Reuters research shows that the Meta AI picture detector can’t find some of its own cropped AI images

A new AI detection tool from Meta, which the tech company previewed this week alongside the launch of its image-generation model, Muse Image, did not successfully identify some of its own AI-generated images after they were cropped, as reported by a Reuters analysis.

The finding underscores the difficulties in verifying AI-generated images following common modifications, a constraint that may complicate the identification of deepfakes online during a hectic election year that encompasses the U.S. midterms.

In an analysis of 40 images generated using Muse Image, Reuters discovered that the detection tool successfully verified all of the original AI-generated images. But it couldn’t verify 55% of the same images after cropping them to one-third to one-half of their original size.

On its website, Meta states that the preview detection tool can recognize its own AI-generated images, even when cropped, through an invisible watermarking system known as Content Seal. This system is embedded in every image produced by Muse Image and is intended to assist users in verifying whether the image was created by Meta’s AI models.

In response to inquiries regarding the findings of the Reuters analysis on the detection tool, Meta indicated that the tool was still in a preview phase. The company stated that typical edits should keep the watermark intact, but significant cropping may compromise it.

Rival tech companies Google and OpenAI have warned that their detection tools are not infallible when it comes to image-alteration techniques.

In March, Meta’s Oversight Board, a group of experts responsible for making binding decisions and providing recommendations on content issues across the company’s social media platforms, urged the company to take further action to tackle the “proliferation of deceptive AI-generated content” on its platforms and to invest in more robust detection tools.

Siwei Lyu, a computer science professor at the State University of New York at Buffalo who researches AI image forensics, mentioned that he had not assessed Meta’s tool but acknowledged the limitations of watermark-based systems. “Watermark-based methods can be highly effective when the watermark remains intact, but any modification that removes or weakens the embedded signal — such as cropping, resizing, heavy compression, or editing — may reduce their effectiveness, depending on how the watermark is designed,” Lyu stated.

Sarah Barrington, an AI researcher and Ph.D. candidate at the UC Berkeley School of Information, expressed that watermarking shows potential for the future of AI-generated content, but its effectiveness may be limited.

“Similar to various preventive cybersecurity or physical security measures, it may not be completely foolproof, but even if we manage to address only 90% of cases, that represents a significant improvement from zero,” she stated.