Does AI Detection Work in Other Languages? Our Humanizer Roadmap

AI detection outside English is patchier than inside it: Turnitin's AI writing detection covers English, Spanish, Japanese and Arabic, and most detectors were built around English first. The Humanizer accepts English input today. Korean, Spanish, Chinese, Japanese and French are on the way, and the input guide in the app already lists them as coming soon. The Academic Humanizer website and this blog are already available in 10 languages.

We are not attaching dates to the new input languages, and this post explains why. Humanizing well in a language is not a translation task. It depends on how detectors behave in that language, how text is split into units, and what good academic prose looks like there.

Where language support stands today

English input covers a large share of academic publishing, and it is also the language most AI detectors were built around. Many researchers who publish in English write it as a second or third language, so English input already serves a multilingual group of authors.

Why multilingual humanizing is harder than it looks

A sentence-level humanizer like our Humanizer has to do three things in every language: split the draft into sentences correctly, find the sentences detectors flag, and rewrite only those in a register that fits the field. Each step gets harder outside English.

1. Detectors were mostly built for English

Most widely used AI detectors were developed and tested mainly on English text. Their behaviour on other languages is less studied, and research has already shown that detectors can misjudge writing that departs from the patterns they were trained on. Work on non-native English writing found that detectors flagged it as AI-generated far more often than writing by native speakers, largely because simpler, more predictable phrasing looks machine-like to them. We cover that study in GPT detectors are biased against non-native English writers.

If detection is uneven within English, it is reasonable to expect more unevenness across languages. Before we accept a new input language, we need to know which detectors handle it, how they disagree, and what a false flag looks like there.

2. Text is split differently

Detectors and language models do not read words. They read tokens, which are pieces of text that a model's vocabulary happens to contain. In English, common words are often single tokens. In languages with rich word endings, or with no spaces between words, the same idea can be cut into more and smaller pieces.

The results were significant / 결과가 유의미했다

Illustrative only. Green: a familiar whole-word piece. Red: a word broken into smaller pieces. Actual splits depend on each model's vocabulary.

This matters for two reasons. Statistical detectors judge how predictable each token is, so a different split changes the signal they see. And sentence splitting itself varies: Chinese and Japanese use different end punctuation and no spaces, while Korean and Japanese use sentence endings that carry politeness and register. A humanizer that splits sentences wrongly will rewrite the wrong spans.

3. Academic register differs by language

"Sounds human" is not the same target in every language. Good academic prose follows different conventions:

• Korean research writing typically uses the plain declarative ending and a formal, impersonal tone.
• Japanese papers usually use the formal written style rather than the polite spoken style, and the choice must stay consistent across a text.
• Spanish and French academic prose often favours longer sentences and impersonal constructions that would look wordy in English.
• Chinese academic writing has its own set of formal connectives and set phrases, and overusing them is a known marker of templated text.

A rewrite that reads as natural English style transplanted into another language would fail readers and probably detectors too. Each language needs its own sense of what a clean scholarly sentence looks like, plus care with field terminology, which often mixes local terms with English loanwords.

Do Turnitin and GPTZero detect AI in other languages?

Some do, for a limited set of languages. Turnitin's AI Writing Report guide lists English, Spanish, Japanese and Arabic as supported, and a file in another language does not get an AI writing report. That is not the same as a clean result. Turnitin's detection of AI-paraphrased and AI-bypassed text works on English only.

GPTZero's FAQ says its model supports English, Spanish, French, German and other languages, and Copyleaks says it covers more than 30. These are vendor statements. Independent testing outside English is much thinner than inside it.

If you write in another language, three points follow:

• Check whether the detector your institution or journal uses supports your language at all.
• Expect scores in a newly supported language to rest on less testing than English scores.
• Be careful with translation. In a 2023 study of 14 tools, human writing machine translated into English was classified correctly about 20 points less often than native English text (see what that study found).

What we will check before switching a language on

1. Sentence splittingDrafts in the language must split into sentences the way a reader would split them.
2. Detector coverageWe need to know how the detectors we check against behave on that language.
3. Register qualityRewrites must read as formal academic prose in that language, not translated English.
4. Meaning and termsNumbers, citations and field terminology must survive the rewrite.

Only when a language passes these checks does it move from "coming soon" to available. That is also why we are not promising dates: the work is done when it is done properly.

What to do in the meantime

If you write in English as an additional language, the Humanizer works for you today. Your draft is scanned sentence by sentence, only the flagged sentences are rewritten, and the rest stays as you wrote it. For writers whose plain English is misread by detectors, that is often the most useful setting. Our case study of a non-native English writer looks at this situation in more detail.

If your manuscript is in another language, wait for native support rather than translating it into English and back. A round trip through translation tends to flatten the register you are trying to keep. And whichever language you write in, check the AI policy of your institution or target journal before submitting AI-assisted work.

Frequently asked questions

Can I humanize Korean text with Academic Humanizer now?

Not yet. Input is English only today. Korean is one of the five languages coming soon, along with Spanish, Chinese, Japanese and French.

When will the new input languages launch?

We have not announced dates. Each language goes live after it passes the checks above.

Is the website available in my language?

The website and blog are available in English, Korean, Japanese, Simplified Chinese, Spanish, French, German, Hindi, Italian and Portuguese.

Why are AI detectors less reliable on non-English text?

Most were developed and tested mainly on English. Different tokenization and different writing conventions change the signals they rely on. For background, see how AI detectors work.

Building toward more languages

Academic Humanizer exists so that researchers' own ideas are not mislabelled as machine text. That problem does not stop at English, and neither will the Humanizer. For now, if your draft is in English, you can run it today and see sentence by sentence what detectors pick up.