On Thursday 5 March, the European Commission published the second version of the ‘Code of Practice on Transparency of AI-Generated Content’ (see EUROPE 13775/22), which is intended to provide model providers with details on compliance with the transparency rules laid down in Article 50 of the AI Act.
Divided into two parts, the ‘Code’ details rules for providers of generative AI systems, which focus on marking AI-generated content and technical solutions, and rules that apply to deployers of the systems, covering the disclosure of deepfakes and AI-generated or manipulated content on matters of public interest.
The ‘Code’ covers all the technologies that can be considered ‘marking’ technologies to indicate whether content is generated or manipulated by AI: watermarks, labelling, digital signatures, cryptographic metadata, fingerprints, etc.
In this new version, several elements that were mandatory in the first version become optional, in the name of “greater flexibility and clarity”, according to the European Commission’s press release.
Providers and deployers are only “encouraged” to implement certain technical features, such as a “common symbol” for reporting AI-generated deepfakes, or to provide channels for individuals or third parties (trusted flaggers, independent fact-checkers, etc.) to flag missing or incorrect information.
The ‘Code’ itself is already organised as a general recommendation for compliance, and not as a binding text. This weakening of obligations could render its very usefulness obsolete.
The Commission will invite comments on the second draft by 30 March. The final version of the ‘Code’ should be completed by early June this year. Unless changed by the revision of the AI Act (see EUROPE 13822/13), the rules on transparency of AI-generated content will come into force on 2 August 2026.
See the ‘Code’: https://aeur.eu/f/l28 (Original version in French by Isalia Stieffatre)