Which AI detector works best for Hungarian?
We tested seven well-known AI detectors on the same 1,000 Hungarian texts. Preds 3, a detector built specifically for Hungarian, came first with 99.6% accuracy and did not flag a single one of the 500 human-written texts as AI.
The dataset is open (CC BY 4.0) and free to download from Hugging Face, so anyone can repeat the test.
All results in one table
Accuracy is the share of texts classified correctly. FPR is the share of human texts flagged as AI, and FNR is the share of AI texts that passed as human. For those two, lower is better.
| 1 | Preds 3 Top result | 99.6% | 99.60% | 0.9996 | 0.0% | 0.8% |
|---|---|---|---|---|---|---|
| 2 | GPTZero | 90.1% | 89.04% | 0.9923 | 0.2% | 19.6% |
| 3 | Originality.ai | 82.0% | 84.29% | 0.9574 | 32.6% | 3.4% |
| 4 | Plag.hu | 76.5% | 69.44% | 0.9480 | 0.4% | 46.6% |
| 5 | isgen.ai | 65.5% | 48.43% | 0.7317 | 1.4% | 67.6% |
| 6 | ZeroGPT | 64.7% | 64.52% | 0.6911 | 34.8% | 35.8% |
| 7 | Smodin | 62.9% | 41.39% | 0.7655 | 0.4% | 73.8% |
Click a column name to sort. Full results are available as CSV and JSON. Measured in September 2026.
Disclosure: the dataset and the benchmark were created by the developer of Preds. That is why every text, label and result is public, so anyone can check them.
What the numbers mean
Percentages hide real people. These are the raw counts behind them, out of 500 human-written and 500 AI-written texts.
Zero false accusations
Preds 3 flagged none of the 500 human-written texts as AI. ZeroGPT flagged 174 and Originality.ai flagged 163, roughly one in three.
Almost nothing slips through
Preds 3 missed only 4 of the 500 AI texts. Smodin missed 369, isgen.ai 338 and Plag.hu 233, even though they rarely flagged humans.
Hungarian training data matters
ZeroGPT, isgen.ai and Smodin score an AUROC between 0.69 and 0.77, so their scores barely separate human from AI text. No threshold can fix that, and it usually points to little Hungarian training data. Preds 3, built for Hungarian, scores 0.9996.
Transparent and repeatable
The 1,000 texts come from 9 domains, with the same number of human and AI texts in each, so the topic cannot give the label away. Lengths are matched within each domain, from 80 to 995 words, with a median of 198.
The AI texts were generated fresh for this test and were never used to train Preds, so our own model had no head start.
AI texts were written by
- GPT-5.6
- Gemini 3.8 Flash
- DeepSeek V4 Flash
- Kimi K2.6
- Llama 4 Scout
- Qwen 3.7 Flash
- GLM 5.3 Flash
- Mistral Small
- Claude Haiku 4.5
- Grok 4.3
- gpt-oss-120b
Domains
- Creative writing
- Essays and academic
- General web
- Technical and IT
- News
- Reference and education
- Legal and administrative
- Social and Q&A
- Science and medicine
Common questions
What is the best AI detector for Hungarian?
On the open HuDetect-1k benchmark, Preds 3 led every metric: 99.6% accuracy, a 0% false positive rate and 0.9996 AUROC. GPTZero came second with 90.1% accuracy.
Do English-first AI detectors work on Hungarian text?
Results vary widely. In the benchmark, ZeroGPT and Originality.ai flagged roughly one in three human-written Hungarian texts as AI, while Smodin and isgen.ai missed most AI-written texts.
Who ran the benchmark?
The dataset and the benchmark were created by the developer of Preds. Because Preds is one of the detectors compared, every text, label and per-detector result is public, so anyone can check and repeat the test.
Check the results yourself
The 1,000 texts, the labels and the per-detector results are public. Download the dataset, run any detector on it and compare the numbers.