Developer digs into Microsoft Paint and Photos, finds invisible watermark on AI-made images
Windows users may not realize how much happens behind the scenes when they use Microsoft's AI image tools in Paint and Photos, according to Tom's Hardware .
Reverse-engineering by a developer suggests the Windows apps embed a hidden watermark in addition to adding a visible Copilot logo to AI-made images.
Here's what to know
Developer Xusheng Li says his examination of Windows 11's Paint app turned up signs of an invisible watermarking layer tied to AI-generated images. The visible Copilot logo already showed that Microsoft marks AI-created pictures, as Tom's Hardware reported. But Li also found a second form of labeling that the app writes into the image data itself.
Li started by looking at Paint's AI capabilities. In the process, he identified several local model files, including what Tom's Hardware described as an ONNX-style model file and three encrypted variants.
"Paint does more than alter the pixels," Li noted, as relayed by Tom's Hardware. "It also attaches C2PA Content Credentials to the saved file. The code responsible for this lives in ProvenanceHelper.dll, backed by provenancesdk.dll."
He said a function called AddPerceptibleWatermark appears to control the visible mark while another, WmkWriteWatermark, appears to write the invisible watermark. In Paint, that hidden watermark seems to be required. If the app cannot add it, image generation reportedly fails.
Li also said Microsoft Photos uses the same general GUID watermarking method with its AI tools.
More background
Beyond Microsoft's apps, the discovery adds to the overarching argument over how people and apps should identify AI-generated material.
Li added that the system "might be related to Article 50 of the EU AI Act, whose transparency rules took effect on Aug. 2, 2026, and require AI-generated content to carry a detectable, machine-readable mark — but not a prompt-specific GUID."
This might also mean that prompts may still be sent to Microsoft servers for moderation even when image generation appears to happen locally on a device.
AI also comes with infrastructure tradeoffs. It can improve forecasting, streamline industrial systems, and optimize clean energy networks, but those benefits carry costs.
High-powered, generative AI systems can use large amounts of electricity and water and put pressure on the grid. And even if it does become more eco-friendly, many people argue that it devalues the time, labor, and thought that artists put into their work.
What's being done?
Watermarking is helping people at least understand what they're looking at and where it came from. Standards efforts like C2PA are part of that push as well.
As AI features spread across creative software, hidden labeling may become more common, and server-side checks may remain part of tools that otherwise seem local. As Tom's Hardware stated, Li's findings indicate Microsoft is already building those safeguards into its apps.
Get TCD's free newsletters for easy tips, smart advice, and a chance to earn $5,000 toward home upgrades. To see more stories like this one, change your Google preferences here .
