AI detectors ranked by study records on non-English and non-native text
GPTZero has five indexed study records that print a result on text outside native English prose, the most of any detector in the index: machine-translated human text in Weber-Wulff 2023, TOEFL essays by non-native writers in Liang 2023, Spanish coursework in Orenstrakh 2023, five translation languages in Chaka 2023 and non-native graduate papers in Van Vlasselaer 2026, each read in the primary source on 8 or 9 November 2026 (DOI 10.1007/s40979-023-00146-z, DOI 10.1016/j.patter.2023.100779, arXiv 2307.07411, DOI 10.37074/jalt.2023.6.2.12, DOI 10.1007/s40979-026-00226-w). The GPT-2 Output Detector has four and Copyleaks three. Nineteen of the 72 cards have at least one such record, drawn from six of the 35 records (GPTKit and PlagiarismCheck joined on 10 November 2026, then Writefull and Check For AI on the Weber-Wulff 2023 machine-translated class later the same day, then AI Writing Check on the Liang 2023 TOEFL class with the eleventh wave); the other 53 cards, including eight of the twelve that state 20 or more languages, have none.
Updated Studies dated 2023 to 2026
Short answer
The sort is the count of indexed records that print a result for the card on non-English, machine-translated or non-native English text, a count of evidence and not a verdict on any language: GPTZero 5; the GPT-2 Output Detector 4; Copyleaks 3; ZeroGPT and Turnitin 2 each, ZeroGPT first on study records; then nine cards with one record each, from Originality.ai to Compilatio, ordered on total study records and name. The records are Weber-Wulff 2023 (9 cards), Liang 2023 (5), Orenstrakh 2023 (4), Van Vlasselaer 2026 (4), Chaka 2023 (2) and Ibrahim 2023 (1), all read 8 and 9 November 2026. The results point both ways: Copyleaks flagged all 10 Spanish human theses and GPTZero called every translated ChatGPT response human. Vendor multilingual figures, such as Pangram's model card and Copyleaks' per-language table, are excluded because the vendor measured them. No detector score should be the only evidence in a misconduct decision, least of all on a writer's second language.
Which AI detectors have the most study records on non-English or non-native text?
| Rank | Detector | Records | Records and the text they cover | Result as the record prints it | Best for | Main limitation |
|---|---|---|---|---|---|---|
| 1 |  GPTZero | 5 records | Weber-Wulff 2023 (machine-translated human text), Liang 2023 (TOEFL essays by non-native writers), Orenstrakh 2023 (Spanish), Chaka 2023 (five translation languages), Van Vlasselaer 2026 (human papers by non-native graduate students) | 6 of its 9 false accusations fell on the 9 machine-translated texts (Weber-Wulff 2023); 30 of 30 ChatGPT responses translated into German, French, Spanish, Southern Sotho and isiZulu read as human (Chaka 2023); every Spanish ChatGPT text read as human (Orenstrakh 2023); 0 of 40 non-native human papers flagged (Van Vlasselaer 2026); TOEFL rate in a chart only (Liang 2023) | Readers who want the widest non-English record, whichever way it points | Five records point both ways: translated human text was flagged, translated AI text was missed, and the TOEFL per-tool rate is printed only as a chart |
| 2 |  GPT-2 Output Detector | 4 records | Weber-Wulff 2023, Liang 2023, Orenstrakh 2023, Ibrahim 2023 (120 ESL essays) | 2 false accusations in 18 human documents, both on machine-translated texts (Weber-Wulff 2023); every Spanish ChatGPT text read as human (Orenstrakh 2023); 16 of 120 ESL essays (13.3%) below a 95% originality cut (Ibrahim 2023); TOEFL rate in a chart only (Liang 2023) | Researchers comparing a 2019 baseline on non-native and translated text | A GPT-2 era model: Ibrahim's 13.3% carries the author's inverted label and is not entered as a rate |
| 3 |  Copyleaks | 3 records | Orenstrakh 2023 (Spanish), Chaka 2023 (German, French, Spanish), Van Vlasselaer 2026 | All 10 Spanish human theses and all 10 Spanish ChatGPT texts flagged as AI (Orenstrakh 2023); 3 of 6 German, 5 of 6 French and 6 of 6 Spanish translated ChatGPT responses classified AI (Chaka 2023); 0 of 40 non-native human papers flagged (Van Vlasselaer 2026) | Institutions with Spanish or French submissions who want the one card with a per-language translation result | The Spanish result is 10 of 10 human theses flagged, the sharpest non-English false positive figure in the index |
| 4 |  ZeroGPT | 2 records | Weber-Wulff 2023, Liang 2023 | 0 false accusations in 18 human documents, 9 of them machine-translated (Weber-Wulff 2023); TOEFL rate in a chart only (Liang 2023) | Readers who want a zero on translated human text from a free tool | The Liang 2023 per-tool rate is a chart; the Weber-Wulff zero covers 9 translated documents |
| 5 |  Turnitin | 2 records | Weber-Wulff 2023, Van Vlasselaer 2026 | 0 false accusations in 18 human documents, 9 of them machine-translated (Weber-Wulff 2023); 0 of 40 human papers by non-native graduate students flagged (Van Vlasselaer 2026) | Institutions with many non-native writers who want two zeros on human text | Both zeros are on human text; the same Van Vlasselaer paper scored all 40 fully AI papers under 20% |
| 6 |  Originality.ai | 1 record | Liang 2023 | One of seven detectors on 91 TOEFL essays; the paper prints the per-tool rate only in a chart (Liang 2023) | Readers who want the card named in the TOEFL study | No number: the per-tool TOEFL rate is a bar in Figure 1a |
| 7 |  Writer | 1 record | Weber-Wulff 2023 | 1 false accusation in 18 human documents, on a machine-translated text (Weber-Wulff 2023) | Readers of the 2023 record on the withdrawn tool | The tool was withdrawn on 22 December 2025 |
| 8 | DetectGPT | 1 record | Weber-Wulff 2023 | 5 of 9 machine-translated human documents classified correctly, 9 of 9 untranslated (Weber-Wulff 2023) | Researchers who want the method's translated-text split | A research method run through the 2023 demo; 4 of 9 translated texts misread |
| 9 |  GLTR | 1 record | Orenstrakh 2023 | Every Spanish ChatGPT submission called human (Orenstrakh 2023) | Researchers who want the visual tool's Spanish result | One record, and it is a miss on all 10 Spanish AI texts |
| 10 |  Pangram | 1 record | Van Vlasselaer 2026 | 0 of 40 human papers by non-native English-speaking graduate students flagged (Van Vlasselaer 2026) | Teams that want a zero on non-native academic prose from a 2026 record | One record; the vendor's multilingual figures sit on its model card and are excluded here |
| 11 |  BrandWell | 1 record | Weber-Wulff 2023 | 0 false accusations in 18 human documents, 9 of them machine-translated (Weber-Wulff 2023) | Readers who want the translated-text zero beside the 36 of 36 misses | The same record missed all 36 AI documents |
| 12 |  Sapling | 1 record | Liang 2023 | One of seven detectors on 91 TOEFL essays; per-tool rate in a chart only (Liang 2023) | Readers who want the card named in the TOEFL study | No number printed for the tool |
| 13 |  GPTKit | 1 record | Orenstrakh 2023 (Spanish: 10 human theses and 10 ChatGPT texts) | 80% of 10 human and 20% of 10 ChatGPT Spanish submissions classified correctly, Table 12; the vendor states English only | Readers who want a Spanish result for a detector that claims one language | 10 texts per set; the Spanish set is the paper's smallest |
| 14 |  Winston AI | 1 record | Weber-Wulff 2023 | 0 false accusations in 18 human documents, 9 of them machine-translated (Weber-Wulff 2023) | Readers who want a translated-text zero from a card stating 11 languages | One record, from 2023 |
| 15 |  Writefull | 1 record | Weber-Wulff 2023 (machine-translated English from Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish, class 02-MT, 9 of 18 human documents) | 1 false accusation in 18 human documents, including the 9 machine-translated ones, Table 9; the per-class accuracy for the tool is in the preprint's Table 5 | Readers who want the one translated-text result for a discontinued tool | 9 translated documents; the vendor stated no language; the tool is discontinued |
| 16 |  AI Writing Check | 1 record | Liang 2023: 91 TOEFL essays by non-native English writers against 88 US eighth-grade essays | One of seven detectors whose average false positive rate on the TOEFL essays was 61.22% and on the US essays 5.19%; per-tool rates in Figure 1a only | Readers who meet a 2023 verdict on a non-native writer's essay | The averaged rate is the study's, not the tool's; withdrawn since August 2023 |
| 17 |  Check For AI | 1 record | Weber-Wulff 2023 (machine-translated English from Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish, class 02-MT, 9 of 18 human documents) | 1 false accusation in 18 human documents, including the 9 machine-translated ones, Table 9; the per-class accuracy for the tool is in the preprint's Table 5 | Readers who want the one translated-text result for a sunset tool | 9 translated documents; the vendor stated English only; the tool was sunset in 2023 |
| 18 |  Compilatio | 1 record | Weber-Wulff 2023 | 0 false accusations in 18 human documents, 9 of them machine-translated (Weber-Wulff 2023) | European institutions who want a translated-text zero from a vendor stating 35 languages | One record, on the 2023 free tool; 7 of 9 paraphrased texts missed |
| 19 |  PlagiarismCheck | 1 record | Weber-Wulff 2023 (machine-translated English from Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish, class 02-MT, 9 of 18 human documents) | 0 false accusations in 18 human documents, including the 9 machine-translated ones, Table 9; the per-class accuracy for the tool is in the preprint's Table 5 | Readers who want the one translated-text result for an LMS-integrated suite | 9 translated documents; the vendor claims English and Spanish and the record tested English |
The six records and what they cover
| Record | Non-English or non-native text | Cards with a printed result | Source |
|---|---|---|---|
| Weber-Wulff 2023 | Class 02-MT: 9 human documents written in Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish and machine-translated to English with DeepL or Google Translate, half of the 18 human documents | 12: GPTZero, GPT-2 Output Detector, ZeroGPT, Turnitin, Writer, DetectGPT, BrandWell, Winston AI, Compilatio, PlagiarismCheck, Check For AI, Writefull | DOI 10.1007/s40979-023-00146-z, read 8 November 2026 |
| Liang 2023 | 91 TOEFL essays by non-native English writers against 88 US eighth-grade essays; per-tool rates in Figure 1a only, average 61.22% | 6: GPTZero, GPT-2 Output Detector, ZeroGPT, Originality.ai, Sapling, AI Writing Check | DOI 10.1016/j.patter.2023.100779, read 8 November 2026 |
| Orenstrakh 2023 | 10 human theses and 10 ChatGPT texts in Spanish, from a South American institution | 4: GPTZero, GPT-2 Output Detector, Copyleaks, GLTR | arXiv 2307.07411, read 9 November 2026 |
| Chaka 2023 | 6 ChatGPT responses machine-translated into German, French, Spanish, Southern Sotho and isiZulu (30 texts to GPTZero, 18 to Copyleaks) | 2: GPTZero, Copyleaks | DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026 |
| Ibrahim 2023 | 120 essays by ESL first-year composition students in Kuwait, written before ChatGPT | 1: GPT-2 Output Detector | DOI 10.1186/s40468-023-00260-2, read 9 November 2026 |
| Van Vlasselaer 2026 | 40 human papers by non-native English-speaking graduate students at a Belgian faculty, at least 4,000 words each | 4: GPTZero, Pangram, Copyleaks, Turnitin | DOI 10.1007/s40979-026-00226-w, read 9 November 2026 |
The remaining 28 cards have no record on non-English or non-native text: QuillBot, Grammarly, Scribbr, Phrasly, JustDone, Proofademic, Walter Writes, Hive, AIDetectPlus, Codequiry, WasItAI, Reality Defender, Undetectable AI, AI or Not, Grammarly Authorship, Turnitin Clarity, SynthID, Smodin, Surfer, Quetext, TruthScan, Sightengine, Deepware, Sensity, Codeleaks, Illuminarty, Pindrop Pulse, Resemble Detect. Twelve of them state a language count on the languages ranking, from Smodin's 100 and Undetectable AI's 50 down; the count there is the vendor's statement and the count here is the evidence, and the two pages share nine cards (Copyleaks, Originality.ai, Pangram, Compilatio, Sapling, Turnitin, Winston AI, GPTZero and BrandWell).
One of the evidence rankings in the rankings hub, beside languages, which counts the languages a vendor states, and independently tested, which counts every record. The machine-translated text question and the Spanish question read two of the six records in detail. The home board answers which AI detectors have been tested on non-English text?
Which AI detectors should you check first?
- GPTZero, rank 1Best for: Readers who want the widest non-English record, whichever way it points. Main limitation: Five records point both ways: translated human text was flagged, translated AI text was missed, and the TOEFL per-tool rate is printed only as a chart.
- GPT-2 Output Detector, rank 2Best for: Researchers comparing a 2019 baseline on non-native and translated text. Main limitation: A GPT-2 era model: Ibrahim's 13.3% carries the author's inverted label and is not entered as a rate.
- Copyleaks, rank 3Best for: Institutions with Spanish or French submissions who want the one card with a per-language translation result. Main limitation: The Spanish result is 10 of 10 human theses flagged, the sharpest non-English false positive figure in the index.
- ZeroGPT, rank 4Best for: Readers who want a zero on translated human text from a free tool. Main limitation: The Liang 2023 per-tool rate is a chart; the Weber-Wulff zero covers 9 translated documents.
- Turnitin, rank 5Best for: Institutions with many non-native writers who want two zeros on human text. Main limitation: Both zeros are on human text; the same Van Vlasselaer paper scored all 40 fully AI papers under 20%.
What does the evidence say, tool by tool?
GPTZero
- Verdict
- Fits readers who want five records on text outside native English, with results that run in both directions.
- Value
- 5 records on non-English or non-native text: the five records print a false accusation count on translated human text, a miss count on translated AI text, a zero on 40 non-native human papers, an all-human verdict on 10 Spanish ChatGPT texts, and a non-native TOEFL rate shown only in a chart; records read 8 and 9 November 2026.
- Best for
- Readers who want the widest non-English record, whichever way it points
- Main limitation
- Five records point both ways: translated human text was flagged, translated AI text was missed, and the TOEFL per-tool rate is printed only as a chart
GPT-2 Output Detector
- Verdict
- Fits researchers who want the baseline's record on ESL, translated and Spanish text.
- Value
- 4 records on non-English or non-native text: two of its 18 human documents were flagged, both machine-translated (Weber-Wulff 2023, Table 9); 16 of 120 ESL essays fell under a 95% originality cut (Ibrahim 2023); every Spanish ChatGPT text was called human (Orenstrakh 2023); records read 8 and 9 November 2026.
- Best for
- Researchers comparing a 2019 baseline on non-native and translated text
- Main limitation
- A GPT-2 era model: Ibrahim's 13.3% carries the author's inverted label and is not entered as a rate
Copyleaks
- Verdict
- Fits institutions with Spanish, French or German submissions who want the one card with per-language results.
- Value
- 3 records on non-English or non-native text: Orenstrakh 2023 flagged all 10 Spanish human theses and all 10 Spanish ChatGPT texts (Table 12); Chaka 2023 classified 3 of 6 German, 5 of 6 French and 6 of 6 Spanish translated responses as AI; Van Vlasselaer 2026 flagged 0 of 40 human papers by non-native graduate students; records read 8 and 9 November 2026.
- Best for
- Institutions with Spanish or French submissions who want the one card with a per-language translation result
- Main limitation
- The Spanish result is 10 of 10 human theses flagged, the sharpest non-English false positive figure in the index
ZeroGPT
- Verdict
- Fits readers who want a free tool's zero on translated human text.
- Value
- 2 records on non-English or non-native text: 0 false accusations in 18 human documents, 9 of them machine-translated (Weber-Wulff 2023, Table 9); the Liang 2023 TOEFL rate is printed only in Figure 1a; records read 8 and 9 November 2026.
- Best for
- Readers who want a zero on translated human text from a free tool
- Main limitation
- The Liang 2023 per-tool rate is a chart; the Weber-Wulff zero covers 9 translated documents
Turnitin
- Verdict
- Fits institutions with many non-native writers who want two zeros on human text.
- Value
- 2 records on non-English or non-native text: 0 of 18 human documents flagged, 9 machine-translated (Weber-Wulff 2023, Table 9), and 0 of 40 non-native human papers flagged (Van Vlasselaer 2026, Table 5); records read 8 and 9 November 2026.
- Best for
- Institutions with many non-native writers who want two zeros on human text
- Main limitation
- Both zeros are on human text; the same Van Vlasselaer paper scored all 40 fully AI papers under 20%
Originality.ai
- Verdict
- Fits readers who want to know the card was in the TOEFL study.
- Value
- 1 record on non-English or non-native text: one of seven detectors on 91 TOEFL essays by non-native writers, per-tool rate in a chart only (Liang 2023, Figure 1a); records read 8 and 9 November 2026.
- Best for
- Readers who want the card named in the TOEFL study
- Main limitation
- No number: the per-tool TOEFL rate is a bar in Figure 1a
Writer
- Verdict
- Fits readers of the 2023 record on a tool that no longer runs.
- Value
- 1 record on non-English or non-native text: 1 false accusation in 18 human documents, on a machine-translated text (Weber-Wulff 2023, Table 9); records read 8 and 9 November 2026.
- Best for
- Readers of the 2023 record on the withdrawn tool
- Main limitation
- The tool was withdrawn on 22 December 2025
DetectGPT
- Verdict
- Fits researchers who want the method's own translated-text split.
- Value
- 1 record on non-English or non-native text: 5 of 9 machine-translated human documents classified correctly against 9 of 9 untranslated (Weber-Wulff 2023, per-class table); records read 8 and 9 November 2026.
- Best for
- Researchers who want the method's translated-text split
- Main limitation
- A research method run through the 2023 demo; 4 of 9 translated texts misread
GLTR
- Verdict
- Fits researchers who want the visual tool's Spanish result.
- Value
- 1 record on non-English or non-native text: every one of the 10 Spanish ChatGPT submissions was called human (Orenstrakh 2023, Section 5); records read 8 and 9 November 2026.
- Best for
- Researchers who want the visual tool's Spanish result
- Main limitation
- One record, and it is a miss on all 10 Spanish AI texts
Pangram
- Verdict
- Fits teams that want a 2026 zero on non-native academic prose.
- Value
- 1 record on non-English or non-native text: 0 of 40 human papers by non-native English-speaking graduate students flagged, at least 4,000 words each (Van Vlasselaer 2026, Table 5); records read 8 and 9 November 2026.
- Best for
- Teams that want a zero on non-native academic prose from a 2026 record
- Main limitation
- One record; the vendor's multilingual figures sit on its model card and are excluded here
BrandWell
- Verdict
- Fits readers who want the translated zero in its context.
- Value
- 1 record on non-English or non-native text: 0 false accusations in 18 human documents, 9 machine-translated, beside 36 of 36 AI documents undetected (Weber-Wulff 2023, Tables 9 and 10); records read 8 and 9 November 2026.
- Best for
- Readers who want the translated-text zero beside the 36 of 36 misses
- Main limitation
- The same record missed all 36 AI documents
Sapling
- Verdict
- Fits readers who want to know the card was in the TOEFL study.
- Value
- 1 record on non-English or non-native text: one of seven detectors on 91 TOEFL essays by non-native writers, per-tool rate in a chart only (Liang 2023, Figure 1a); records read 8 and 9 November 2026.
- Best for
- Readers who want the card named in the TOEFL study
- Main limitation
- No number printed for the tool
Winston AI
- Verdict
- Fits readers who want a translated-text zero from a card that states 11 languages.
- Value
- 1 record on non-English or non-native text: 0 false accusations in 18 human documents, 9 machine-translated (Weber-Wulff 2023, Table 9); records read 8 and 9 November 2026.
- Best for
- Readers who want a translated-text zero from a card stating 11 languages
- Main limitation
- One record, from 2023
Compilatio
- Verdict
- Fits European institutions who want a translated-text zero from a vendor stating 35 languages.
- Value
- 1 record on non-English or non-native text: 0 false accusations in 18 human documents, 9 machine-translated, on the free tool in March 2023 (Weber-Wulff 2023, Table 9); records read 8 and 9 November 2026.
- Best for
- European institutions who want a translated-text zero from a vendor stating 35 languages
- Main limitation
- One record, on the 2023 free tool; 7 of 9 paraphrased texts missed
GPTKit
- Verdict
- Fits readers who want one Spanish result for a detector whose vendor claims English only.
- Value
- 1 record on non-English text: Orenstrakh 2023 ran 10 human and 10 ChatGPT submissions in Spanish and GPTKit classified 80% and 20% correctly (arXiv 2307.07411, Table 12, read 9 November 2026); the vendor's FAQ states "Currently English is supported".
- Best for
- Readers who want to see a detector tested outside its claimed scope.
- Main limitation
- 10 texts per set, the paper's smallest sample; no other record tested GPTKit in another language.
PlagiarismCheck
- Verdict
- Fits readers who want to know how the suite's detector read machine-translated human text.
- Value
- 1 record: Weber-Wulff 2023's 18 human documents include 9 written in Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish and machine-translated into English (class 02-MT), and PlagiarismCheck made 0 false accusations across the 18 (arXiv 2306.15666, Table 9, read 9 November 2026), the same counting the ranking applies to the other eight cards in that study.
- Best for
- Institutions with many students writing in a second language who weigh a false accusation first.
- Main limitation
- 9 translated documents; the vendor's stated Spanish support has no indexed result, because the study tested English text.
Writefull
- Verdict
- Fits readers who want to know how a discontinued tool read machine-translated human text.
- Value
- 1 record: Weber-Wulff 2023's 18 human documents include 9 machine-translated into English (class 02-MT), and Writefull made 1 false accusation across the 18 (arXiv 2306.15666, Table 9, read 10 November 2026), the same counting the ranking applies to the other cards in that study.
- Best for
- Readers reading a 2023 Writefull percentage on a second-language writer's text.
- Main limitation
- 9 translated documents; the vendor stated no language list and discontinued the tool before 7 March 2025.
Check For AI
- Verdict
- Fits readers who want to know how an English-only tool read machine-translated human text.
- Value
- 1 record: Weber-Wulff 2023's 18 human documents include 9 machine-translated into English (class 02-MT), and Check For AI made 1 false accusation across the 18 (arXiv 2306.15666, Table 9, read 10 November 2026).
- Best for
- Readers reading a March 2023 risk score on a translated text.
- Main limitation
- 9 translated documents; the vendor's learn more page stated "Our models are only trained for English text" and the service was sunset in July 2023 (web.archive.org, learn more page, 11 December 2023, read 10 November 2026).
AI Writing Check
- Verdict
- Fits readers who want the one non-native English record of a withdrawn nonprofit tool.
- Value
- 1 record: Liang 2023, where the tool was one of seven whose average false positive rate on 91 TOEFL essays was 61.22% against 5.19% on 88 US eighth-grade essays, accessed 15 March 2023 (arXiv 2304.02819v3, read 10 November 2026).
- Best for
- Teachers reading a spring 2023 verdict on an English learner's essay.
- Main limitation
- The paper prints per-tool rates only in a chart, so the card carries the averaged figure; the tool has been withdrawn since August 2023 (aiwritingcheck.org, read 10 November 2026).
How we ranked
- Field sorted
- Count of indexed study records that print a result for the card on non-English text, machine-translated text or English by non-native writers; vendor multilingual figures excluded
- Direction
- Higher first
- Tie rule
- More study records in all first, then name
- Qualification
- At least one such record; a record that names a language without scoring the card on it does not count
- Qualified
- 19 of 72 detectors
Field sorted: the non-English records field on each card, written on 9 November 2026 from the study records in the data file, each read in the primary source on 8 or 9 November 2026. A record counts for a card when it prints a result for that card on text that is not native English prose, in one of three forms: written in another language (Orenstrakh 2023's Spanish theses and ChatGPT texts; Chaka 2023's translations into German, French, Spanish, Southern Sotho and isiZulu); machine-translated into English from another language (Weber-Wulff 2023's class 02-MT, 9 of the 18 human documents, so every card's false accusation count in that record covers the translated half); or written in English by non-native writers (Liang 2023's 91 TOEFL essays, Ibrahim 2023's 120 ESL essays, Van Vlasselaer 2026's 40 human papers by non-native graduate students). A record that names a language without scoring a card on it does not count: RAID's Czech and German news in RAID-extra, Perkins 2024's "non-native English voice" as an evasion prompt applied to AI text, Jabarian and Imas 2025 and Epoch AI 2026, which are English-only. Vendor measurements are excluded (Pangram's multilingual model card, Copyleaks' six-language table, Originality.ai's 30-language internal test). Direction: more records rank higher. Tie rule: more study records in all first, then product name, so ZeroGPT (7 records) sits above Turnitin (5) at two, and the nine cards at one run Originality.ai (7), Writer (5), DetectGPT (4), GLTR (4), Pangram (4), BrandWell (3), Sapling (3), Winston AI (2), Compilatio (1). 19 of 72 qualify. Full rules: methodology page. The count says how many records exist, not whether they favour the tool; the result column prints what each record found.
Why numbers differ between sources
The languages ranking and this page describe different things with similar words. The languages ranking counts what a vendor states: Smodin 100, Resemble Detect 54, Undetectable AI 50, Pindrop Pulse 40, Compilatio 35, read 8 and 9 November 2026. This page counts what a study printed, and none of those five vendors except Compilatio has a record on any non-English text; Compilatio's one record covers machine-translated text, not any of its 35 languages by name (DOI 10.1007/s40979-023-00146-z, read 8 November 2026). Liang 2023 is counted for five cards although it prints no per-tool number: the per-detector rates are bars in Figure 1a and the paper states a 61.22% average across seven tools (DOI 10.1016/j.patter.2023.100779, read 8 November 2026), so the record counts as evidence while the result column says "chart only". Weber-Wulff 2023 counts for nine cards because its 18 human documents include 9 machine-translated ones and the per-tool false accusation count covers them; the record prints the translated split for GPTZero (6 of 9 false accusations), the GPT-2 Output Detector (both), Writer (one) and DetectGPT (5 of 9 correct) and a zero for the other five (arXiv 2306.15666, Table 9, read 8 November 2026). Ibrahim 2023's 13.3% carries the author's label "false negatives" for human essays scored below a 95% originality cut and is quoted as printed, not entered as a false positive rate (DOI 10.1186/s40468-023-00260-2, read 9 November 2026). All read 8 and 9 November 2026; the Updated line is the date the page last changed.
Turnitin's 2 records here cover machine-translated English and non-native English, and none of its six non-English models has an indexed record, as the Turnitin languages question, added 9 November 2026, sets out. The localisation teams view, added the same day, pairs this ranking's counts with the 14 cards that state 10 or more languages.
Questions about AI detectors and non-English text
Which AI detector has the most study evidence on non-English text?
GPTZero, with five records: Weber-Wulff 2023, Liang 2023, Orenstrakh 2023, Chaka 2023 and Van Vlasselaer 2026, read 8 and 9 November 2026 and listed above. The GPT-2 Output Detector has four and Copyleaks three.
Which AI detectors have been tested on Spanish text?
Four, in Orenstrakh 2023: Copyleaks flagged all 10 Spanish human theses and all 10 Spanish ChatGPT texts, while GPTZero, the GPT-2 Output Detector and GLTR called every Spanish ChatGPT text human (arXiv 2307.07411, Table 12, read 9 November 2026). Chaka 2023 adds Spanish translations: Copyleaks classified 6 of 6 as AI and GPTZero 0 of 6 (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026).
Which AI detectors have been tested on non-native English writing?
Eight: Liang 2023 ran GPTZero, the GPT-2 Output Detector, ZeroGPT, Originality.ai and Sapling on 91 TOEFL essays with a 61.22% average false positive rate across seven tools (DOI 10.1016/j.patter.2023.100779, read 8 November 2026); Ibrahim 2023 scored the GPT-2 Output Detector on 120 ESL essays, 16 of them under a 95% originality cut (DOI 10.1186/s40468-023-00260-2, read 9 November 2026); Van Vlasselaer 2026 found 0 of 40 non-native human papers flagged by GPTZero, Pangram, Copyleaks and Turnitin (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).
Does a vendor's language count predict its study record?
No. Smodin states 100 languages and has no record, Undetectable AI 50 and none, Pangram 24 and one, read 8 and 9 November 2026; Weber-Wulff 2023 tested every tool on machine-translated English, not on the languages the vendors state. The languages ranking and this page share nine cards.
Why does machine-translated English count as non-English evidence?
Because the texts were written in Bosnian, Czech, German, Latvian, Slovak, Spanish or Swedish and reached the detector through DeepL or Google Translate, and Weber-Wulff 2023 reports that accuracy fell by 20% on that class and that most false accusations fell on it (DOI 10.1007/s40979-023-00146-z, read 8 November 2026). The machine-translated text question reads the class in full.
Why are Pangram's and Copyleaks' multilingual figures left out?
Because the vendor measured them: Pangram's model card and Copyleaks' six-language accuracy table are vendor benchmarks, recorded on the vendor benchmarks record and the cards, read 8 and 9 November 2026, and this ranking counts outside studies only, by the rule the independently tested ranking uses.
Which of these records is the most recent?
Van Vlasselaer 2026, published 29 June 2026, which scored GPTZero, Pangram, Copyleaks and Turnitin on 40 human papers by non-native English-speaking graduate students and flagged none (DOI 10.1007/s40979-026-00226-w, read 9 November 2026). The other five records date from 2023.
How often is this ranking updated?
The ranking was created on 9 November 2026 from the six records above, read 8 and 9 November 2026, and changes whenever a study record that prints a non-English or non-native result is added to the index. New records are read in the primary source before they are indexed (methodology).
Sources and revisions
Sources
- Weber-Wulff et al. 2023, Testing of detection tools for AI-generated textdoi.org/10.1007/s40979-023-00146-z
- Weber-Wulff et al. 2023, arXiv version with per-class tablesarxiv.org/pdf/2306.15666
- Liang et al. 2023, GPT detectors are biased against non-native English writersdoi.org/10.1016/j.patter.2023.100779
- Orenstrakh and Liut 2023, Detecting LLM-generated text in computing educationarxiv.org/abs/2307.07411
- Chaka 2023, Detecting AI content in responses generated by ChatGPT, YouChat, and Chatsonicdoi.org/10.37074/jalt.2023.6.2.12
- Ibrahim 2023, Using AI-based detectors to control AI-assisted plagiarism in ESL writingdoi.org/10.1186/s40468-023-00260-2
- Van Vlasselaer, Van Droogenbroeck and Spruyt 2026, Who wrote this?doi.org/10.1007/s40979-026-00226-w
Revision log
- Ranking created from the non-english records field written at index version 2026
- Eighth-wave pages linked
- Nine wave 9 cards added
- Twelve wave 10 cards added
- Nine wave 11 cards added
). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify.: register plate](/img/independently-tested-480.webp)
