Do AI detectors work on machine-translated text? What two studies measured, with bases
Not well, on the two records that tested it. In Weber-Wulff 2023, the 14 detectors tested averaged 69% binary accuracy on 9 human-written texts that had been machine-translated into English with DeepL or Google Translate, against 94% on the 9 untranslated human texts, and 6 of GPTZero's 9 false accusations in that study fell on the translated class (DOI 10.1007/s40979-023-00146-z, Tables 5 and 9, read 8 November 2026). In Chaka 2023, GPTZero classified 30 of 30 ChatGPT responses as human-written after Google Translate had rendered them into German, French, Spanish, Southern Sotho and isiZulu, while Copyleaks classified 14 of 18 translations into German, French and Spanish as AI-generated (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026). Translation moves text in both directions: human text translated into English picks up false flags, and AI text translated out of English loses them.
The two studies tested opposite cases, and the table keeps them apart. Weber-Wulff 2023 asked whether a detector would wrongly flag a human author whose text reached English through a machine; Chaka 2023 asked whether a detector would still catch AI text once a machine had moved it into another language. A third record, Orenstrakh 2023, tested original Spanish texts rather than translations and is listed below as the nearest non-English comparison, labelled as such. This page is one of the questions answered from the index; its siblings ask whether a 100% accurate detector exists and which detectors claim under 1% false positives, and the languages ranking lists what each vendor says it supports.
| Study and test | Detector | Result on translated text | Base | Comparison in the same study | Record |
|---|---|---|---|---|---|
| Weber-Wulff 2023, human text machine-translated into English (class 02-MT) | All 14 tools, average | 69% binary accuracy | 9 documents, 14 tools | 94% on the 9 untranslated human documents (01-Hum) | Weber-Wulff 2023 |
| Weber-Wulff 2023, class 02-MT | GPTZero | 6 of 9 false accusations fell on translated texts | 18 human documents, 9 translated | 9 false accusations in 18 overall (50%), the highest of 14 | Weber-Wulff 2023 |
| Weber-Wulff 2023, class 02-MT | GPT-2 Output Detector | Both false accusations fell on translated texts | 18 human documents, 9 translated | 2 false accusations in 18 (11.1%) | Weber-Wulff 2023 |
| Weber-Wulff 2023, class 02-MT | Writer | The one false accusation fell on a translated text | 18 human documents, 9 translated | 1 false accusation in 18 (5.6%); tool withdrawn 22 December 2025 | Weber-Wulff 2023 |
| Weber-Wulff 2023, class 02-MT | Turnitin, ZeroGPT, Winston AI | 0 false accusations, translated or not | 18 human documents, 9 translated | 0 of 18 overall | Weber-Wulff 2023 |
| Chaka 2023, ChatGPT text machine-translated out of English | GPTZero | 30 of 30 translations classified human-written | 6 responses in each of 5 languages | 8 of 21 English responses classified AI-generated, 10 human, 3 indecisive | Chaka 2023 |
| Chaka 2023, ChatGPT text machine-translated out of English | Copyleaks | 14 of 18 translations classified AI-generated: German 3 of 6, French 5 of 6, Spanish 6 of 6 | 6 responses in each of 3 languages | 15 of 21 English responses classified AI-generated, the most of five tools | Chaka 2023 |
| Orenstrakh 2023, original Spanish texts (not translated) | Copyleaks | 10 of 10 human Spanish theses flagged as AI; 10 of 10 ChatGPT Spanish texts flagged | 10 human, 10 ChatGPT | 1 false positive in 114 English human submissions | Orenstrakh 2023 |
| Orenstrakh 2023, original Spanish texts (not translated) | GPT-2 Output Detector, GPTZero, GLTR | 0 of 10 ChatGPT Spanish texts flagged as AI | 10 ChatGPT texts | 95%, 45% and 95% on 20 ChatGPT English submissions | Orenstrakh 2023 |
Results as each paper prints them: arXiv 2306.15666 for Weber-Wulff 2023, DOI 10.37074/jalt.2023.6.2.12 for Chaka 2023 and arXiv 2307.07411 for Orenstrakh 2023, read 8 and 9 November 2026. The Orenstrakh rows are original Spanish, not machine translation, and are included because no third record tested translation.
Does machine translation make human text look like AI?
In Weber-Wulff 2023 it did, by 25 points on average. The authors wrote 9 texts in Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish and machine-translated them into English with DeepL or Google Translate; across the 14 tools, binary accuracy on that class was 69% against 94% on the 9 untranslated human texts, and the journal version puts the drop at about 20% (DOI 10.1007/s40979-023-00146-z; Weber-Wulff 2023 record, read 8 November 2026). The drop was not spread evenly. GPTZero's 9 false accusations in 18 human documents, the highest of the 14 tools, included 6 on the translated class, and both of the GPT-2 Output Detector's 2 false accusations and Writer's 1 fell on translated texts (false positive rate ranking). Turnitin, ZeroGPT, Winston AI and five other tools recorded 0 false accusations on translated and untranslated texts alike. The sample is 9 translated documents, so a single text moves a tool's rate by 11 points, and the authors restrict their claims to English output.
Does translating AI text hide it from detectors?
The non-English evidence ranking, added 9 November 2026, counts the Weber-Wulff 2023 translated class for nine cards and the Chaka 2023 translations for two, beside the Spanish, TOEFL, ESL and non-native graduate records.
In Chaka 2023 it hid it completely from one tool and partly from another. The author ran 6 ChatGPT responses through Google Translate into German, French, Spanish, Southern Sotho and isiZulu and submitted the 30 translations to GPTZero, which classified every one as human-written; the same 6 responses translated into German, French and Spanish went to Copyleaks, which classified 14 of 18 as AI-generated, with Spanish at 6 of 6, French 5 of 6 and German 3 of 6 (DOI 10.37074/jalt.2023.6.2.12; Chaka 2023 record, read 9 November 2026). The author attributes GPTZero's result to perplexity: the translated texts scored far higher than the English originals, the highest at 2,478 for German, 1,715 for Southern Sotho and 938 for isiZulu, and a high perplexity reads as human to a perplexity-based detector. The study contains no human-written texts, so it measures missed AI text and nothing about false flags, and the base is 30 translations of 6 responses.
What happens to non-English text that was never translated?
The nearest record is Orenstrakh 2023, which is not a translation test and is labelled as such. The authors submitted 10 human final-thesis papers written in Spanish and 10 ChatGPT texts in Spanish to the detectors that accepted them: Copyleaks and the OpenAI AI Text Classifier flagged all 20 as AI, human or not, while the GPT-2 Output Detector, GPTZero, CheckForAI and GLTR flagged none of the 10 ChatGPT texts, for an average ChatGPT accuracy of about 17.50% across the tools (arXiv 2307.07411, Tables 11 and 12; Orenstrakh 2023 record, read 9 November 2026). The same Copyleaks recorded 1 false positive in 114 English human submissions in the same study. Original Spanish produced the same two failure modes the translation studies show, every human text flagged by one tool and every AI text missed by four others, which is why the row sits in the table with its label.
Liang 2023 belongs beside these results without being a translation study: seven detectors averaged a 61.22% false positive rate on 91 TOEFL essays by non-native English writers against near-perfect accuracy on 88 US eighth-grade essays (DOI 10.1016/j.patter.2023.100779; Liang 2023 record, read 8 November 2026). A translated text and a non-native text share the property the detectors appear to punish, English with lower perplexity than a native writer's, and the three records point the same way.
No vendor in the index publishes a figure for translated text. The languages ranking records that Undetectable AI states 50+ languages, Copyleaks and Originality.ai 30+, Pangram 24 and GPTZero 5, each as the vendor's own statement, and the ranking notes that no study measures accuracy per language. A listed language means the tool accepts it; the question page on Spanish support sets every card's Spanish statement beside the Orenstrakh 2023 result above.
No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a score on translated text rests on two studies with bases of 9 and 30.
The localisation teams view, added 9 November 2026, reads these records beside the 14 cards that state 10 or more languages and finds that 6 of the 14 have any non-English record.
Questions about AI detectors and translated text
Do AI detectors flag translated text as AI?
Some do, on the one study that tested it. Weber-Wulff 2023 recorded 6 of GPTZero's 9 false accusations, both of the GPT-2 Output Detector's 2 and Writer's 1 on 9 machine-translated human texts, while Turnitin, ZeroGPT and Winston AI recorded none (DOI 10.1007/s40979-023-00146-z, read 8 November 2026). Average accuracy across 14 tools fell from 94% to 69% on that class.
Can translating AI text beat a detector?
In Chaka 2023 it beat GPTZero on 30 of 30 translations into five languages and beat Copyleaks on 4 of 18 translations into three (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026). The base is 6 ChatGPT responses, and the index records no later test of the same method.
Which detectors did not flag translated human text?
Turnitin, ZeroGPT, Winston AI and five tools without a card recorded 0 false accusations in all 18 human documents of Weber-Wulff 2023, translated texts included (Weber-Wulff 2023 record, read 8 November 2026). Zero on 9 translated texts rules out a high rate, not a rate of several percent.
Has any study tested detectors in Spanish?
Orenstrakh 2023 tested 10 human and 10 ChatGPT texts written in Spanish, not translated: Copyleaks flagged all 20 as AI and the GPT-2 Output Detector, GPTZero and GLTR flagged none of the 10 ChatGPT texts (arXiv 2307.07411, read 9 November 2026). Chaka 2023's Spanish rows are translations of English ChatGPT text.
Do vendors publish accuracy for translated text?
No card in the index carries one. The vendors publish language lists and counts (languages ranking), and Writer's disclaimer stated the tool was "better at detecting English text than other languages" before the tool was withdrawn (archived disclaimer, read 9 November 2026).
Sources
- Weber-Wulff et al. 2023, DOI 10.1007/s40979-023-00146-z: https://doi.org/10.1007/s40979-023-00146-z (read 8 November 2026)
- Weber-Wulff et al. 2023, arXiv: https://arxiv.org/pdf/2306.15666 (read 8 November 2026)
- Chaka 2023, DOI 10.37074/jalt.2023.6.2.12: https://doi.org/10.37074/jalt.2023.6.2.12 (read 9 November 2026)
- Orenstrakh et al. 2023: https://arxiv.org/abs/2307.07411 (read 9 November 2026)
- Liang et al. 2023, DOI 10.1016/j.patter.2023.100779: https://doi.org/10.1016/j.patter.2023.100779 (read 8 November 2026)
- Writer AI content detector disclaimer, archived 9 December 2025: http://web.archive.org/web/20251209115654/https://writer.com/disclaimer-aicd/ (read 9 November 2026)
Revision log: 2026-11-09: page created from the Weber-Wulff 2023, Chaka 2023 and Orenstrakh 2023 records at index version 2026.11.4; 2026-11-09: Spanish question page linked; 2026-11-09: non-English evidence ranking linked (seventh wave); 2026-11-09: eighth-wave pages linked
Related pages: Which AI detectors claim a false positive rate under 1%, and who checked each claim?, Is an AI detector itself an AI? The four detection methods behind the cards, each sourced, AI detector false positive rates, ranked and Questions about AI detectors, answered from the index.