OpenAI on May 19 said it will embed Google DeepMind's SynthID invisible watermark into every image generated by ChatGPT, Codex, and the OpenAI API, layered on top of C2PA Content Credentials metadata. The company is also previewing a public verification tool that detects either signal — initially limited to OpenAI-produced images, with cross-platform expansion promised in the coming months.
A two-layer provenance stack
The pairing is deliberate. C2PA Content Credentials write structured metadata into the file at generation time: model, edit history, and origin. That metadata is rich but fragile — a screenshot, a re-encode, or a social network's compression pipeline strips it. SynthID, developed by Google DeepMind, perturbs pixels in a way that survives those same transformations but carries far less context. OpenAI framed the combination as making provenance more durable than either layer alone — watermarks survive transformations, while metadata carries richer context.
For builders shipping image features against the OpenAI API, this changes the artifact returned by the generation endpoints. Outputs will now ship with embedded signals whether or not your application surfaces them, and downstream pipelines — moderation, ad review, newsroom workflows, e-discovery — gain a primitive they can interrogate without a model call.
The verification tool's deliberate limits
The public verification preview accepts an uploaded image and reports whether OpenAI-origin signals are present. It does not, at launch, identify content from other generators. That matters: a "clean" result tells you the image was not produced via OpenAI, not that it was produced by a human. Cross-industry verification depends on other providers adopting SynthID or C2PA themselves. OpenAI noted that Kakao and ElevenLabs are already shipping SynthID, and OpenAI has joined the C2PA Steering Committee — the cross-industry body that maintains the spec.
What changes for production systems
Three practical consequences for teams building on OpenAI image endpoints. First, compliance and trust-and-safety stacks now have a deterministic provenance check they can run server-side without a model in the loop. Second, any pipeline that strips EXIF or rewrites images for delivery needs to preserve C2PA manifests deliberately — or fall back to SynthID detection. Third, regulators in jurisdictions standing up disclosure regimes (the EU AI Act's transparency obligations, China's content-labeling rules, several US state laws) gain a more enforceable hook: "AI-generated" can be inspected at the file level rather than asserted by a publisher.
The missing piece remains video. OpenAI has reportedly committed to extending Content Credentials to Sora outputs when the model returns to public availability — Sora was shut down in April 2026 — but did not commit a date. Until then, the verification tool is a still-image story — useful for product feeds, ad creative, and journalism integrity, less so for the synthetic-video disclosure problem policymakers are actually most worried about.
