Meta is working on tools to identify images and videos created with Muse Image, a new image generation model. The company showed off a preview of a web-based tool that allows users to check for invisible watermarks used on new models.
This watermarking system, called Content Seal, remains intact “even when cropped, compressed, resized, or screenshotted,” Mehta explained in a blog post. “We’re previewing a detection tool that lets you see if an image contains a Content Seal watermark. This provides a first way to help you better understand whether an image was created with Meta AI.”
Content Seals seem to be a somewhat new approach for Meta. Meta told Engadget that while the version that is part of Muse Image is proprietary, the company has previously released open source versions of the technology. Meta’s new model does not include a visible watermark, similar to previous versions of Meta AI that added a small logo in the bottom right corner.
For now, Meta AI’s detection capabilities are limited to images created or edited with Muse Image, but the company said it plans to extend Content Seal watermarks to AI-generated and edited videos as well. Meta is also working on another video generation model called Muse Video, which is “coming soon.”
Today I tried out a new detection feature for images created using Meta AI. The web-based tool was able to detect watermarks in edited images and works created entirely by AI (like the one pictured above). Watermarks were also found in screenshots of my images. “A positive result means the image was generated or edited using the Meta AI app or meta.ai,” the company explains in its FAQ. “A negative result means the image is unlikely to have been processed using the Meta AI app or meta.ai.”
Interestingly, Meta AI’s new detection capabilities don’t seem to be part of the Meta AI app yet. When I asked Meta’s app-based assistant about images that the web tool had determined were created by AI, it said it didn’t have the ability to check. “It is impossible to determine whether this particular image was created with Meta Al just by looking at it.” “Meta Al does not automatically watermark images and does not have tools that can detect which Al model created an existing image.”
Meta has previously faced some criticism for how it labels and identifies AI-generated material within its app. Earlier this year, the Oversight Committee told the company it was “concerned” that Meta was “inconsistently implementing” digital watermarks on AI content created using its tools.
However, it appears that this new feature still has other limitations. Content Seal is not compatible with SynthID or C2PA Content Credentials, two established watermarking methods used by other companies. In my testing, the web-based feature was unable to identify images created or edited with previous versions of Meta’s AI model. When adding an image created in an old chat using Meta AI, it was not possible to tell whether the image was created with AI or not. This feature appears to be subject to Meta rate limiting for some reason. After uploading a few examples, I was warned that I had reached my “daily limit for ID checks.”
