🏠 News Empire
tech

Meta's AI Image Detector Fails to Spot Its Own Cropped AI Images, Analysis Finds

Published on: 11 Jul 2026, 04:02 AM
Meta's AI Image Detector Fails to Spot Its Own Cropped AI Images, Analysis Finds

A Reuters analysis has found that Meta's new artificial intelligence detection tool, previewed alongside the launch of its image-generation model Muse Image, failed to identify some of its own AI-generated images after they were cropped. The finding underscores the ongoing difficulties in verifying AI-generated content after common modifications, a challenge that could complicate efforts to detect deepfakes online, particularly during election periods.

In an analysis of 40 images generated using Muse Image, Reuters reported that the detection tool verified all original AI-generated images but failed to verify 55% of the same images after they were cropped to approximately one-third to one-half of their original size. Meta's detection tool relies on an invisible watermarking system called Content Seal, embedded in every Muse Image output, designed to remain intact after common edits. However, the company acknowledged that the watermark may be lost if an image is heavily cropped.

Meta noted that the tool is still in preview and that the watermark is intended to withstand typical alterations. Competitors Google and OpenAI have similarly cautioned that their detection tools are not foolproof against image-manipulation techniques. In March, Meta's Oversight Board urged the company to strengthen its detection tools to address the spread of deceptive AI-generated content on its platforms.

Experts in AI forensics have pointed out limitations of watermark-based systems. Siwei Lyu, a computer science professor at the State University of New York at Buffalo, stated that watermark methods can be effective when the watermark remains intact, but modifications like cropping, resizing, heavy compression, or editing can reduce their effectiveness depending on the watermark's design. Sarah Barrington, an AI researcher at UC Berkeley, added that while watermarking holds promise, it may not be completely watertight, but catching even 90% of cases would be a significant improvement from zero.

Latest in Tech 10
→ View All Tech News