Meta Develops Tool to Identify Images and Videos Created with Its New AI Models

Meta has unveiled a preview of a web-based tool designed to detect images and videos created with its latest image generation model, Muse Image. The tool relies on a watermarking system called Content Seal, which embeds invisible markers that persist even after cropping, compression, resizing, or screenshotting.

Content Seal is a proprietary technology for Meta, though the company has previously released open-source versions. Unlike earlier Meta AI models that added a visible logo to images, the new models do not include any visible watermarks. In tests, the web tool successfully identified watermarks in fully AI-generated images, edited versions, and even screenshots.

Currently, the tool only works for images created or edited with Muse Image, but Meta plans to extend Content Seal to AI-generated and edited videos. The company is also developing a separate video generation model called Muse Video, which is expected to launch soon.

However, the detection tool has several limitations. It is not compatible with other watermarking methods such as SynthID or C2PA Content Credentials. Additionally, it cannot identify images made with older versions of Meta's AI models, and it is subject to daily rate limits—after uploading a few samples, the test indicated that the daily limit had been reached.

Notably, the Meta AI app does not yet include this detection capability. When the author asked the app's assistant about an image that the web tool had identified as AI-generated, the assistant replied that it could not determine its origin. Meta has faced criticism in the past for inconsistent labeling of AI-generated content; the Oversight Board earlier this year expressed concern that Meta was inconsistently implementing digital watermarks on AI content created by its own tools.