Meta's AI detector missed 55% of its own cropped AI images, Reuters finds
An analysis of 40 images revealed a weakness.
Image by Getty
- Meta’s AI detector missed 55% of its own Muse Image pictures after Reuters cropped them.
- The tool verified all 40 original AI-generated images before cropping changed one-third to one-half of each picture.
- Meta said the detector is in preview and heavy cropping can weaken its invisible Content Seal watermark.
- The findings show watermark-based AI detection can fail when images are edited, shared, or compressed.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
An AI detector released by Meta alongside its new image-generation model, Muse Image, was supposed to help users identify whether an image was generated using its tool. But a Reuters analysis revealed that it might not be as reliable as advertised – failing to identify its own images after cropping in more than half of the tests.
Reuters analyzed 40 images generated using Muse Image and uploaded into Meta's detection tool. It correctly verified all of the AI-created images.
But after the same images were cropped to approximately one-third to one-half of their original size, the tool failed to identify 55% of them. None of the other parameters were changed.
Meta says its AI detection tool can verify whether an image is Muse-generated thanks to Content Seal, its invisible watermarking system. The detector checks whether an uploaded image contains a Content Seal watermark.
The watermark is designed to remain intact, Meta says, even when images are cropped, compressed, resized, or screenshotted.
In response to the analysis, Meta said that the tool was in preview, and although the watermark is designed to remain intact after common edits, heavy cropping can cause the signal to be lost.
Watermark-based systems are widely considered not a foolproof way of detecting whether an image is AI-generated. They may not survive if an image is shared or edited.
"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, told Reuters.
Several of Meta’s AI rivals have similar provenance or verification tools. For example, OpenAI's verification tool allows you to check whether an uploaded image contains provenance signals associated with images generated by OpenAI tools.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
However, the company openly states that a negative result doesn’t mean that the image wasn’t AI-generated. It could be that the metadata was stripped or has evidence of tampering, the watermark was degraded, the image comes from a legacy image generation model, or it was created before provenance signals were available.
Similarly, Google SynthID Detector allows you to upload an image, audio track, video or piece of text and check it for a SynthID watermark.
“While this technology isn’t perfect, our internal testing shows that it’s accurate against many common image manipulations,” Google's blog post said.
