Meta Built an AI Detection Tool That Only Sees Its Own Watermarks
A watermark designed to survive cropping and compression
Meta is working on a detection tool that could identify images and videos created or edited with Muse AI models by embedding a special watermark in the content. The watermark is designed to resist common forms of image manipulation and can be removed with relative ease, according to the company.
A Watermark That Can Survive Cropping
Content Seal, the name of Meta’s watermarking technology, is supposed to be more resilient than most digital watermarks that typically vanish when an image is altered. By simply uploading the image to a specific website, it becomes possible to determine if the content has been manipulated or generated by AI because the watermark will still be visible even if the picture is cropped or compressed. Such an approach addresses the issue of proving the source of AI-made images because they frequently circulate through various messaging platforms and social networks, where they can be additionally modified.
Focusing on Meta’s AI Tools
The latest test results show that the proposed solution can recognize watermarks embedded in AI-generated and edited pictures. However, the current implementation has some significant limitations. For instance, the tool only works with the Meta AI application, does not support older AI models made by the company, and cannot read other watermark standards, such as Google’s SynthID. This problem is part of a larger context, and the fact that different firms are developing their approaches to detecting synthetic media presents a problem as well. In other words, users will be required to rely on distinct tools to identify AI-generated content from various sources.
A Similar Approach for Video Content
It is expected that Meta will adopt a similar strategy for its upcoming Muse Video model, indicating that watermarking AI-generated videos will soon be available. In the future, such measures might become essential in ensuring that AI content can be reliably identified by ordinary users. However, it is not clear how this task will be accomplished when multiple companies provide analogous but incompatible solutions.
Get the latest tech news on WhatsApp
Comments are closed.