A detection tool developed by Meta to spot its own AI-generated images fell short in a key test — failing to flag more than half of those images after they were simply cropped, a Reuters analysis has found.
The analysis examined 40 images created through Meta’s newly launched image-generation model, Muse Image. While the detection tool successfully identified all of the original, unaltered AI-generated images, it failed to recognize 55% of those same images after they were trimmed down to roughly one-third to one-half of their original size.
Meta’s website states that the preview detection tool can identify its own AI-generated content — even after cropping — through an invisible watermarking system known as Content Seal. That system is embedded in every image produced by Muse Image and is meant to help users confirm whether a photo was created by Meta’s AI.
When Reuters presented its findings to Meta, the company acknowledged that the tool is still in preview. Meta said the watermark is built to survive typical edits, but that heavy cropping can cause the embedded signal to be lost.
The issue carries broader implications. With a busy election year underway — including the U.S. midterms — the ability to detect manipulated AI images, or deepfakes, has become increasingly important. Other tech giants, including Google and OpenAI, have also acknowledged that their own detection tools are not perfect against image-alteration techniques.
Back in March, Meta’s Oversight Board — an independent body of experts that issues binding decisions and recommendations on content matters across Meta’s social media platforms — urged the company to take stronger action against what it called the “proliferation of deceptive AI-generated content” and to invest in better detection technology.
Siwei Lyu, a computer science professor at the State University of New York at Buffalo who specializes in AI image forensics, said he had not personally tested Meta’s tool, but noted that watermark-based systems come with inherent limitations.
“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 said.
Sarah Barrington, an AI researcher and Ph.D. candidate at the UC Berkeley School of Information, offered a measured take on the technology’s promise. She said watermarking could play an important role in the future of AI content verification, even if it isn’t perfect.
“Like many preventive cybersecurity or physical security measures, it may not be fully watertight, but even if we catch only 90% of cases, that’s still a great leap from 0,” Barrington said.
