Do AI detectors work on essays written in Spanish, German or Arabic, not English?
Mostly no, or only for a handful of languages. Turnitin runs separate models for English, Spanish and Japanese and does not score other languages at all. Benchmarks show detectors trained on English lose accuracy on languages they have not seen. A zero on German text usually means the tool had nothing useful to say, not that the text is human.
Cite
Global 100 Forum, "Do AI detectors work on essays written in Spanish, German or Arabic, not English?", https://global100.org/forum/q/do-ai-detectors-work-on-essays-written-in-spanish-german-or-arabic-not/, accessed 2026-11-10.I have a question about languages other than English, because I could not find a clear answer anywhere.
Most of my courses are in English, but this term I also take an Arabic literature module where we write essays in Arabic, and a German language course where the final paper is in German. One friend told me the detectors are "English only", so there is nothing to worry about. Another friend said her Spanish department definitely gets AI scores on Turnitin, so they must work for other languages too.
So I did a small experiment. I took one paragraph I wrote myself in German and one paragraph I asked ChatGPT to write in German, and I put both into two free checkers. Both came back 0% for both paragraphs. Human and AI, the same result. This made me more confused, not less. Does 0% mean the text is safe, or does it mean the tool did not understand anything?
I worry in both directions. I do not want to be flagged for my own Arabic essay, and I also do not want a system where English essays are checked and German essays are not, because that is not fair either. Can someone explain what actually happens with non-English text?
Write something first.
Give people something to work with: at least 30 words on what happened and what you tried.
That is too long. Keep it under 6,000 characters.
Write the question as the title, 15 to 140 characters, no links.
Pick a category.
Add a name (2 to 40 characters, no links).
That email address does not look right.
That was quick. Read the thread, then try again.
The form expired. Reload the page and post again.
Something went wrong with the form. Reload and try again.
Please complete the check and post again.
Limit reached for now. Try again later.
This thread is closed to new replies.
This thread no longer accepts replies.
Something went wrong with the form. Reload and try again.
Post a reply or question first (name and email), then this browser can vote, edit and accept answers.
You cannot vote on your own post.
Only the author (within 30 days) or the forum team can do that.
That email already has an account. Use your sign-in link instead.
New accounts are paused for the moment. Try again later.
That was already posted.
7 answers
- Accepted answerMaya LindqvistStaffEdits the AI text detection and academic integrity sections ·
Mostly no, or only for a handful of languages. Turnitin runs separate models for English, Spanish and Japanese and does not score other languages at all. Benchmarks show detectors trained on English lose accuracy on languages they have not seen. A zero on German text usually means the tool had nothing useful to say, not that the text is human.
The vendor your Spanish-department friend is describing is the clearest documented case. Turnitin's AI writing detection model release notes record Spanish detection arriving in September 2024 and Japanese in April 2025, each described as a different model from the English one, trained on writing in that language against specific generators (GPT-3.5 and GPT-4 for Spanish; GPT-4 and GPT-4o variants for Japanese). The May 2026 note is a further update to the Spanish model. There is no German or Arabic model listed. The same notes record that the detector surfaces no percentage at all below 20 percent, precisely to limit false positives. So for your Arabic essay and your German paper, the tool most universities use produces no score. Your friend's Spanish essays are scored by a younger, less tested model.
On the research side, the benchmark to know is MULTITuDE, published at EMNLP 2023. It assembled 74,081 human and machine-generated texts across eleven languages (Arabic, Catalan, Czech, German, English, Spanish, Dutch, Portuguese, Russian, Ukrainian and Chinese) from eight multilingual models, and tested statistical, black-box and fine-tuned detectors on how well they generalise to languages they were not trained on and whether training on several languages helps. Its starting point answers your friend's "English only" claim: before it there was very little authentic non-English data to evaluate detectors on, which is why free tools behave the way yours did.
Two practical points. Rob's translation idea does not work, and Hana explains why below. And your fairness concern is one reason several institutions decided policy should not depend on a score; see the non-native English thread and the accuracy thread.
- Priya N.Member ·
I tried something similar with Spanish last year for a bilingual section I was assisting. Three free tools, ten paragraphs each from students (with permission) and from a model. The results were noisy to the point of uselessness: one tool scored everything near zero like yours, one scored everything high including the student work, and one gave numbers that changed when I pasted the same paragraph twice. Length mattered too; anything under a few hundred words seemed to produce essentially random output.
Aisha, how long were your test paragraphs? If they were short, that alone could explain the flat zeros. But I would not read "0% on the AI paragraph" as the tool working. It looks more like the tool having no opinion and reporting that as a number.
- Rob M.Member ·
Plain question from a parent who is still learning this. Could you not just translate the German essay into English with Google Translate and then run the English version through the detector? Then at least the tool is working on a language it knows, and you would get a real answer instead of zeros.
- Tobias K.Member ·
From the administrative side, for completeness. Our platform's indicator displays an error state rather than a score when a submission is in an unsupported language. We have instructed staff in writing that an absent score is absent evidence, in both directions: it neither clears nor implicates the student, and no inference may be drawn from the tool declining to run. Aisha, if your institution uses the same platform, your Arabic essay will simply not be scored. Ask your module convenor how the department handles originality for Arabic submissions; the answer should be about the assessment design, not about software.
- Dan OkaforMember ·
I teach a heritage Spanish section alongside my English classes, and the Spanish detector going live was a strange experience. Suddenly a group of students who had been invisible to the tool were visible, scored by a model nobody at the school had evaluated, with no one able to say how it handled dialect, regional vocabulary, or kids who learned the language at the kitchen table rather than from a textbook. Nothing bad happened, as far as I know. But Aisha's fairness point cuts both ways: being unscored is a kind of protection, and being newly scored by a less mature model is a kind of exposure. Neither is equal treatment.
- Hana SatoMember ·
@Rob M. it is a natural idea, but it defeats itself. Machine translation output is machine-generated text. If you translate Aisha's German essay into English and score that, you are scoring the translator's writing, not hers, and the detector will quite reasonably see model-shaped prose. It would look like evidence against her when it is evidence of nothing. The reverse is also true: an AI-written German paragraph machine-translated into English does not necessarily keep the statistical features the English detector was trained on. So the test produces a number, but the number has no relationship to who wrote the original. Scores only mean anything when the detector was trained on the language and the kind of text it is looking at.
- Aisha R.Member ·
@Maya Lindqvist thank you, this is very clear. So the zeros meant "no opinion", not "safe", and the Arabic essay will probably not be scored at all. I will ask my convenor what the department actually does for Arabic. And thank you to Hana for the explanation about translation; I had wondered the same thing as Rob.