Meta introduced a new verification tool in the preview stage that will help identify content prepared with the new artificial intelligence-supported visual production model Muse Image. This web-based tool can detect invisible watermarks in images created or edited by Muse Image. The company states that the same technology will be used for videos produced with artificial intelligence in the future, and that this step aims to more easily verify the source of digital content.
According to the information shared by Meta, the basis of the system is the invisible watermark technology called Content Seal. The company says that this watermark can be preserved even if the image is cropped, compressed, resized or screenshoted. The web-based verification tool also provides a first-level verification of whether the content was created by Meta AI by searching for this watermark on the uploaded image. Meta emphasizes that this should be considered as a tool that will allow users to better understand the source of the content, rather than a definitive verification system.
Content Seal system is currently limited to Muse Image
Unlike the watermark technologies that Meta previously shared as open source, Content Seal is this time offered as a special solution developed by the company itself. In addition, the visible logo in the lower right corner, seen in previous Meta AI versions, is no longer included in the content created with Muse Image. Instead, a digital marking method that cannot be noticed by users is preferred.
The company notes that its current verification system only supports images created or edited using Muse Image. Despite this, it was announced that Content Seal technology will be extended to videos created and edited with artificial intelligence in the future. The video production model called Muse Video, which Meta is working on, is also planned to be available soon. Thus, the aim is to use the same verification infrastructure in both visual and video content.
In the tests conducted by Engadget, it was reported that the verification tool correctly detected the edited content as well as the images created entirely with Muse Image. In addition, it was stated that the Content Seal watermark can be detected in the screenshots of the created images. According to Meta’s support page, a positive result means that the image was produced or edited through the Meta AI application or meta.ai. A negative result indicates that it is more likely that the relevant content has not been processed by Meta AI.
On the other hand, it seems that the same verification ability has not yet been added to the chat bot in the Meta AI application. In the trials reported in the news, the Meta AI assistant stated that it could not make a definitive assessment about an image marked as artificial intelligence production by the verification tool. The assistant stated that he could not understand which artificial intelligence model was produced by just looking at the image and that he did not have a detection tool used for this purpose.
Meta’s approach to labeling artificial intelligence content has been subject to criticism before. In its assessment published this year, the company’s independent Oversight Board expressed concerns that Meta could not consistently implement digital watermarking in content created with its own artificial intelligence tools. The new verification system can be considered as one of the technical responses to these criticisms. However, it seems that the current version does not yet include all Meta AI content.
In addition, the system also has some technical limitations. Content Seal is not compatible with other common digital watermarking solutions, such as Google’s SynthID system or the C2PA Content Credentials standard. In the tests, images created with Meta’s old artificial intelligence models could not be detected by the verification tool. When the images produced in old chats were uploaded to the system, it could not be determined whether they were created by Meta AI.
In addition, the verification tool is also limited by usage quota. After consecutive uploads in a short time, the system shows a warning that the daily verification limit has been reached. This may restrict the use of the current preview version, especially for users who want to verify a lot of content. However, as Meta expands its Content Seal technology to a wider range of products and adds video support, the company’s approach to detecting AI-derived content is expected to become more comprehensive.