OpenAI Adds C2PA Metadata and SynthID Watermarks to All Generated Images

OpenAI has committed to embedding two distinct provenance signals into every image its models generate. Going forward, outputs from ChatGPT, Codex, and the API will include C2PA content credentials - structured metadata that records how and when content was created or modified - alongside SynthID, the imperceptible watermark system developed by Google DeepMind.

The reason for using both is practical: C2PA metadata travels with a file but can be stripped when an image is screenshotted, re-uploaded, or converted. SynthID, by contrast, is embedded in the pixel data itself and survives most common transformations. OpenAI described the two as complementary - "C2PA helps content carry detailed context; SynthID helps preserve a signal when metadata does not survive."

This move also marks a meaningful expansion of SynthID beyond Google's own ecosystem. Nvidia has also announced plans to adopt the watermarking standard, suggesting SynthID is gaining traction as an industry-wide baseline rather than a proprietary Google tool. The C2PA standard, backed by a coalition that includes Adobe and Microsoft, has been the more widely adopted of the two until now.

For the broader effort to identify AI-generated imagery to work in practice, adoption across major generators is considered necessary. OpenAI producing a significant share of AI images in circulation makes this a consequential step, though the watermarks are not foolproof - adversarial removal techniques exist, and detection still requires access to verification tools most users do not currently have.