Meta AI Image Detection Tool Fails in Cropped Image Tests
Meta's newly released tool for detecting AI-generated images has shown significant limitations. Reuters tests revealed the tool failed to identify images created by Meta's own AI model once they were cropped.
Meta Platforms' recently unveiled tool for detecting AI-generated images exhibits notable limitations, according to test results disclosed by Reuters. The tool struggled to identify images produced by Meta's own Muse Image model when those images were cropped.
Meta introduced the Muse Image model alongside the detection tool, claiming it could recognize content generated by Meta AI even after images were cropped. However, Reuters' tests demonstrated that common editing operations can weaken the detection capability.
Reuters generated 40 images using Muse Image. The detection tool successfully identified all original images. When the same images were cropped to approximately one-third to one-half of their original size, 55% of the images could not be verified as AI-generated.
The Muse Image model embeds an invisible watermark named Content Seal into each generated image. Meta has stated on its website that the detection tool can confirm if an image was generated by Meta's AI models through this watermark system, even after cropping.
Responding to Reuters' findings, Meta emphasized that the detection tool is still a preview version. The company acknowledged that while Content Seal is designed to retain its signal through common edits, significant cropping can still cause the signal to be lost.