⚡ Quick Summary
Published by the European Commission, this opinion of 1.8.2025 sets out the Commission's assessment of the General-Purpose AI Code of Practice under Article 56 of Regulation (EU) 2024/1689 (the AI Act). It responds to Article 56(6), under which the AI Office and the European Artificial Intelligence Board must assess whether the Code covers the obligations in Articles 53 and 55; the Board's assessment is published separately.
The document first sets out its legal basis and the Code's drafting process, which opened on 30 September 2024 with a plenary of nearly 1,000 participants and ran through four working groups. It records the Code's structure: 12 commitments in two categories and three documents, with Commitment 1 in the Transparency and Copyright Chapters applying to all providers of general-purpose AI models and Commitments 1 to 10 in the Safety and Security Chapter applying to providers of models with systemic risk, together with a Model Documentation Form and four Appendices.
The assessment covers Article 53(1) documentation and copyright obligations, Article 55(1) model evaluation, systemic risk assessment and mitigation, serious incident reporting and cybersecurity, the clarity of the Code's objectives, its reporting and key performance indicators, and the interests of affected parties. The Commission concludes that the Code adequately covers the obligations in Articles 53 and 55 and meets the aims of Article 56.
🧩 What’s Covered
- Legal basis (Section 1): the Code is drawn up under Article 56 of the AI Act, and Article 56(6) requires the AI Office and the Board to publish their assessment of whether it covers Articles 53 and 55; this document is the Commission's assessment, which does not prejudge enforcement or bind the Commission's interpretation.
- Background (Section 2): the drafting process opened on 30 September 2024 with nearly 1,000 participants and four working groups of independent chairs and vice-chairs; the Code comprises 12 commitments in two categories and three documents, a Model Documentation Form and four Appendices.
- Assessment criteria (Section 3.1): tests drawn from Article 56(2), points (a) to (d) — keeping Article 53(1)(a) and (b) information up to date, the detail of the training-content summary, identification of systemic risks at Union level, and measures for their assessment and management — plus Article 56(4) and (5) on objectives, commitments, indicators and reporting.
- Coverage of Article 53(1) (Section 3.2): Transparency Chapter Measures 1.1 to 1.3 and the Model Documentation Form; Copyright Chapter Measures 1.1 to 1.5 on a single copyright policy, non-circumvention of technological measures, exclusion of piracy websites, rights reservation protocols such as robots.txt, mitigation of infringing outputs and a complaint mechanism.
- Coverage of Article 55(1) (Section 3.3): Safety and Security Chapter Commitments 1 to 10, including Appendix 3 evaluation requirements (rigour, elicitation, robustness, team qualification, external evaluation), Appendix 1 risk identification, Commitments 5 and 6 mitigation, Appendix 4 cybersecurity objectives and Commitment 9 serious incident reporting.
- Objectives, reporting and indicators (Sections 3.4–3.5): Objectives A and B and purpose-clauses in each Chapter; no dedicated reporting commitments for the Transparency and Copyright Chapters, which the Commission considers unnecessary, and no key performance indicators.
- Interests of interested parties (Section 3.6): proportionate documentation, confidentiality references, rightsholder measures and value-chain input such as end-user feedback and downstream incident reporting.
- Review and conclusion (Sections 4–5): formal updates at least every two years, possible rapid guidance or Code update in cases of imminent large-scale irreversible harm, and the conclusion that the Code adequately covers Articles 53 and 55.
💡 Why it matters?
For providers of general-purpose AI models and of models with systemic risk, the opinion maps where voluntary adherence to the Code can demonstrate compliance with Articles 53 and 55 of the AI Act, tying specific Measures to specific obligations — for example Measure 1.1 and the Model Documentation Form for documentation duties, or Appendix 3 for model evaluations. Supervisors, downstream integrators and auditors can see what evidence the Commission expects. It also clarifies that the Copyright Chapter measures operate as a market entry condition and do not settle compliance with Union copyright law, and that the assessment may be revisited.
❓ What’s Missing
The opinion assesses adequacy rather than reproducing the Code, so the commitments, Measures and Appendices are summarised and their text is not included. The Commission records that the Code contains no key performance indicators and no dedicated reporting commitments for the Transparency and Copyright Chapters, which it considers currently unnecessary. The level of detail for the training-content summary is left to a separate AI Office template not reproduced here. The Copyright Chapter measures are stated not to constitute compliance with Union law on copyright and related rights, and the opinion does not bind the Commission's interpretation or prejudge enforcement. The Board's separate assessment is not included.
👥 Best For
Compliance and legal teams at providers of general-purpose AI models mapping Articles 53 and 55 obligations onto the Code's commitments and Measures; policy analysts and regulators tracking how the AI Act's code-of-practice mechanism is assessed; and downstream providers or auditors who need to know what documentation, evaluation and incident-reporting evidence the Commission expects.
📄 Source Details
Commission Opinion of 1.8.2025 on the assessment of the General-Purpose AI Code of Practice within the meaning of Article 56 of Regulation (EU) 2024/1689, reference C(2025) 5361 final, issued by the European Commission in Brussels and signed for the Commission by Henna Virkkunen, Executive Vice-President. Twelve pages; document language English. No URL is printed in the document. The input was the text extracted from all 12 pages.
