OpenAI Ties C2PA Provenance to EU Code of Practice for AI Content

OpenAI started embedding C2PA metadata into DALL-E 3 images in 2024, and now it's formally backing the European Commission’s Code of Practice on Transparency of AI-Generated Content. That's not a PR gesture—it’s a signal that the company expects provenance requirements to stick, and it wants its technical choices to become the de facto standard.

Why Provenance Is a Multi-Layered Problem

Metadata alone doesn't cut it. C2PA (Coalition for Content Provenance and Authenticity) metadata can be stripped by a simple re-upload, lost during file format conversion, or trivially removed with a screenshot. OpenAI knows this, which is why they layer SynthID watermarks—imperceptible signals that survive transformations—on top of the metadata. Watermarks carry less information but persist where metadata fails.

What OpenAI's Stack Actually Does

Since 2024, every image generated or edited by DALL·E 3 in ChatGPT and the API carries both C2PA metadata and a SynthID watermark. That dual approach is the core of their provenance stack. They also run a public verification tool at openai.com/verify that checks whether a given image contains OpenAI-originated provenance signals. In 2024, OpenAI joined the C2PA Steering Committee alongside news orgs, camera makers, and platforms. A year earlier, in 2025, they became the first US company to sign the EU’s General-Purpose AI Code of Practice—setting the stage for this new Code.

The Limits of Current Techniques

Provenance is still a nascent field. OpenAI admits that signals degrade across the web: metadata gets removed, watermarks fade, and labels only help where people actually encounter content. Their contribution to the Code is essentially a bet that mandatory transparency requirements will push the ecosystem to adopt interoperable standards—even if no single technique is bulletproof. The Code stays grounded in what works today while leaving room for better methods.

OpenAI will keep strengthening verification tools and working with the AI Office and Member States, but the real test is whether these signals survive the upload-download sausage grinder long enough to actually inform a viewer's trust decision.

Source: Supporting Europe's work in ensuring a trustworthy AI ecosystem
Domain: openai.com