OpenAI strengthens the provenance of its images with C2PA and SynthID

OpenAI adopts the C2PA standard and Google DeepMind's SynthID watermark to trace images from DALL·E 3, Sora, and ChatGPT using a new public verification tool.

OpenAI, in turn, is adopting a multi-layered approach to trace the origin of its content, echoing Google's announcement on the same topic, which specifically mentioned this alignment. The first component is C2PA compliance: the company, which has been adding Content Credentials metadata to its images since 2024, from DALL·E 3 to ImageGen and Sora, is becoming a generator compliant with this open standard, which allows platforms to read and transmit this provenance information beyond the point of creation.

The second component is Google DeepMind's SynthID watermark, integrated through a partnership with Google, initially on images produced by ChatGPT, Codex, or the API. The two mechanisms complement each other: C2PA metadata carries detailed context but can be removed during a format change or a screenshot, whereas the invisible watermark better resists these transformations.

OpenAI is also providing a first version of a public verification tool, which indicates whether an uploaded image was generated via its products by detecting Content Credentials and SynthID signals. The approach remains cautious: in the absence of a detected signal, the tool does not draw a conclusion, as these markers may have been removed. Initially limited to OpenAI content, it is intended to expand to other platforms and media types, in line with the interoperability logic also promoted by the C2PA committee.