Meta's New AI Watermark Is Already Failing Basic Tests
Content Seal, Meta's invisible stamp on AI-generated images, can't be read by rival detection tools and failed to flag more than half the images in one real-world test. Experts and users deserve better.
Key points
- Meta launched Content Seal in July 2025, an invisible watermark applied only to images made by its new Muse image generator.
- Reuters tested Content Seal in 2025 and found it failed to detect more than half of Muse-generated images after they had been cropped.
- When a test image was fed into Google's Gemini AI assistant and the official C2PA detection portal, neither tool confirmed the image was AI-generated.
- Meta has offered AI image generation tools since 2023, meaning years of output cannot be identified by its own new system.
- Google's SynthID watermarking technology has already been adopted by OpenAI, yet Meta chose to build a separate, incompatible system instead.
Meta, the company behind Facebook and Instagram, quietly announced a new tool in July called Content Seal. The idea is straightforward: every image created by Meta's AI image generator, Muse, gets an invisible stamp baked into the picture. You cannot see it, but a detection tool can read it and tell you the image was made by AI.
Invisible watermarking is not new. Google already runs a system called SynthID that works the same way, and OpenAI uses it too. A separate industry standard called C2PA Content Credentials, which is a kind of digital certificate of origin that travels with an image file, also exists and Meta already sits on the steering committee that promotes it.
So why build yet another system from scratch?
Meta has not given a clear answer. What it has given is a tool with some notable gaps.
Right now, Content Seal only covers images from Muse. It does not cover video. It does not cover anything made with Meta's older AI tools, which have been generating images since 2023. The detection tool is a separate website users must visit on their own. It is not built into Meta AI, the company's chatbot, the way Google's detection is built into Gemini.
Meta also caps how many times a day someone can check images. That daily limit is understandable from a security angle, but it works against the goal of making detection easy at scale. C2PA has no such cap.
The bigger problem is compatibility. When The Verge ran a practical test and fed a Muse-generated image into Google's Gemini and the official C2PA checking portal, neither tool flagged it as AI-made. Content Seal, for now, essentially only talks to itself.
Meta spokesperson Faith Eischen told The Verge the company is "exploring ways to bring detection closer to where people encounter AI-generated content" and is "determined to work with industry peers." That language signals the work is not done.
Reuters found the system's real-world performance even weaker: Content Seal failed to identify more than half of Muse-generated images once they had been cropped. Cropping is one of the most basic edits anyone makes to a photo.
Should ordinary users trust Content Seal to protect them?
Not yet, and not on its own. The system is too narrow in what it covers, too isolated from the tools other companies use, and too easy to break with a simple crop.
If you see an image on Facebook or Instagram with an "AI" label, that label may come from metadata signals rather than Content Seal alone. On other platforms, Muse-generated images may carry no label at all.
What to watch for:
- An AI label on a social post does not mean every AI-made image has been caught. Absence of a label is not proof an image is real.
- Reverse image search in Google or a dedicated fact-checking site is still your most reliable first check.
- Cropped or resized images are harder for any watermark system to read, so treat heavily edited images with extra scepticism.
Researchers at Reuters and the team at The Verge AI are the ones who surfaced these gaps. Credit where it is due.
