OpenAI adds provenance signals to its AI images

OpenAI says it will start attaching provenance signals to the images its models generate, using two methods at once. The first is C2PA, the open Content Credentials standard that writes an "AI-generated" marker into a file's metadata. The second is SynthID, Google's invisible watermark, which OpenAI says survives screenshots, resizing, and ordinary editing. The two are paired because they fail in different ways: metadata carries more detail but is easy to strip, while the watermark is harder to remove but says less.

Alongside this, OpenAI is previewing a public verification tool that checks an image for either signal, so anyone can test whether it came from an OpenAI model. The honest limit, which OpenAI states plainly, is reach. At launch the tool only recognizes images made by OpenAI's own products, and it does nothing about the much larger flood of images from other generators. The scope is narrow on purpose: the goal is to make sure OpenAI's own output can be traced, not to detect every AI image.

You can read OpenAI's post here.

Why it matters

If you moderate a platform or verify images for a living, you now have a way to confirm an OpenAI origin, but the coverage gap means a clean check proves nothing about images from other tools, so treat it as a signal, not proof.