OpenAI · 2026-05-19 · major
OpenAI Adds Google DeepMind's SynthID Watermark to AI-Generated Images — Public Verification Portal Now Checks Both C2PA Metadata and the Invisible Pixel Layer
OpenAI is layering Google DeepMind's SynthID watermark on top of its existing C2PA content credentials for images, and previewing a public portal anyone can use to check whether an image came from a ChatGPT, Codex, or API model.
Two rival labs share a watermark stack so anyone can check whether an image was generated by OpenAI's models.
Key specs
| Watermark layers | 2 |
|---|---|
| Models covered | ChatGPT, Codex, OpenAI API |
| Partner | Google DeepMind |
What is it?
OpenAI is shipping two layers of content provenance on every image its models produce. The first is C2PA metadata, an industry-standard tag attached to the file. The second is SynthID, an invisible pixel-level signal developed by Google DeepMind that survives screenshots, cropping, compression, and re-uploads.
How does it work?
C2PA writes signed metadata into the file header that says where the image came from. SynthID embeds a statistical pattern directly into pixels using a neural network at generation time; a paired detector then reports whether the watermark is present and how confident it is. OpenAI is also previewing a public verification portal that runs both checks against any uploaded image and returns the result.
Why does it matter?
Metadata gets stripped the moment an image is screenshotted or pasted into Slack; an invisible pixel watermark is harder to remove. Stacking the two raises the cost of laundering an AI image as real photography. OpenAI and Google calling out each other's tech in the same release is also the first significant cross-lab cooperation on provenance.
Who is it for?
trust-and-safety teams, journalists, election integrity researchers
Try it
https://openai.com/index/advancing-content-provenance/