Meta’s recently launched AI image detection tool, designed to identify images generated by its own Muse Image model, has demonstrated significant limitations. When AI-generated images are cropped, the tool often fails to recognise them as its own creations. This issue has been observed during tests conducted shortly after the tool’s introduction, raising concerns about the reliability of watermark-based detection methods for AI-generated content.
According to The Indian Express, a Reuters analysis found that Meta’s detection tool successfully identified all original AI-generated images but failed to verify 55% of the same images after they were cropped to about one-third or one-half of their original size. The tool relies on an invisible watermarking system called Content Seal, which is embedded in every image generated by Muse Image.
Meta has stated that the watermark is intended to remain intact after common edits, including cropping. However, coverage revealed that the signal may be lost if an image is heavily cropped, reducing the effectiveness of the detection tool. This limitation is not unique to Meta, as other technology companies such as Google and OpenAI have also acknowledged that their detection tools are not foolproof against image alteration techniques.
In March, Meta’s Oversight Board called for stronger detection tools to address the proliferation of deceptive AI-generated content on its platforms. Analysis showed that the board’s recommendations included investing in more robust systems to ensure the authenticity of images shared online, especially during periods of heightened misinformation risk such as election cycles.
“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,” said Siwei Lyu, a computer science professor at the State University of New York at Buffalo.
Sarah Barrington, an AI researcher at the UC Berkeley School of Information, noted that watermarking holds promise for the future of AI-generated content verification. However, reporting indicated that even with advanced watermarking, detection rates may not reach 100%, and some manipulated images could evade identification.
Meta has acknowledged the limitations of its preview detection tool and indicated that improvements are ongoing. The company emphasised that the tool is still in development and that further updates are expected to enhance its reliability. The issue remains critical as AI-generated images become more prevalent and the need for effective detection grows as details emerged.
“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.
Industry experts agree that while watermarking is a step forward, it is not a comprehensive solution. Ongoing research and development are required to address the evolving challenges posed by AI-generated content, particularly as image manipulation techniques become more sophisticated in the current landscape.
Note: This article is produced using AI-assisted tools and is based on publicly available information. It has been reviewed by The Quint's editorial team before publishing.