Microsoft links AI watermarks in Paint and Photos to user IDs
Researcher finds that images created with AI in Microsoft apps contain unique identifiers tied to user prompts.
1Microsoft's AI watermark implementation
Microsoft is adding an invisible watermark to images created with AI in its Paint and Photos apps. This watermark contains a GUID (Global Unique Identifier) linked to the prompt used to generate the image. Researcher Xusheng Li shared details of this practice in a LinkedIn post. Microsoft aims to go beyond the EU's transparency rules for AI safety with this method.
2EU's AI transparency rules
The EU's rules on transparency of AI-generated content require signatory companies to label content created or manipulated by AI. However, the rules do not specify compliance methods. As a founder of C2PA (Coalition for Content Provenance and Authenticity), Microsoft has developed a practice that exceeds these rules. The company links user prompts to GUIDs in images.
3How GUIDs work
Microsoft Paint and Photos add a 16-byte GUID to AI-generated images. This GUID is generated as a response after the user's prompt is sent to Microsoft's servers. According to Li's analysis, Paint also sends the previous promptGenerationId in the next moderation request, enabling explicit linking of successive requests. This theoretically allows user identities to be traced.
4Privacy concerns
Microsoft's practice is compared to similar tracking methods used by laser printer manufacturers. Such practices have been criticized by privacy advocates in the past. Microsoft aims to verify the source of images by linking user prompts to GUIDs. However, this method could compromise user anonymity. The company has not yet commented on this issue.
5Approaches by other companies
Besides Microsoft, companies like Meta and OpenAI are developing similar technologies to track AI-generated content. Meta is working on a watermarking technology called Content Seal. OpenAI uses Google DeepMind's SynthID and C2PA metadata. These companies are experimenting with different methods to ensure the authenticity and reliability of AI content.
6Open-source alternatives
Users who don't want tracking numbers in their AI-generated images can opt for open-weight models like Stable Diffusion and on-device open-source tools. These methods don't require sending user data to company servers. However, such tools may not offer as advanced features as those provided by Microsoft or OpenAI.
7Microsoft's AI safety measures
Microsoft aims to verify the origin of content using the C2PA manifesto for AI safety. By linking user prompts to GUIDs in Paint and Photos, the company enhances the traceability of AI-generated content. However, the implications of this practice for user privacy are not yet fully clear. Microsoft is expected to provide more transparency on this issue.
8The future of AI transparency
Microsoft's GUID embedding practice is seen as a significant step in AI transparency. However, such practices need to balance user privacy. The EU's AI transparency rules offer flexibility to companies while aiming to protect user rights. In the future, there may be more debates on the reliability and privacy of AI-generated content.