AI or Not says its detector identified all 103 original Meta AI images in a company benchmark, and 98.0% of 102 cropped or tampered versions, the company said on 20 July 2026.
In the same test, Meta AI’s native content-detection labelling identified 98.1% of the originals but 35.3% of the cropped or tampered images, according to AI or Not.
The results varied by content type for Meta’s labelling: 9.6% of cropped landscape and scene images were identified, against 62.0% of cropped portraits, food and objects.
The benchmark was run by AI or Not itself, not an independent tester. The company pointed to a Reuters analysis published on 10 July 2026, which found that Meta’s detector failed to identify some cropped AI images.
Cropping and re-encoding can strip watermarks and metadata, which is why the company argues detection should be trained against manipulation.