Meta this week unveiled a new detector intended to identify AI-generated images, announced alongside its Muse Image generative model. A Reuters analysis, however, found that the tool does not always recognize images produced by Meta once they are modified by common post-processing such as cropping.
Reuters examined 40 images created with Muse Image. The detector correctly identified all original AI-generated files, but failed to recognize 55% of the same images after they were cropped to roughly one-third or one-half of their original size.
Meta's response and the Content Seal
Meta notes on its website that the released tool is currently a preliminary, test-phase version. The company says Muse Image embeds an invisible watermark called Content Seal into every image it produces so users can verify whether a file was generated by Meta models. Meta states that the watermark survives typical edits, but acknowledges that severe cropping can remove the mark — a point underscored by Reuters' test results.
Competitors and expert views
Meta's competitors Google and OpenAI have similarly warned that their detection tools are not fully reliable against various image modification techniques.
Lju Szi-vej (Ljü Szi-vej), a computer science professor at the University at Buffalo, said watermark-based systems can be very effective while the marker remains intact, but any intervention — such as cropping, resizing, heavy compression, or editing — can weaken an embedded signal. Sarah Barrington, a researcher at the University of California, Berkeley, described watermarking as a promising approach for the future even if it is not yet perfect; she said that detecting 90 percent of cases would already be a major improvement over past capabilities.
Why this matters
The Reuters findings highlight how difficult it can be to detect AI-generated images after routine modifications. That limitation could complicate efforts to spot deepfakes and other manipulated content, a concern that gains urgency during dense political cycles such as U.S. midterm elections.
Source: Reuters. Lead image is illustrative.
