EU General-Purpose AI Code of Practice
2025-07-10 · Publisher : Commission européenne · Domain : Artificial Intelligence
The Commission publishes the Code of Practice for general-purpose AI models: a voluntary three-chapter tool that lets providers demonstrate compliance with the AI Act ahead of the August 2, 2025 deadline.
What changes
- Three chapters: transparency and copyright for all providers, safety and security for models with systemic risk.
- Drafted by 13 independent experts with input from more than 1,000 stakeholders.
- Signatories benefit from a reduced administrative burden and greater legal certainty.
- Most major providers sign; Meta refuses, xAI signs only the safety and security chapter.
- The code turns still abstract legal obligations into concrete practices.
On July 10, 2025, a few weeks behind the initial schedule, the European Commission publishes the General-Purpose AI Code of Practice. It is meant to help model providers meet the AI Act obligations that apply from August 2, 2025. ### What happened - **Drafting**: thirteen independent experts wrote the text over several drafts, with input from more than 1,000 stakeholders (model providers, SMEs, researchers, safety experts, rightsholders, civil society). - **Transparency**: a model documentation form gathers the information to provide to the AI Office and to downstream integrators. - **Copyright**: a copyright compliance policy, respect for machine-readable rights reservations (including robots.txt), exclusion of sites recognized as infringing, complaint handling. - **Safety and security**: reserved for models with systemic risk (more than 10^25 FLOP of training compute), this chapter requires a risk assessment and mitigation framework and reporting of serious incidents. - **Validation**: the Commission and the European AI Board later confirmed the code as an adequate compliance tool. ### Why it matters The code acts as an operating manual for the AI Act for foundation models. Signatories, including Anthropic, Google, Microsoft, Mistral AI and OpenAI, benefit from a presumption of good faith. Others will have to demonstrate compliance by other means, under closer scrutiny. ### Caveats The code remains voluntary. Meta refused to sign, arguing that it created legal uncertainty and went beyond the regulation. Some rightsholders found it insufficiently protective, and European companies separately called for a pause in AI Act enforcement. ### Key takeaways For model providers, the code turns the AI Act into a list of verifiable practices; for user companies, signing the code becomes a criterion when choosing a provider.
Links
Tags : IA, Régulation, AI Act, Europe, Droit d'auteur