Microsoft Paint and Photos embed invisible watermarks containing server-issued GUIDs into AI-generated images created locally, according to reverse engineering by Xusheng Li. The watermark is added even when images are generated on the user’s device, with the GUID supplied by a remote moderation server. This process was detailed in a blog post published today on xusheng.dev.
The two apps support both local and cloud-based image generation, but always send the user’s prompt to a remote moderation server. The server returns a moderated prompt along with a unique GUID, which Paint and Photos embed invisibly into the pixels of the locally generated image. This watermark is separate from any visible watermark settings and is embedded regardless of user preference. On devices with Copilot+ integration, image generation is fully local, but prompt moderation remains remote, ensuring the GUID is still embedded. Microsoft also adds C2PA metadata to AI-generated images, which are saved only in formats that preserve this metadata, such as PNG, JPEG, GIF, and .paint files, according to the blog.
This discovery highlights Microsoft’s approach to AI content moderation and provenance tracking within its native image apps. Embedding a server-issued GUID invisibly into images allows Microsoft to trace the origin and moderation status of AI-generated content, even when generation occurs offline. This method complements the use of C2PA metadata, a standard for content authenticity. The approach contrasts with other AI tools that rely solely on visible watermarks or metadata, marking a nuanced strategy for content verification in AI-generated images.
The research was published on August 24, 2026, by Xusheng Li on xusheng.dev, providing detailed technical insights into the watermarking mechanism used by Microsoft Paint and Photos. The findings contribute to understanding how major software integrates AI moderation and provenance features in local image generation workflows.