In response to Reuters' findings, Meta said the detector is only a preview and that the watermark is intended to survive common edits, but may be lost if an image is heavily cropped.
The Reuters report highlights the difficulty of reliably identifying AI-generated images after common alterations, a challenge that matters during a busy election year that includes the U.S. midterms.
Experts quoted in the story said watermark-based detection can be effective when the embedded signal remains intact, but cropping, resizing, compression, or editing can weaken it.
Meta's Oversight Board had earlier called on the company to do more to address the spread of deceptive AI-generated content and to invest in stronger detection tools.
A Reuters analysis found that Meta's new AI detection tool identified original images generated by its Muse Image model, but missed 55% of the same images after they were cropped to roughly one-third to one-half of their original size.
Meta previewed the detector this week alongside Muse Image, saying it uses an invisible watermarking system called Content Seal to help verify whether an image was created by Meta's AI models.
