AI detectors for localisation and translation teams

Machine translation moved detectors in the one record built to test it. Weber-Wulff 2023 wrote 9 human documents in Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish, machine-translated them into English with DeepL or Google Translate, and scored them beside 9 human documents written in English: average binary accuracy across 14 tools was 94% on the English human class (01-Hum) and 69% on the machine-translated class (02-MT), and the authors describe a drop of about 20 points on the translated class (arXiv 2306.15666, Table 5 and Section 4, read 8 November 2026; Weber-Wulff 2023 record). Two of the 14 tools' false accusations in that study fell on translated texts: the GPT-2 Output Detector's 2 of 18 were both machine-translated documents, and Writer's 1 of 18 was too; Winston AI and Compilatio, the two cards in this view that were tested, recorded 0. A human translation scored by a detector trained on English prose is the document class the index has the least evidence on, and a machine-translated draft is the class with the measured drop.

Non-native writing is the class with the worst false positive record, and a localisation team's source texts are often written by non-native authors before translation. Liang 2023 reported an average false positive rate of 61.22% across seven detectors on 91 TOEFL essays against near-perfect accuracy on 88 US 8th-grade essays, and attributed the gap to lower lexical variety and simpler syntax (arXiv 2304.02819, read 8 November 2026; Liang 2023 record); Sapling and Originality.ai in this view were among the seven. Ibrahim 2023 found that the GPT-2 Output Detector gave 16 of 120 ESL essays (13.3%) an originality score under 95% (DOI 10.1186/s40468-023-00260-2, read 9 November 2026; Ibrahim 2023 record). Orenstrakh 2023 is the one record with text written in another language rather than translated: on 10 Spanish theses and 10 Spanish ChatGPT texts, Copyleaks and the OpenAI classifier called every submission AI-generated while the GPT-2 Output Detector, GPTZero, CheckForAI and GLTR called every ChatGPT text human (arXiv 2307.07411, Sections 4.1 and 5, read 9 November 2026; Orenstrakh 2023 record). Chaka 2023 translated six ChatGPT responses into five languages and found Copyleaks classified 3 of 6 German, 5 of 6 French and 6 of 6 Spanish translations as AI (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026; Chaka 2023 record).

The gap between the two fields in the table is the finding. The 14 cards state 488 language slots between them and hold 8 non-English records between 6 of them, and no indexed record scored any detector in more than five languages; Compilatio's 24-language measurement is the vendor's own and Pangram's multilingual model card likewise. Turnitin, outside this view at 7 languages, publishes release dates for its Spanish (12 September 2024) and Japanese (2 April 2025) models and states that paraphrase detection is English only, the one dated language roadmap in the index (Turnitin languages question). A localisation team should read a language count as a list of inputs the tool accepts, not of languages it has been shown to score, and no detector score should be the only evidence in a decision about a translator's work.

Sources

• https://smodin.io/ai-content-detector (read 9 November 2026)
• https://docs.smodin.io/reference/v2-ai-detection-single (read 9 November 2026)
• https://www.resemble.ai/benchmarks (read 9 November 2026)
• https://undetectable.ai/ai-detector (read 9 November 2026)
• https://www.pindrop.com/company-news/pindrop-launches-pulse-inspect (read 9 November 2026)
• https://www.compilatio.net/en/ai-detector-info/ (read 9 November 2026)
• https://support.compilatio.net/hc/en-us/articles/17408769118993 (read 9 November 2026)
• https://copyleaks.com/ai-content-detector (read 8 November 2026)
• https://originality.ai (read 8 November 2026)
• https://truthscan.com/faq (read 9 November 2026)
• https://www.quetext.com/ai-content-detector (read 9 November 2026)
• https://www.pangram.com/ (read 8 November 2026)
• https://proofademic.ai/multilingual-detection/?utm_source=aidetectorindex.com&utm_medium=owned&utm_campaign=for_localisation_teams (read 8 November 2026)
• https://walterwrites.ai/ai-detector/?utm_source=aidetectorindex.com&utm_medium=owned&utm_campaign=for_localisation_teams (read 8 November 2026)
• https://sapling.ai/ai-content-detector (read 8 November 2026)
• https://gowinston.ai/ai-content-detector/ (read 8 November 2026)
• https://arxiv.org/pdf/2306.15666 (read 8 November 2026)
• https://arxiv.org/pdf/2304.02819v3 (read 8 November 2026)
• https://doi.org/10.1186/s40468-023-00260-2 (read 9 November 2026)
• https://arxiv.org/pdf/2307.07411 (read 9 November 2026)
• https://doi.org/10.37074/jalt.2023.6.2.12 (read 9 November 2026)

Revision log: 2026-11-09: view created at index version 2026.11.9