Detector accuracy

How accurate is the detector?

We test the detector on real and generated files every batch and publish the results with the sample size and the date. No single accuracy number, because the catch rate differs by generator, by file type, and by whether the file kept its signed record.

Free, no sign-up
We don't keep your files
Most checks finish in under a minute

What we publish and why

The accuracy page lists every generator we have tested, the number of files, the percentage that came back very likely or likely AI, and the test date. The false-alarm rate on held-out real files is shown separately, measured on photographs.

Debunk.org ran an independent test in 2024 of commercial detectors on 1,975 images that circulated online; the best scored 0.82. We do not claim to beat that number. We claim we publish our rates with the sample size and the date so you can check them.

What a verdict means and does not mean

A verdict is evidence, not proof. The detector reads the file's own labels and signed records, then judges the picture. A high score means the file shows the marks AI generators leave behind; it does not name the tool unless the file itself carries a signed record.

A low score means no sign of AI generation was found. It does not prove the file is real, because edits like face swaps or lip-sync can slip through, and some generators are harder to read from the image alone.

More on what we check, where it can go wrong and what happens to your file: how it works.

FAQ

Questions, answered honestly.

Something else? hello@findtheai.app

Why is there no single accuracy number?

The catch rate differs by generator. A file from one model may read very differently from a file from another. Publishing a single number would hide that difference, so we publish the per-generator rate with the sample size and the date instead. Per-generator rates

How do you measure the false-alarm rate?

We run the detector on held-out real files that were not part of the training set and report the percentage that came back likely or very likely AI. The current figure is on the accuracy page, with the sample size and the test date. Current false-alarm rate

Also