Meta has built a tool to identify images and videos created with its new AI image generator, called Muse Image. The company is letting people test it for free at meta.ai/identification.

Here's how it works: Meta embeds an invisible digital stamp — essentially a hidden marker — into every image its AI creates. The system is called Content Seal. When you upload an image to Meta's checker, it looks for this invisible stamp. If it finds one, the tool tells you the image was made by Meta's AI.

The key claim Meta is making is that this invisible stamp survives what normally destroys watermarks. When images travel across the internet, they get compressed, cropped, resized, and shared so many times that visible watermarks usually fall apart. Meta says its hidden watermark stays intact through all of that, which is why it switched away from the old approach — a visible logo in the corner of generated images.

That's a big engineering bet. If the invisible watermark works as Meta describes, it could help track AI-generated images even after they've been edited or shared multiple times.

Right now, the detection tool only works for images made with Muse Image. Meta says it will eventually add the same watermark system to AI-generated video, and the company is building a new video generator called Muse Video, but that hasn't shipped yet. The web checker also has limits: it caps how many images you can check per day, and testing by Engadget found it doesn't recognize images from Meta's older AI tools. The detector also doesn't work with images created by other companies' AI systems.

There's a larger problem here that's worth thinking about. Google, OpenAI, Adobe, and Microsoft have been working on a shared standard for labeling AI media, called C2PA Content Credentials. The idea is that any company could use the same format to mark their AI images, and any checker could read that mark, the same way all browsers can read the same websites. Google has also built its own watermarking system, called SynthID, and has started adding it across its own products.

Meta's invisible watermark doesn't work with either of these systems. An image marked with Meta's seal won't be recognized by a SynthID checker, and vice versa. That means researchers and platforms trying to catch AI-generated images at scale now have to use multiple, separate detection tools instead of one shared standard.

The real question is whether a watermark that only Meta's own tool can detect actually solves the problem it's meant to. Misinformation spreads across platforms Meta doesn't control, and it's checked by people using tools Meta doesn't build. A watermark that Meta can detect through its own website helps, but only if people remember to go there and check. A shared standard that works everywhere would be more powerful, because it would let any platform, news site, or social network check images on their own without sending people off to Meta's website.

Meta has said it plans to add video watermarking once its video generator launches, and the detection tool will probably move from a standalone website into Meta's main AI app. Whether Meta eventually connects its watermarking system to the broader industry standard, or keeps it separate, will say a lot about how AI provenance tools develop in the coming years.