Understanding the Provenance of Content You See and Hear Online

OpenAI is introducing new tools to help identify the authenticity of online content and is joining the C2PA (Coalition for Content Provenance and Authenticity) steering committee. This is a measure to transparently disclose the origin of content as audiovisual content produced by generative AI surges.

In its text provenance technology research, OpenAI is focusing on reviewing text watermarking and metadata methods. Watermarking shows high accuracy, but is vulnerable to extensive manipulation such as translation or reconstruction using another model, and carries the risk of causing bias against non-English speaking users. On the other hand, the metadata method, which allows cryptographic signing, is evaluated as a promising alternative in that it has no false positives.

In the image field, C2PA standard metadata is being integrated into products. When an image generated with DALL-E 3 is edited in ChatGPT, the entire history—including the apps and tools used and the edits performed—is transparently recorded through C2PA certification, making it possible to trace the content's creation and modification process.