A tool designed to detect AI-generated images has failed a basic real-world test and the images it missed were ones it created itself. A Reuters analysis published on July 10, 2026, found that Meta’s newly previewed AI detection tool could not identify more than half of its own AI-generated images once those images had been cropped. The finding arrives at an uncomfortable moment, just days after Meta launched Muse Image, its first publicly available image generation model, and at the start of a year that includes the US midterm elections.
In a test of 40 images generated using Muse Image, Reuters found the detection tool correctly flagged all 40 originals as AI-generated. But when those same images were cropped to approximately one-third to one-half of their original dimensions, the tool failed to identify 55% of them as AI-generated content returning no warning to users who might rely on the detector to make judgements about authenticity.
“A new AI detection tool from Meta failed to identify some of its own AI-generated images once they were cropped, a Reuters analysis found. The tool missed 55% of Muse Image-generated photos after they were cropped to approximately half their original size.”~Reuters
The Content Seal Promise And Where It Breaks Down:
Meta had made specific claims about how the system was designed to work. The company embedded an invisible watermarking technology called Content Seal into every image generated by Muse Image, describing it as a signal that would remain intact even after an image was cropped, compressed, resized, or screenshotted by users. The accompanying detection tool was presented as a way for anyone — journalists, fact-checkers, ordinary users to verify whether an image had been created by Meta’s AI.
When Reuters shared its findings with the company, Meta did not dispute the results. The company acknowledged that while the watermark is designed to withstand common edits, the signal can be lost if an image is heavily cropped. It emphasised that the detection tool is currently in preview meaning it is not yet a finished, fully released product.
That distinction between a preview tool and a released product matters, but it does not fully address the concern. A detection tool that is publicly available even in preview will be used by people making real decisions about real images. If the tool says an image is not AI-generated because its watermark was disrupted by cropping, a user who trusts the result has been given false assurance.
“Meta’s new AI detector can’t detect images it generated itself, a Reuters report finds. The tool correctly flagged all 40 original images from Muse Image as AI-generated, but missed 55% of the same images after they were cropped — undermining the company’s Content Seal watermarking claims.”~Gizmodo
A 900% Rise In Deepfakes And Detection Tools That Haven’t Kept Pace:
The practical stakes of this technical limitation are significant. According to cybersecurity firm DeepStrike, the volume of AI-generated deepfakes online has experienced roughly 900% annual growth between 2023 and 2025. Yet detection capabilities have not advanced at a comparable pace. Commercial AI detection tools are still plagued with mistakes, and the average person’s ability to independently identify AI-generated content is, according to previous studies, no better than a coin toss.
Siwei Lyu, a computer science professor at the State University of New York at Buffalo who researches AI image forensics, put the limitation clearly without having directly evaluated Meta’s specific tool: watermark-based methods are highly effective when the watermark remains intact, but any modification that removes or weakens the embedded signal – cropping, resizing, heavy compression, editing can reduce their effectiveness significantly depending on how the watermark is designed.
Sarah Barrington, an AI researcher and PhD candidate at UC Berkeley’s School of Information, acknowledged the gap while arguing watermarking still represents progress: “Like many preventive cybersecurity or physical security measures, it may not be fully watertight, but even if we catch only 90% of cases, that’s still a great leap from 0.”
“Watermark-based methods for AI image detection can be highly effective when the watermark remains intact, but any modification that removes or weakens the embedded signal — such as cropping, resizing, heavy compression, or editing — may reduce their effectiveness depending on how the watermark is designed.”~Siwei Lyu (
Instagram’s Public Photo Feature Pulled And Muse Video Still To Come:
The detection tool controversy was not the only problem Meta faced around the Muse Image launch. Instagram users reacted with alarm when it emerged that the Muse Image model could generate images using photos from any public profile without explicitly asking for the account owner’s consent. That feature was quickly pulled after public backlash, with Meta stating: “We’ve heard the feedback that this feature missed the mark, so it’s no longer available.”
Meanwhile, Meta’s Oversight Board urged the firm in March to do more to combat the spread of misleading AI-generated material on its platforms, as well as to invest in improved detection technologies. According to Reuters, the detection equipment deployed with Muse Image has failed to meet that grade. Meta’s next planned release in the generative AI field is Muse Video, a video creation model with more complicated detection and watermarking difficulties than those shown in the picture tool test.
