A tool developed by Meta to detect images generated by artificial intelligence demonstrated failures when trying to identify its own creations after they were subjected to a simple crop, according to Reuters.
Results of the Test with Generated Images
The test involved 40 images produced by Muse Image, the company's latest image generation model. Initially, the tool correctly identified all 40 original images. However, when these same images were reduced to about one-third or half their original size through cropping, the detector failed in 55% of cases.
System Functioning and Limitations
Meta explains that its detection system uses an invisible watermark called Content Seal, which is embedded in every image generated by Muse Image. This technology was designed to allow users to verify whether an image was created by the company's AI models, even after undergoing usual edits.
When questioned about the findings of the Reuters analysis, Meta emphasized that the tool is still in a preview phase. The company admitted that the watermark was designed to withstand common edits but acknowledged that the signal can be lost in very drastic crops.
Opinions from AI Experts
Siwei Lyu, a computer science professor at the State University of New York at Buffalo and an expert in AI image forensics, told Reuters that systems relying on watermarks have known limitations. He stated that while watermark-based methods are highly effective when the signal remains intact, any alteration that weakens or removes this signal—such as heavy compression, resizing, or cropping—can decrease performance, depending on the watermark's design.
Sarah Barrington, an AI doctoral student at UC Berkeley's School of Information, agreed with the limitations but argued that, like other cybersecurity or physical safety measures, it cannot be entirely infallible. She added that even if detection only reaches 90% of cases, this represents a significant advance compared to a zero rate.
Regulatory Context and Competition
This discovery comes at a sensitive time, as the year includes US midterm elections, a period when the dissemination of manipulated images and deepfakes tends to increase. In March, Meta's Oversight Board, an independent body responsible for binding decisions on content moderation, requested that the company intensify efforts against the proliferation of misleading AI-generated content on its platforms and invest in more robust detection tools.
Meta does not face this challenge alone; both Google and OpenAI have also issued public warnings that their own detection tools are not immune to image manipulation techniques.
