What a false positive is, and why it matters more than accuracy

A false positive is a piece of human writing that a detector labels as AI-generated. In a classroom, it can mean a student defending work they wrote themselves. In publishing, it can cost a freelancer a client. It's the mistake we work hardest to avoid.

Why they happen

Detectors learn the statistical fingerprints of model-written text: predictable word choices, an even sentence rhythm, a narrow range of structures. Some human writing shares those traits, especially when it's formal, formulaic or heavily polished.

• Very short submissions, where there isn't enough text to judge.
• Highly templated writing, like lab reports, legal summaries or product descriptions.
• Writers working in a second language, who may favor safer, more common phrasing.
• Drafts run through aggressive grammar or rewriting tools.

How we keep them rare

We train on a large, deliberately varied set of human writing, including student essays, non-native English, technical documentation and transcribed speech. We tune for a low false positive rate first and accept a small cost in recall to get there. On our current benchmark, 0.6% of human documents are wrongly flagged.

If you think a result is wrong

Look at which sentences were highlighted and whether they share a pattern, like a quoted definition or a boilerplate introduction. Check the confidence band. Then rely on what you already know about the writer and their process.

If you believe a scan is wrong, report it from the results page. Every reported case is reviewed by a person and goes into the set we test each new model against.