Meta’s AI detection tool failed to verify 55% of its own AI-generated images after cropping, according to a Reuters analysis of pictures made with Muse Image.
Reuters said the tool verified all 40 original Muse Image pictures in its test. However, it failed on more than half of the same images after they were cropped to about one-third to one-half of their original size.
Meta says the preview tool uses an invisible watermarking system called Content Seal. The company says Muse Image embeds the watermark in every generated picture to help verify Meta-made AI content.
Meta said the tool remained in preview. It said the watermark should survive common edits, but the signal may be lost after heavy cropping.
The finding points to a weakness in AI image verification as altered synthetic images spread online during an election year that includes the U.S. midterms.
Google and OpenAI have also cautioned that their image detection tools are not foolproof against alteration techniques.
Siwei Lyu, a computer science professor at the State University of New York at Buffalo, said watermark systems can work when the signal remains intact. He said cropping, resizing, compression or editing may weaken detection.
Read: Meta Muse Image Rolls Out With Instagram Privacy Opt-Out
Sarah Barrington, an AI researcher and Ph.D. candidate at the UC Berkeley School of Information, said watermarking could still help even if it does not catch every case.
