Meta ‘admits’ its AI detection tool can fail to identify cropped AI-generated photos created using Muse Image; says: Signal may get lost if an image is …

​Facebook-parent Meta has acknowledged that its new AI image detection tool may fail to identify some AI-generated images. According to a report by the news agency Reuters, the social media giant has revealed that its AI image detection tool may fail to identify some images generated by its Muse Image model when heavily cropped. The company said the tool’s invisible watermarking system is designed to survive common edits, but the embedded signal can be lost if an image is cropped significantly.​The disclosure comes after a Reuters analysis found that Meta’s preview AI detection tool successfully identified all original AI-generated images created with Muse Image but failed to verify 55% of those same images after they were cropped to around one-third to one-half of their original size. Responding to Reuters’ findings, Meta said the tool is still in preview. The company added that while the watermark is designed to remain intact after common edits, “the signal may be lost if an image is heavily cropped.”​Meta introduced the preview detection tool alongside Muse Image earlier this week. According to the company, every image generated by Muse Image includes an invisible watermark called Content Seal, which is intended to help users determine whether the image was created using Meta’s AI models.​

Researchers highlight AI tool’s limitations in watermark-based detection

​The findings come as technology companies continue to develop tools to identify AI-generated content, particularly as concerns over manipulated images and deepfakes grow during election periods. Google and OpenAI have also said their own AI image detection systems are not immune to image-alteration techniques.​Earlier this year, Meta’s Oversight Board urged the company to strengthen its approach to AI-generated content. In March, the board called on Meta to do more to address the “proliferation of deceptive AI-generated content” across its platforms and invest in stronger detection tools.​Researchers say watermark-based detection systems have technical limitations when images are modified.​“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,” Siwei Lyu, a computer science professor at the State University of New York at Buffalo, who researches AI image forensics told Reuters.​Sarah Barrington, an AI researcher and PhD candidate at the UC Berkeley School of Information, said watermarking remains a useful approach despite its limitations.​“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,” she said.