On May 19, 2026, OpenAI announced that it is strengthening its efforts to make the provenance of AI-generated images and videos easier to verify[1]. The plan rests on three pillars: conformance with the industry standard C2PA, adoption of Google's SynthID watermarking technology, and a preview of a verification tool that anyone can use, all aimed at making it possible to trace the origin of generated content through multiple layers[1]. The backdrop is the growing need for a way to confirm, after the fact, where a piece of media came from and how it was created, as AI-based creation and editing of images and audio become commonplace[1].
C2PA conformance makes provenance information readable
The first pillar is conformance with C2PA, the industry standard for recording content provenance[1]. C2PA is a mechanism that attaches metadata (Content Credentials) to images and videos, indicating when and with what they were created and how they were edited. OpenAI joined the C2PA steering committee in 2024 and has been adding metadata to outputs such as DALL·E 3 images[2].
This time, OpenAI says it is shaping the provenance signals it applies into a form that other tools and platforms can recognize more easily[1]. Videos generated with Sora carry both a visible watermark and C2PA metadata[1], and by making such signatures easier to handle across the industry, provenance information becomes more likely to be carried along as content circulates.
SynthID, a watermark that is hard to erase
The second pillar is the adoption of SynthID watermarking technology through a partnership with Google[1]. SynthID, developed by Google DeepMind, embeds a watermark invisible to the human eye into images, and OpenAI says it will roll it out starting with images generated via ChatGPT, Codex, and the OpenAI API[1].
While C2PA metadata can convey detailed context, it has the weakness of being lost through operations such as taking a screenshot or converting a file[1]. A watermark like SynthID tends to survive such modifications, so OpenAI positions the combination of the two as a multi-layered mechanism that can preserve provenance signals even when the metadata disappears[1].
A verification tool to check an image's origin enters preview
The third pillar is a preview of a public verification tool that ordinary users can use[1]. It checks whether an uploaded image contains provenance signals such as Content Credentials or SynthID, allowing users to confirm whether the image was generated with ChatGPT, the OpenAI API, or Codex[1].
At the time of release, the tool can only assess content generated by OpenAI[1]. However, OpenAI says that within the coming months it hopes to support cross-industry efforts that make verification possible across platforms, setting a goal of improving transparency across the whole ecosystem rather than something that is complete within OpenAI alone[1].
Summary
OpenAI's announcement is an effort to make the origin of AI-generated content easier to trace through three layers: C2PA conformance, SynthID watermarking, and a public verification tool. By stacking multiple measures rather than relying on a single technology, it aims for a mechanism that can withstand missing metadata and modifications. For provenance information to actually work, industry-wide cooperation involving platforms and creators is essential, and how widely that spreads looks set to be the focus going forward.
出典:[1] https://openai.com/index/advancing-content-provenance/
