How Multilingual AI Detection Lags Behind English (and Why That Matters)

AI detection in non-English languages runs years behind English: the training data is thinner, the baseline statistics are miscalibrated, and the accuracy numbers vendors publish for English mostly don’t exist for anything else. That gap isn’t a footnote. It decides whether a student in Madrid or a copywriter in Jakarta gets a fair score — and right now, they often don’t.

Key takeaways

• Detectors are statistical models, and the statistics were overwhelmingly learned from English. Everything else inherits a handicap.
• Even OpenAI’s own (now retired) classifier came with an explicit warning: don’t use it on non-English text.
• Vendor language support claims vary wildly — from “English plus a couple of languages” (Turnitin) to “dozens” (Copyleaks) — and independent verification is rare in every case.
• The harm is asymmetric: non-native English writers get over-flagged, while AI text in lower-resource languages sails through.
• A score on translated text is close to meaningless, because translation replaces every word the detector would judge.

Why AI detection in non-English languages lags

AI detectors don’t understand language; they measure it. They learn what “typical human writing” and “typical machine writing” look like statistically — how predictable each word is, how much sentence rhythm varies — and then score new text against those baselines. Every part of that pipeline was built on English first.

Training data. The big public corpora of AI-generated text, the benchmark datasets, the human-writing baselines — English dominates all of them. A detector’s accuracy in any language tracks how much labeled data it has seen there, and for most languages the answer is “a sliver of the English pile.” The RAID benchmark, one of the broadest public evaluations of detector robustness, illustrates the pattern in a different dimension: detectors that look strong on familiar test data degrade fast on text that differs from what they trained on. Language is the biggest domain shift there is.

Tokenization. Most detection models inherit tokenizers optimized for English. Morphologically rich languages — Finnish, Turkish, Korean — get carved into fragments those models handle less naturally, which distorts the perplexity signal detectors lean on. It’s like measuring a poem with a ruler marked in the wrong units.

Baseline calibration. “Suspiciously uniform sentence length” is a useful signal in English. But languages differ in natural rhythm, clause structure, and repetition norms. Formal academic German runs long and subordinate-clause-heavy; conversational Japanese omits subjects English detectors expect. A threshold tuned on English prose will misread those patterns as anomalies.

What vendors actually claim

The honest summary of the market: claims outrun evidence everywhere, and the gap is widest outside English.

Turnitin launched its AI writing detection in April 2023 for English only, and has since extended it to a small set of additional languages, including Spanish and Japanese. Copyleaks advertises AI detection across dozens of languages. GPTZero and most free tools are English-first, with softer language about “support” elsewhere. What almost no vendor publishes: per-language false-positive rates on independent test sets. The 98-99% accuracy figures that circulate come from English evaluations, then quietly stand in for the whole product.

OpenAI set the honest precedent here, oddly enough. When it launched its AI Text Classifier in January 2023, it stated plainly that the tool should not be used on non-English text because performance was significantly worse. That classifier was retired six months later for low accuracy — in English.

Who pays for the gap

The damage lands in two opposite directions at once.

Over-flagging non-native writers. The best-documented harm isn’t about foreign-language text at all — it’s about non-native speakers writing *in English*. The 2023 Stanford study by Liang and colleagues found detectors flagged TOEFL essays by non-native writers at rates above 60% while barely flagging native-written essays. Simpler vocabulary and more uniform sentence structure — the natural fingerprint of writing in a second language — happens to overlap with the fingerprint of machine text. We’ve written a practical guide on why ESL students get flagged and what to do about it.

Under-detecting AI text elsewhere. Flip the coin: if a detector has thin baselines in a language, AI-generated text in that language gets through more easily. Content farms figured this out early. AI-spun articles in lower-resource languages face far less scrutiny — from detectors and from platforms — than the same operation would in English.

And translation sits in the middle, confusing everything. Run human-written Portuguese through a translation engine and the English output often scores AI-like, because translation engines produce statistically smooth, high-probability prose. The detector is scoring the translator, not the author. The same logic means “translate to evade” also works — which tells you how little signal survives the round trip.

Why the gap persists

You’d expect market pressure to close this by now. Three reasons it hasn’t:

1. The economics don’t reward it. English is where the paying customers are — US and UK universities, English-language SEO. Building and validating per-language baselines is expensive, and the buyers rarely demand per-language evidence.
2. The data problem compounds. Newer models write increasingly fluent non-English prose, so vendors are chasing a moving target in dozens of languages simultaneously with a fraction of the data they have for English.
3. Nobody’s checking. Independent evaluations — the academic studies, the journalism — test English almost exclusively. Without third-party scrutiny, “supports 30 languages” never has to mean “works equally well in 30 languages.” That’s a big part of why the myth that detectors work perfectly in other languages survives.

What to do with this, practically

If you write, teach, or edit across languages, a few rules of thumb hold up:

• Discount any score on non-English text unless the vendor publishes accuracy figures for that specific language. Most don’t.
• Never treat a score on translated text as evidence about the original author, in either direction.
• If you’re a non-native writer working in English, your risk is over-flagging, not under-detection. Check your own drafts before a gatekeeper does, and revise the passages that read as statistically flat — you can try it on your own text for free.
• If you’re an educator, treat detector output on multilingual students’ work as the weakest form of evidence you have, not the strongest. The documented bias runs against exactly the students with the least power to contest it.

For the broader landscape of how these tools work and fail, you can browse the rest of our writing on detection, or see what each plan handles if you’re checking text at volume.

Frequently asked questions

Does AI detection work in languages other than English? It exists, but with a real accuracy gap. Detectors learn statistical patterns from training data, and that data is overwhelmingly English. Some vendors support other major languages — Copyleaks advertises support for dozens, and Turnitin has extended its model beyond English to a small set including Spanish and Japanese — but published accuracy figures for non-English text are scarcer, and independent testing is thinner still.

Why would a detector be worse in Spanish or Hindi than in English? Three compounding reasons: less training data in that language, tokenizers built primarily for English that carve other languages into awkward pieces, and baseline statistics — like typical sentence rhythm — that were calibrated on English prose. A pattern that means ‘machine-like’ in English may simply be how a language normally sounds.

Are non-English speakers more likely to be falsely flagged? The evidence points that way in two respects. Stanford researchers showed in 2023 that detectors flagged essays by non-native English writers at dramatically higher rates than native writers. And for text written in other languages, smaller training corpora generally mean less reliable scores in both directions — more false flags and more misses.

Which languages have the best AI detection coverage? English is far ahead of everything else. Major European languages — Spanish, French, German, Portuguese — form the second tier, since vendors have the most non-English data there. Coverage drops off sharply for morphologically rich or lower-resource languages, and for many of the world’s languages no vendor publishes any accuracy claim at all.

Can I trust an AI score on translated text? Be extra careful. Translation rewrites every token, so a detector is really scoring the translator’s style, not the original author’s. Machine-translated human writing frequently scores as AI-like because translation engines produce statistically smooth output. A score on translated text says very little about who wrote the original.

The bottom line

Multilingual AI detection lags English because everything detectors are built from — data, tokenizers, baselines, benchmarks — was built for English first. The result is a technology that over-accuses non-native writers in English and under-scrutinizes AI text everywhere else, while marketing pages round both problems away. Until vendors publish per-language error rates, the only safe posture is skepticism scaled to distance from English: the further a text sits from the detector’s home turf, the less its score means.