MS Paint and Photos inivisibly watermark even locally generated output with GUID
Points and comments are a snapshot, not live.
Microsoft Paint and Photos embed a server-issued GUID as an invisible watermark in locally generated AI images.
Reverse engineering reveals Paint and Photos send prompts to a remote moderation server that returns a GUID. This GUID is embedded into locally generated images via Watermarker.dll as an invisible pixel-level watermark, separate from the optional visible Copilot logo. On Copilot+ PCs, image generation is local but prompt moderation remains remote. The watermark is also recorded in C2PA metadata as a soft-binding. The watermark cannot be disabled and uses a 16-byte GUID that can link images to prompts and Microsoft accounts.
Paint also restricts AI-generated image saves to PNG, JPEG, GIF, and .paint formats, excluding BMP which would strip C2PA metadata.
What commenters are saying
Commenters express alarm that the invisible watermark functions like digital yellow printer dots but can identify the exact user, device, and prompt. Several practical workarounds are suggested: converting to BMP removes C2PA metadata, though pixel-level watermarking remains. Some argue watermarking will only affect legitimate users while adversaries will strip it easily. One commenter notes that if the public comes to trust such markers, sophisticated actors could exploit that trust by creating unmarked deepfakes or by falsely marking real content as AI-generated.