Meta's new tool fails to identify cropped AI images
What's the story
Meta's latest artificial intelligence (AI) detection tool, unveiled this week with its image-generation model Muse Image, has faltered in identifying some of its own AI-generated images after they were cropped. The revelation was made by Reuters after analyzing 40 images created with Muse Image. The study found that while the detection tool verified all original AI-generated images, it failed to do so for 55% of them once they were cropped to one-third or half their original size.
Tool details
Detection tool recognizes Muse Image creations, even when cropped
Meta claims its preview detection tool can recognize its own AI-generated images, even when cropped. This is done through an invisible watermarking system called Content Seal. The system is embedded in every image generated by Muse Image and is intended to help users confirm if it was created by Meta's AI models. However, the company acknowledged that while the watermark is designed to survive common edits, it could be lost if an image is heavily cropped.
Verification hurdles
Complications in detecting deepfakes online
The difficulty of verifying AI-generated images after common alterations could complicate the detection of deepfakes online, especially during a busy election year like the US midterms. Rival tech companies Google and OpenAI have also warned that their own detection tools aren't foolproof against image-alteration techniques. Siwei Lyu, a computer science professor at SUNY Buffalo who studies AI image forensics, said watermark-based systems can be effective but modifications such as cropping or resizing can weaken their reliability.
Watermarking potential
Watermarking holds promise for AI-generated content verification
Sarah Barrington, an AI researcher and Ph.D. candidate at UC Berkeley School of Information, said watermarking holds promise for the future of AI-generated content but could only do so much. She compared it to preventive cybersecurity measures that may not be fully watertight but still make a significant difference in detection rates. This highlights the ongoing challenges and potential solutions in the field of AI image verification.