Microsoft applies watermarks to AI-generated images in its Windows Paint and Photos applications. This compliance follows European Union transparency rules. Anthropic previously noted that at least 190 providers agreed to these rules.
EU regulations require machine-readable formatting for AI content. The rules do not specify compliance methods. Signatories choose technologies ensuring effectiveness and reliability.
Microsoft embeds a globally unique identifier in these images. Xusheng Li of Vector 35 recently analyzed this approach. Li explained that Microsoft sends prompts for moderation and encodes the returned GUID into the pixels.
This 16-byte integer watermark differs from the visible watermark option. The C2PA manifest contains a GUID linked to user prompts. Li argues Microsoft did not sufficiently clarify this connection.
This tracking method resembles historical laser printer tracking. Privacy advocates previously raised concerns about printer tracking. Microsoft could theoretically identify users through the image watermark if prompts tie back to accounts.
Other firms pursue similar strategies. Meta is developing Content Seal. OpenAI utilizes Google DeepMind SynthID and C2PA metadata.
Users seeking to avoid tracking numbers can run open weight models. Stable Diffusion and on-source tools offer alternatives. Microsoft did not immediately respond to a request for comment.
