Microsoft Implements Hidden Identifiers in AI-Generated Imagery
Microsoft has integrated invisible watermarks into images created using artificial intelligence (AI) through its Paint and Photos applications. These hidden identifiers contain information that could potentially link back to the user who initiated the image generation request.
Details of the Invisible Watermark Implementation
- Dual Generation Mechanism: Microsoft Paint supports both local and cloud-based image generation. Both Paint and Photos applications also come equipped with local AI models.
- Remote Prompt Moderation: All user prompts are submitted to a remote server for moderation before image generation.
- GUID Embedding: Following moderation, the server returns a unique Global Unique Identifier (GUID), which is then embedded into the locally generated image as an invisible watermark.
- Independent of Visible Settings: Crucially, a toggleable visible watermark feature in the settings does not control this invisible identifier.
- Copilot+ PC Specifics: On Copilot+ PCs, image generation is performed locally, but the prompt moderation process remains remote.
- Metadata Standards: Microsoft has disclosed that Paint adds C2PA (Coalition for Content Provenance and Authenticity) metadata to AI-generated images.
- Supported Save Formats: AI-generated images can only be saved in C2PA-supported formats, including PNG, JPEG, GIF, and the proprietary .paint format.
These measures allow for the traceability of content created using Microsoft’s AI tools, providing an additional layer of identification for each generated image.
Microsoft’s implementation of invisible watermarks, specifically the GUID embedding post-remote prompt moderation, is a significant step for AI content provenance. While C2PA metadata is standard, the direct linking to a user via a GUID raises interesting questions about data privacy and the potential for misuse, despite its clear intent for traceability and accountability in AI-generated media. The distinction between local generation and remote moderation is critical here.