Open AI-Text Detection

We're building open Pangram.

The most widely used AI-text detectors today are closed — you can't see how they were trained, audit their biases, or independently verify their accuracy claims.

What we're building

• 🧠 An open detector model, with weights, training code and recipes all public
• 📚 Open datasets of human and machine text across many models, domains and languages
• 📏 Honest benchmarks, including false-positive rates on non-native English writers, short texts and edited AI text

Principles

• Transparency We publish every score, including the bad ones.
• Low false positives first. Wrongly flagging a human does real harm.
• A signal, not a verdict. Detection should inform a human decision, never replace it.

Built in the open, on the Hub. 🤗