Meta is developing a web-based detection tool designed to identify images and videos created with its image generation model, known as Muse Image. This tool verifies the presence of invisible watermarks referred to as Content Seal that are embedded within generated content.
According to Meta, Content Seal remains intact regardless of image manipulation, including cropping, compression, resizing, or screenshots. The company stated that the tool provides an initial mechanism for users to determine if an image carries a Content Seal watermark.
The watermarking approach utilized by Muse Image is proprietary and differs from earlier open-source versions released by Meta. The current version of Meta’s AI models does not employ visible watermarks, a departure from previous models that included logos in the images.
Currently, the detection capabilities are limited to images created or edited with Muse Image. Meta plans to extend Content Seal watermarks to AI-generated and edited videos in the future. The company is also developing a separate video generation model called Muse Video, which is expected to launch soon.
Testing of the detection tool confirmed its ability to identify watermarks in both edited images and those generated entirely by AI. A positive detection indicates the image was produced using the Meta AI app or meta.ai, while a negative result suggests it was unlikely processed with Meta AI. However, the detection feature is not integrated into the Meta AI app itself; the app’s assistant confirmed it cannot determine if a specific image was created by Meta AI.
Meta has faced scrutiny regarding the inconsistent application of digital watermarks for AI-generated content. The Oversight Board expressed concerns earlier in the year about the company’s practices in labeling AI-produced material. Content Seal is incompatible with established watermarking methods like SynthID and C2PA Content Credentials.
During recent tests, the detection tool failed to identify images created or edited with previous versions of Meta’s AI models. Additionally, users of the tool encounter rate limits, receiving alerts once they reach their daily threshold for identification checks.
