Detection
How accurate is it?
Accurate enough to be useful, not enough to be a verdict. Signatures and watermarks are near-certain; classifiers and writing patterns are probabilities the card always attributes.
Updated
Two kinds of evidence
Near-certain when it fires. A verified Content Credentials signature, a matching watermark check through a provider's own endpoint, or metadata that declares AI generation. These read what the file says about itself, cryptographically where possible. They rarely fire wrongly; their weakness is that platforms strip them, so they miss a lot.
Probabilities. The generator classifier, the model recheck and the writing-pattern analysis estimate. They catch stripped images and plain text but are wrong some of the time in both directions. The card always shows which detector produced a flag and its number.
Why this beats guessing
People are poor at this unaided. In a 2025 Royal Society Open Science study, typical adults picked the AI-generated face in a pair only 30% of the time, and trained super-recognisers managed 41%, both below chance. A tool that reads the metadata people cannot see, and shows its working, does better than a hunch.
What we do not publish
We do not publish a single accuracy percentage, because it would depend entirely on the mix of content and on which signals survived upload, and a headline number would be misused. The terms say plainly that verdicts are automated opinions about content, not facts about people.
Making it more accurate for you
- Use the default Balanced preset, or Strict if you want only the strongest evidence.
- Report wrong flags with Not AI so the community tally corrects them.
- Pro users can trigger the classifier and recheck on borderline posts, and every result is cached for everyone.