How we evaluate AI detection
A detector you'll defend to a student, an editor or a regulator has to show its work. Here's how we test, what we measure, and where the limits are.
How we test
We evaluate against a balanced benchmark of human-written and AI-generated text. The human set is drawn from sources predating widespread generative AI and from verified original writing; the AI set is produced across multiple models and prompting styles, including paraphrased and “humanized” variants designed to evade detection.
Crucially, the benchmark is built per language. We don't translate test data into English - each Romance language is evaluated on native text, because translation erases the signals detection relies on.
Models in the benchmark
GPT-4o · Gemini · Claude · Copilot · Llama · Mistral
Why false positives matter most
For the people who rely on detection - teachers, editors and compliance teams - the costly error is wrongly accusing a human. Results should therefore be reviewed with sentence-level evidence, drafts, version history and context.
Honest limitations
No detector is infallible, and anyone who claims otherwise should worry you. Our results are weakest on:
- Very short text. A sentence or two rarely carries enough signal for a confident verdict.
- Heavily edited hybrids. Human text rewritten sentence-by-sentence with AI sits in a genuine grey zone.
- Highly formulaic genres. Boilerplate, legal and technical templates can read as machine-like even when human-written.
We recommend treating a score as strong evidence to investigate - reviewed alongside drafts, version history and context - not as standalone proof.
See the detector behind the numbers
Tell us about your use case, volume, language, and privacy requirements.