Meta's AI Detector Struggles to Identify Cropped AI Images, Reuters Analysis Finds

Meta’s newly introduced AI image detection tool, designed to identify deepfake and AI-generated images, has shown a significant limitation. According to a new analysis by Reuters, images created using Meta’s new Muse Image generator often evade detection after being slightly cropped.

The finding raises fresh concerns about the spread of misleading AI-generated content during election seasons worldwide, including the U.S. midterm elections, where reliable verification of manipulated images has become increasingly important.

Reuters' Findings

Earlier this week, Meta unveiled its latest AI image generation model, Muse Image, alongside a preview version of its AI detection tool.

Reuters tested the detector using 40 AI-generated images created with Muse Image. The results showed that the tool successfully identified 100% of the original images as AI-generated.

However, once the same images were cropped to between one-third and one-half of their original size, the detector failed to recognize them in 55% of cases. In other words, more than half of the cropped AI-generated images bypassed Meta’s own detection system.

Meta Responds

Meta previously stated that the preview detection tool relies on an invisible watermarking technology called Content Seal, which embeds hidden identification data into every image generated by Muse Image. According to the company, this watermark is designed to remain intact even after common image edits, including cropping.

Responding to Reuters’ findings, Meta acknowledged that the tool is still in a preview stage. The company said the watermark is intended to survive typical editing, but excessive cropping may remove or weaken the embedded signal, making detection more difficult.

Experts Highlight Broader Challenges

Meta is not alone in facing this issue. Both Google and OpenAI have previously warned that no AI watermarking or detection technology is completely resistant to image editing or manipulation.

In March this year, Meta’s Oversight Board urged the company to strengthen its safeguards against misleading AI-generated content and invest in more robust detection technologies.

Siwei Lyu, Professor of Computer Science at the University at Buffalo, State University of New York, explained that watermark-based detection works only as long as the embedded watermark remains intact. Cropping, resizing, compression, and other image modifications can weaken or erase the hidden signal, significantly reducing detection accuracy.

Meanwhile, Sarah Barrington, an AI researcher at the UC Berkeley School of Information, noted that although watermarking is a promising approach for identifying AI-generated content online, it will inevitably have limitations. Like many cybersecurity technologies, it is unlikely to be completely foolproof. However, she added that successfully identifying even 90% of AI-generated images would represent a substantial improvement over having virtually no effective detection system at all.

//DBTech/BMT/OR//