Microsoft’s AI features in Paint and Photos appear to do more than add the visible Copilot logo to generated images. Developer Xusheng Li reverse engineered Paint’s AI image-generation path and found what he says is an invisible watermarking mechanism that embeds GUID-based data into generated output, according to his technical write-up and a Tom’s Hardware report on the findings. Li began by examining Paint’s AI image-generation implementation in Windows 11. He expected the feature to rely mainly on a remote API, but reported finding four apparent local model files with an .onnxe extension. In his analysis, one previously known file could be converted into a normal Open Neural Network Exchange model by XORing it with the string Microsoft_2023; Li says the other three used the same algorithm with different keys and could also be checked as valid ONNX models after decryption. The watermarking trail led to Watermarker.dll. Li found the expected visible watermarking path, which places a small Copilot logo at the bottom right of AI-generated images. But he also identified a separate function, WmkWriteWatermark, that Tom’s Hardware describes as previously unknown invisible GUID watermarking functionality inside Microsoft’s Paint and Photos apps. The key reported detail is how tightly Paint couples generation to watermarking. Li says WmkWriteWatermark is called after local Stable Diffusion image generation, and that if the function fails, Paint returns an error rather than handing back the image without the mark. Tom’s Hardware reports the same behavior, saying Paint effectively makes the invisible watermark mandatory for generated output. Li’s technical account says the incoming payload to the watermarking function is 16 bytes and is used to construct an 18-byte, 144-bit message. He identifies that 16-byte payload as a GUID. The embedding logic, as he describes it, rounds usable image dimensions down to multiples of eight, tracks 144 counters, and requires each bit to be placed at least three times. The routine then makes small quantized changes across selected image blocks using matrix operations, which Li characterizes as a content-adaptive, block-domain watermarking approach. Tom’s Hardware adds two important caveats. First, it says the GUID should not be confused with Microsoft’s separate GDID traceable device-specific telemetry controversy. Second, it reports that the watermarking is meant to identify or verify images where AI played a role, not necessarily to expose the visible content to a human viewer. Paint also appears to attach provenance metadata, according to Tom’s Hardware’s account of Li’s findings. The outlet reports that Paint adds C2PA Content Credentials to saved files through ProvenanceHelper.dll backed by provenancesdk.dll. That would put a standards-based file-level provenance layer alongside the pixel-level invisible mark described in the reverse engineering. The Photos behavior is reportedly similar but less strict. Tom’s Hardware says Photos has the same GUID mechanism when AI is used, but if the watermarking process fails, Photos returns the image and logs an error instead of aborting the operation. The provided material does not include Microsoft’s response or an independent confirmation of that Photos-specific implementation. Tom’s Hardware notes that the hidden processing may be related to Article 50 of the European Union AI Act, whose transparency rules took effect on August 2, 2026, and require AI-generated content to carry a detectable, machine-readable mark. The outlet also notes an important limit to that theory: the rule does not require a prompt-specific GUID. With no Microsoft statement in the provided material, the compliance rationale remains an informed possibility rather than a confirmed motive. Who benefits: Microsoft benefits if these mechanisms help it demonstrate that AI-generated content from its consumer apps is detectable and attributable as AI-assisted. Platforms and downstream tools that read C2PA metadata could also benefit if the provenance data remains intact. Who's exposed: Users and organizations generating images in Paint or Photos are exposed to hidden processing they may not notice from the image alone. The exposure is still unclear because the provided sources do not establish what Microsoft can infer from the GUID or how it is stored server-side.