Meta tool stops recognizing AI images after cropping, says agency

Meta's logo.
Daniel Cole/Reuters
A new AI-generated content detection tool from Meta, introduced this week alongside its Muse Image image generation model, failed to identify some of the images created by the technology itself after they were cropped, according to a Reuters analysis.
The conclusion highlights the challenges of verifying AI-generated images after common edits, a limitation that could make it difficult to identify deepfakes on the internet during an intense electoral period in the States Unidos.
Do you have any reporting suggestions? Send to g1
In an analysis of 40 images generated with Muse Image, Reuters found that the tool correctly identified all original versions created by AI. However, it failed to recognize 55% of these images after they were cropped to about a third or half of their original size.
On its website, Meta claims that the preliminary version of the tool can identify images generated by its AI models even after cropping, thanks to an invisible watermarking system called Content Seal, built into all images produced by Muse Image.
Now on g1
The feature was developed to help users check whether an image was created by the company's artificial intelligence.
When asked about the results of the Reuters analysis, Meta highlighted that the tool is still in the preview phase. The company stated that the watermark was designed to resist common edits, but that the signal can be lost when an image undergoes more severe cropping.
Competitors Google and OpenAI have also warned that their detection tools are not capable of identifying all forms of image manipulation.
In March, Meta's Oversight Board - an independent body made up of experts that makes binding decisions and recommendations about content on the company's platforms - asked the company to expand its efforts to combat "proliferation of misleading AI-generated content." The group also advocated investment in more robust detection tools.
Siwei Lyu, a professor of computer science at the State University of New York at Buffalo and a researcher in the field of forensic analysis of AI-generated images, said he had not evaluated Meta's tool, but highlighted that watermark-based systems have limitations.
"Watermark-based methods can be highly effective when the signal remains intact. However, any modifications that remove or weaken it - such as cropping, resizing, intense compression or other edits - can reduce its effectiveness, depending on how the watermark was developed," Lyu said.
Sarah Barrington, an AI researcher and doctoral candidate at the University of California at Berkeley (UC Berkeley) School of Information, said watermarking technology holds promise for the future of AI-generated content, although it has limitations.
"Like many digital or physical security measures, this technology may not be completely foolproof. Still, even if it allows detection. only 90% of cases, this already represents a significant advance compared to having no identification mechanism", she said.

Source: G1