Overview & Key Findings
AI text detectors are routinely benchmarked in English, including in our own 10-tool English accuracy benchmark. The non-English picture is far less studied, and for most detectors it is substantially worse. This benchmark tests two widely-used AI detection tools, AITextDetector.ai (AITD) and GPTZero, across 35 non-English languages using a standardised corpus of AI-generated and human-written passages.
The 35 languages span Indic scripts (Hindi, Bengali, Tamil, Telugu, Gujarati, Punjabi, Marathi, Urdu), East Asian (Chinese Simplified & Traditional, Japanese, Korean), Southeast Asian (Indonesian, Filipino, Malay, Vietnamese, Thai), Middle Eastern (Arabic, Persian, Hebrew), European (Italian, Polish, Czech, Romanian, Hungarian, Ukrainian, Bulgarian, Norwegian, Finnish, Catalan, Croatian, Slovak), and Sub-Saharan African (Swahili).
- AITD achieves 90% AI-recall across all 35 languages. GPTZero achieves 71%. The 19-point gap is almost entirely attributable to GPTZero’s tokenizer and training data failures in non-Latin scripts.
- GPTZero has four distinct failure modes covering approximately 20 of 35 tested languages: tokenizer collapse, under-trained language coverage, wrong-signed verdicts, and a small set of correctly-functioning languages.
- The two detectors produce nearly independent signals: Pearson r = 0.06 across 349 passages. Using both together provides genuine redundancy.
- AITD wins 20 of 35 languages outright. GPTZero wins 7. Eight languages are tied on combined accuracy.
- GPTZero’s human-specificity is unreliable outside its well-trained languages. In 8 languages including Chinese, Polish, and Indonesian, it returns near-100 on every input regardless of whether the text is AI or human.
Methodology
For each of 35 languages we created 10 passages: 4 AI-generated (Llama-3.3-70B, register-appropriate prompts), 4 human-written (pre-LLM-era public domain sources), and 2 mixed. Passage length ranges from approximately 70 to 320 words.
AITD queried via /detect. GPTZero queried via v2/predict/text with multilingual:true. Both at default detection thresholds. Metrics: AI-recall = proportion of AI passages scoring ≥50; human-specificity = proportion of human passages scoring <50; combined accuracy = (AI-recall + human-specificity) ÷ 2. A combined score of 50% = chance.
Overall Accuracy: AI-Recall & Human-Specificity
Averaged across all 35 languages. Higher is better on both axes. A detector that flags everything as AI scores 100% AI-recall but 0% human-specificity: combined accuracy of 50%, no better than chance.
Per-Language Head-to-Head
All 35 languages sorted by AITD combined accuracy descending. Combined = (AI-recall + human-specificity) ÷ 2. Colour coding: green ≥75%, amber ≥62%, red <62%.
| Language | AITD AI-rec. | AITD Hu-spec. | AITD Combined | GPTZero Combined | Comparison | Winner |
|---|
Where GPTZero Quietly Breaks
In approximately 20 of the 35 languages, GPTZero returns degenerate outputs. We observed four distinct failure modes.
Score Correlation: Are the Detectors Independent?
Each point is one passage. X-axis = AITD score; Y-axis = GPTZero score. If the detectors agreed, points would cluster on the diagonal. The cloud shows they do not: Pearson r = 0.06.
Practical Implications
Multilingual academic institutions
An educator using GPTZero to check Chinese, Hindi, or Indonesian student submissions is not receiving valid detection signals: the tool is returning a constant value with no relationship to actual content. This risk is invisible unless you test the tool against your specific language corpus before deployment. Any detector’s published accuracy figure should be treated as an English-only figure unless the vendor explicitly publishes per-language benchmark data.
Content and publishing teams
A tool that achieves 87% accuracy on English text may achieve 50% (random) accuracy on Chinese text while returning the same confidence score format. Platforms verifying content in multiple languages (particularly Asian, African, or Middle Eastern markets) cannot rely on English-calibrated detection thresholds. Test any detector against a sample corpus in every language you use it for before committing to production.
Developers using the API
The near-zero correlation (r = 0.06) means AITD and GPTZero can be combined in an ensemble. A simple voting rule (flag content only when both detectors agree it is AI) will reduce false positives substantially at the cost of some AI-recall. For high-stakes use cases this trade-off is generally appropriate. See the API documentation for integration examples.
Limitations
- Small per-language corpus. 10 passages per language is a minimal sample. Per-language results have wide confidence intervals. The overall 35-language averages and structural failure modes are robust; individual language rankings are directional only.
- Single AI model source. AI passages generated by Llama-3.3-70B only. GPT-4o, Claude, and Gemini-generated text in the same languages may produce different recall figures.
- First-party study. Conducted by the AITextDetector.ai team. Our detector performed better overall. Independent replication would substantially increase confidence in both absolute figures and the comparative ranking.
- GPTZero version at test date. Results reflect the
v2/predict/textendpoint queried on 12 May 2026. GPTZero may improve multilingual coverage in future releases. - No humanized text. Both detectors perform substantially worse on AI text processed by humanizer tools. See our bypass study data.
Frequently Asked Questions
Which AI detector is best for multilingual use?
Based on this benchmark, AITextDetector.ai outperforms GPTZero in multilingual contexts, winning 20 of 35 languages with a 19-point advantage in AI-recall. However both tools have limitations in specific languages. For any critical multilingual deployment, test the specific tool against a sample corpus in your target language before production use.
Does GPTZero work in Chinese, Japanese, or Korean?
Based on our testing with multilingual:true, GPTZero experiences tokenizer collapse in Chinese (Simplified and Traditional), Japanese, and Korean, returning approximately 100 on every input regardless of whether the text is AI-generated or human-written. It cannot meaningfully distinguish between them in these languages.
What is a wrong-signed verdict?
A wrong-signed verdict occurs when a detector’s mean score for human-written text is higher than its mean score for AI-written text in a given language. In Italian, Bulgarian, Ukrainian, Urdu, Czech, and Arabic, GPTZero’s human passages received higher scores than AI passages on average: the detector actively points in the wrong direction in those languages.
Why is the score correlation so low (Pearson r = 0.06)?
GPTZero collapses to constant values (~100 or ~0) in approximately half the languages tested, while AITD produces variable scores. This structural difference creates near-zero overall correlation. On passages in languages where both detectors function correctly, correlation is higher. The low r confirms genuine independence: the two APIs can be combined in an ensemble for real redundancy.
How should I read the combined accuracy metric?
Combined accuracy = (AI-recall + human-specificity) ÷ 2. A score of 50% means the detector is at chance level. Scores above 70% are reliable for screening use; above 85% suitable for higher-stakes applications. Note: combined accuracy of 50% can arise from two very different failures: a detector returning constant 100 (high AI-recall, zero human-specificity) and one returning constant 0 (zero AI-recall, perfect human-specificity). The individual AI-recall and human-specificity columns tell the complete story.
/detect, GPTZero via v2/predict/text (multilingual:true). Length 70–320 words. Test date: 12 May 2026. 349 passages, 700 detector calls.