AI photo detector

What the detector is

The detector is a SigLIP 2 vision model (Google's, Apache-2.0) that its author fine-tuned on 60,000 AI pictures and 60,000 others. We converted it to run in the browser and shrank its weights to 4 bits, which makes it a 57 MB download (kept by your browser) that gives the same results as the full file within the margin we measured. It looks at the picture squeezed to 224 × 224 pixels, so it judges the overall look, not fine detail.

Detector score versus the other checks

• The detector says how much the picture resembles AI pictures it learned from. It can be wrong both ways.
• The file may say who made the picture. When it does, that outranks any score, and the detector is not even run.
• Camera records push toward a real photograph and can overrule a mild score; a very high score with camera records is reported as mixed signals.
• Size and noise are added with a small weight: generators use set sizes, and cameras leave fine noise.

How good it is, measured

Of five public detectors we ran on the same 249 pictures, the others scored 0.63 to 0.72 on AUC (0.5 is a coin toss); this one 0.90. Its author reports 99% accuracy on his own pictures. On ours it flagged 85% of AI pictures and 25% of real ones:

So on photographs it is useful, on drawings it is not (the page gives no score for mostly flat-color pictures), and it misses some photo-realistic services. The full chart is in the guide to comparing AI photo detectors.

Using it on photos

1. Drop in the photo, ideally the original file with its metadata.
2. Check whether the headline came from the file (certain) or from a score (estimate).
3. Look at the reasons: a high score because of the detector alone is weaker than one that agrees with a generator's size and missing noise.

A first run downloads the detector once; later pictures take about two seconds in Chrome and Safari, longer in Firefox. The page sends a copy of each picture and its result to our server (see the privacy policy).