Meta has previewed a web-based detection tool that can identify images, and eventually video, made or edited with its newer AI models by checking for an invisible watermark called Content Seal. The move matters because it is an attempt to improve transparency around AI-generated media after criticism that Meta has not labelled such content consistently.

The tool currently works with content produced by Muse Image and can still detect the watermark after cropping, compression, resizing or screenshots, according to Meta and Engadget’s testing. However, it does not yet work inside the Meta AI app, cannot identify content from older Meta AI models, is not compatible with other watermarking standards such as SynthID or C2PA, and appears to have daily rate limits on checks.