OpenAI Support for EU Code of Practice on Transparency of AI-Generated Content

OpenAI has announced its support for the European Commission’s Code of Practice on Transparency of AI-Generated Content. This commitment aims to implement the EU AI Act by creating a more transparent digital ecosystem where users can identify the source and origin of AI-generated content.

The Role of Content Provenance in AI Governance

Content provenance provides users with context regarding a piece of content’s source, its creation or editing process, and its authenticity. This transparency is critical for protecting the digital ecosystem from disinformation campaigns and supporting election integrity.

OpenAI asserts that effective provenance methods must be interoperable and built collectively across the value chain, involving news organizations, device manufacturers, AI providers, and online platforms.

OpenAI’s Multi-Layered Provenance Strategy

To ensure provenance signals remain resilient even when metadata is stripped or files are modified, OpenAI employs a multi-layered approach combining different technical signals and product safeguards:

Technical Provenance Signals

• C2PA Content Credentials: OpenAI began integrating C2PA metadata into images created or edited by DALL·E 3 via ChatGPT and the OpenAI API in 2024.
• Dual-Signal Marking: Images generated via ChatGPT, Codex, and the OpenAI API utilize both C2PA metadata for rich information and SynthID watermarks to preserve signals across different contexts.
• Public Verification: The tool at openai.com/verify allows users to check if supported images contain provenance signals associated with OpenAI.

Ecosystem Collaboration and Safeguards

• Open Standards: OpenAI joined the C2PA Steering Committee in 2024 to collaborate with software companies, camera manufacturers, and media organizations.
• Product Safeguards: Provenance signals are used alongside classifiers, reporting channels, and enforcement processes to mitigate deceptive AI uses, such as likeness misuse and election-related deception.

Implementation Challenges and Future Outlook

Provenance is described as a nascent field with significant technical limitations. Metadata can be removed, watermarks can degrade, and labels are only effective where users encounter them.

OpenAI emphasizes that the implementation of the Code of Practice should remain flexible and grounded in methods that work in practice to account for these current technical limitations. The company will continue to develop interoperable standards and improve verification tools in coordination with the AI Office and EU Member States.