Developer finds hidden GUID watermark in Microsoft's AI image tools
A developer probing Windows 11's Paint app discovered an invisible watermarking function called WmkWriteWatermark in a DLL. The watermark is applied to all images generated via Paint or Photos AI features, even those processed locally on Copilot+ PCs. Microsoft also adds a visible Copilot logo watermark to such images.
The watermarking system relies on a server-issued GUID embedded directly into image pixels, alongside visible Copilot logos and C2PA Content Credentials attached via separate DLLs. Even when image generation runs locally on Copilot+ PCs, prompts are still transmitted to Microsoft servers for moderation. The process is mandatory in Paint—any failure aborts generation entirely—while Photos logs an error but still returns the image. Li's reverse engineering also revealed four local model files, three of which were encrypted, suggesting additional hidden processing layers.
This discovery follows Microsoft's broader efforts to label AI output, including recent tests of Copilot+ features on discrete GPUs and the controversial GDID telemetry system. The invisible watermark's mandatory nature indicates a deliberate compliance strategy, likely tied to the EU AI Act's transparency rules that took effect in August 2026. The use of both visible and invisible markers, plus server-side moderation, shows a multi-layered approach to AI content provenance.
This hidden watermarking could affect everyday users who create AI images in Paint or Photos, as their outputs carry traceable identifiers that may be used for accountability or auditing. It may also influence trust—people viewing AI-generated images might unknowingly see invisible markers, while creators could face unexpected limits if watermarking fails. For journalists, researchers, and platforms, this could aid in verifying content origins, but it also raises privacy questions about how GUIDs and server interactions are stored or shared. The mandatory nature suggests Microsoft prioritizes compliance over user convenience, potentially shaping industry norms for AI transparency.