AI detector benchmark data
All of our AI detector test data is free to download. It covers 900 texts with their sources and licences, every score each of the seven detectors gave each text, and every published result with its 95% interval. Our scores and results are CC BY 4.0, and each text keeps its own licence.
Tested
Downloads
- results.json: Every published number: rates with 95% intervals, AUROC, breakdowns by model, kind of text and length.
- scores.csv: Every detector's raw score for every text, with the AI/human decision and the split it belongs to.
- texts.jsonl: Every text in the benchmark with its source, date and licence. Human texts keep their original licences.
What is in each file
results.json holds every number shown on this site, with the count behind each rate and its 95% Wilson interval, broken down by detector, AI model, kind of writing and text length.
scores.csv has one row per text and detector: the raw score, whether the detector flagged the text as AI at its threshold, and the split. You can recompute any rate on the site from this file, or try your own thresholds.
texts.jsonl has one line per text with the full text, its kind and genre, word count and split. Human texts carry their source address, publication date, licence and author credit. AI texts carry the model, the exact prompt and the human text it was matched to.
Licences
Our scores, results and metadata are under CC BY 4.0. The human texts keep the licences of their sources: CC BY-SA for Wikipedia and Stack Exchange, CC BY 2.5 for Wikinews, CC0 for arXiv abstracts and public domain for Project Gutenberg passages. The methodology explains how every text was chosen.
Questions people ask
Can I use this data in an article or paper?
Yes. Credit Tool Benchmarks with a link to this page. Human texts from Wikipedia and Stack Exchange are CC BY-SA, so excerpts you republish must keep that licence and credit their authors; the source address of each text is in the file.
How do I tell which texts were used for calibration?
Every row in scores.csv and texts.jsonl has a split field, calibration or test. Published numbers use the test split only.