Meta AI image detector fails on cropped images, Reuters analysis finds

A Reuters analysis has revealed that Meta's new AI image detection tool, recently previewed alongside its Muse Image model, struggles to identify its own AI-generated images once they have been cropped. The analysis examined 40 images created by Muse Image, finding that while the detector successfully verified all original images, it failed to identify 55% of the same images after they were cropped to roughly one-third to one-half of their original size.

Meta's tool utilizes an invisible watermark called Content Seal, designed to remain intact through common edits and help users verify images created by its AI models. However, the company acknowledged that the tool is a preview and that the watermark signal may be lost if an image is heavily cropped. This finding underscores the ongoing challenges in reliably detecting AI-generated content, particularly deepfakes, especially in a year with significant elections.

Other major tech companies like Google and OpenAI have also issued caveats regarding the foolproof nature of their own detection tools against image alteration techniques. Experts in AI image forensics, such as Siwei Lyu from SUNY Buffalo and Sarah Barrington from UC Berkeley, noted that watermark-based systems have inherent limitations. Lyu explained that modifications like cropping, resizing, or heavy compression can weaken the embedded signal, impacting effectiveness depending on the watermark's design. Barrington added that while watermarking holds promise, even catching a high percentage of cases represents a significant improvement over current capabilities.