent with how much training data each language has.

That gap is not a rounding error — it is the difference between a tool you can lean on and a coin flip dressed up as a score. The mechanism is the same one that makes detectors misjudge short essays: they measure how statistically predictable text is, and they have simply seen far less non-English writing to calibrate against. It is worth remembering what an AI-detection score actually represents before treating a non-English result as evidence of anything.

For a multilingual classroom or newsroom the practical rule is blunt: a non-English AI score is a reason to look closer, not a verdict — and the further a language sits from English in the training data, the less that number is worth.

2How few languages detectors actually support

Before accuracy even enters the picture, most detectors simply do not claim to handle your language at all.

This matters because coverage and accuracy compound. Turnitin is the default at most universities, yet its AI checker is the most English-bound of the group, which is part of why Turnitin flags AI when other detectors don’t and vice versa. Copyleaks wins on breadth, though its own results show real limits once you leave English. Pangram posts the highest multilingual claims, but a vendor’s self-reported 99% should always be read against independent testing rather than taken at face value. Writers working across languages often reach for multilingual rewriting tools precisely because the detectors policing them are so unevenly calibrated.

3Low-resource languages are nearly invisible

The headline non-English numbers come from big languages like Spanish and French. For smaller languages, detection is closer to guesswork.

The February 2026 BLUFF benchmark put this on record. It spans 79 languages — 20 high-resource and 59 low-resource — with more than 200,000 samples, and found detection models performing substantially worse on the low-resource set, where there is far less training data to learn from.

There is a fix, but it is not the off-the-shelf detectors. A 2026 case study on Urdu fine-tuned a multilingual transformer (mdeberta-v3-base) to a 91.29% F1 score — English-grade accuracy — proving the gap is solvable with language-specific models rather than one English-trained classifier stretched across the world. This is the same brittleness that explains how humanizers exploit perplexity and burstiness: a system tuned narrowly on one distribution breaks on anything outside it.

4The false-positive trap for multilingual writers

The accuracy gap has a human cost: the writers most likely to be wrongly accused are the ones writing in a second language.

The most-cited evidence is still Liang et al. (2023), which recorded a mean false-positive rate of 61.3% on TOEFL essays by Chinese students against 5.1% for US students — no superseding figure has replaced it as of 2026. In the same study, 97% of the TOEFL essays were flagged by at least one of seven detectors, and 19% were unanimously misclassified as AI.

This is the overlap between language coverage and fairness, and it is why second-language writers need tools and habits built for ESL writing and a clear-eyed view of how detector bias against non-native English actually works. In 2026 testing, Turnitin and Originality.ai tied at the top overall (72% and 74%), but Turnitin produced more false positives on non-native English — a reminder that an “accurate” tool can still be the wrong one for a multilingual classroom. And because translated text largely slips through — neither the similarity checker nor the AI detector reliably catches content generated in one language and translated into English — the same blind spot that punishes honest ESL writers quietly rewards anyone gaming the system, which is exactly why a human-written essay can still look AI-generated.

Methodology

FAQ