mihalykiss/GenAIDetectHU

HuDetect-1k: a Hungarian AI-text detection benchmark (v1.3)

1,000 Hungarian documents, 500 written by people and 500 generated by 11 large language models, for evaluating AI-text detectors on Hungarian. Every text is wholly human or wholly machine-written.

Developed by Preds, the Hungarian AI detector.

Leaderboard

Seven detectors, same 1,000 documents, measured September 2026.

#	Detector	Accuracy	F1	AUROC	FPR	FNR
1	Preds 3	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%

FPR: human texts flagged as AI (lower is better). FNR: AI texts that passed as human (lower is better).

• Interactive leaderboard and submission instructions: Hungarian AI Detector Leaderboard
• Full write-up, charts and methodology (Hungarian): Magyar AI-detektor teszt
• Detailed article on preds.hu: Melyik a legjobb AI-detektor magyar nyelvre?

Disclosure: the dataset was built by the developer of Preds, one of the evaluated detectors. The AI texts were generated fresh and were not used to train Preds.

Generators

GPT-5.6 (88), Gemini 3.8 Flash (86), DeepSeek V4 Flash (58), Kimi K2.6 (58), Llama 4 Scout (56), Qwen 3.7 Flash (52), GLM 5.3 Flash (28), Mistral Small (27), Claude Haiku 4.5 (17), Grok 4.3 (16), gpt-oss-120b (14).

File

hudetect1k_v1.3.csv — UTF-8, RFC 4180 quoting, header + 1,000 rows.

column	content
text	the document (line breaks preserved)
label	0 = written by a person, 1 = generated by a language model
model	the generator for AI rows; human for human rows

Composition

Identical domain counts on both sides, so the domain cannot stand in for the label; lengths are matched within each domain (80–995 words, median 198).

domain	human	AI
creative writing	70	70
essays / academic	70	70
general web	70	70
technical / IT	70	70
news	60	60
reference / educational	60	60
legal / administrative	55	55
social / Q&A	30	30
scientific / medical	15	15

Quality and fairness

• Filters: 80–1,000 words, prose only (no tables or lists), no contact details, page boilerplate, machine metadata, English passages or repeated passages; Latin script.

Licence and attribution

Human documents remain under the terms of their sources; Wikipedia articles are CC BY-SA 4.0. The AI-generated texts, the labels and this compilation are released under CC BY 4.0.

Citation

@misc{hudetect1k2026,
  title        = {HuDetect-1k: Hungarian AI-text detection benchmark},
  author       = {Kiss, Mih{\'a}ly},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/mihalykiss/GenAIDetectHU}}
}