Meta AI detector misses over half of cropped images

Meta's newly previewed tool for detecting AI-generated images has shown significant limitations, failing to identify more than half of its own company's images after they were cropped, according to an analysis published on July -10. The finding underscores the persistent technical challenges in verifying digital content as artificial intelligence becomes more sophisticated.

The detection tool was introduced this week alongside Meta's first image-generation model, called Muse Image. The company announced that all images created by Muse Image would include an invisible watermarking system named Content Seal, designed to remain detectable even after common edits like cropping, compression, resizing, or screenshots. The accompanying preview tool was intended to allow users to check for that watermark.

However, an analysis by Reuters found the tool's performance dropped sharply when images were altered. In tests of 40 images generated by Muse Image, the tool correctly identified all original versions as AI-generated. Yet after those same images were cropped to approximately one-third or one-half of their original size, the tool could only verify 55 percent of them.