Meta's new AI image detector, previewed alongside its Muse Image generator, failed to identify a significant portion of its own AI-generated images after they were cropped, according to a Reuters analysis.
The tool successfully verified all 40 original images created by Meta's Muse Image model. However, it failed to verify 55 percent of the same images once cropped to between one-third and one-half their original size.
Meta's system relies on an invisible watermark called Content Seal, embedded in every Muse Image. The company stated the preview tool is designed to withstand common edits but acknowledged the watermark signal may be lost with heavy cropping.
This limitation surfaces during a critical election year, highlighting challenges in combating deepfakes. Rivals Google and OpenAI have similarly warned their detection tools are not foolproof.
Meta's own Oversight Board recently called for stronger investment in detection tools to address deceptive AI content. Experts note all watermark-based systems have vulnerabilities.
"Any modification that removes or weakens the embedded signal-such as cropping-may reduce their effectiveness," said Siwei Lyu, a computer science professor at SUNY Buffalo.
Researchers see watermarking as a promising, if imperfect, step forward. "Even if we catch only 90 percent of cases, that’s still a great leap from 0," said Sarah Barrington, an AI researcher at UC Berkeley.