Meta recently unveiled its image-generating AI, Muse Image, a release that quickly drew controversy when it emerged the system could produce images of strangers without consent — a capability Meta later disabled.
Alongside the generator, Meta introduced a tool called Content Seal, described by the company as an invisible watermark applied to every Muse Image-generated picture. Meta claimed the marker would be effectively persistent, remaining present even if users crop, compress, resize or take screenshots of the AI-created image.
Meta also said it was building a separate detector to read the Content Seal and verify whether an image was produced by Muse Image. However, a Reuters investigation found the detector’s performance is far from reliable.
In testing, Reuters reported that Meta’s system correctly identified all 40 original Muse Image-generated pictures as AI-made (100 percent detection). After those same images were cropped to half or a third of their original dimensions, the detector recognized the Content Seal in only 55 percent of cases.
This drop in detection accuracy is significant at a time when AI-generated deepfake images are becoming increasingly realistic. According to data from the cybersecurity firm DeepStrike, the volume of such images rose roughly 900 percent annually between 2023 and 2025, while detection capabilities have not kept pace, as noted by Gizmodo.
Publishing Muse Image is an important step for Meta as it seeks to close the gap with other technology companies. Yet without robust detection tools, the risks associated with undetected AI-generated imagery are likely to escalate. The issue is particularly pertinent as Meta prepares to release a video-generating AI, Muse Video — an initiative whose prospects are clouded by the reception to Muse Image and the Content Seal’s current detectability problems.
