Brussels, 19 December 2024
Independent experts present the second draft of the General-Purpose AI Code of Practice, based on the feedback received on the first draft, published on 14 November 2024. Each draft is a work in progress and reflects the views of stakeholders participating in the Code of Practice Working Groups and Provider Workshops, consisting of around 1000 stakeholders, including EU Member State representatives and European and international observers.
The content of the second draft was informed by Working Group meetings which were held across the week of 18 November 2024. Participants had the opportunity to provide verbal feedback, ask the Chairs questions, and answer questions from the Chairs via interactive polls.
A summary of the content discussed at these Working Group meetings was presented to the Plenary. Chairs also received 354 submissions of written feedback on the first draft through a dedicated platform, and considered feedback from providers of general-purpose AI models presented in dedicated workshops, where Minutes were taken and will be shared with all plenary participants to maintain transparency.
Chairs and Vice-Chairs were also invited to a series of interinstitutional meetings:
- On 27 November, they presented the first draft of the Code to the first meeting of the European AI Board Steering Group on General-Purpose AI, which gathers national experts from the 27 Member States.
- On 10 December, they outlined their work at the second meeting of the AI Board, a key advisory body created by the AI Act and composed of representatives of the 27 Member States.
- On 11 December, the Chairs and Vice-Chairs presented the ongoing work per Working Group to the joint IMCO-LIBE Working Group of the European Parliament on the Implementation and Enforcement of the AI Act.
Status of the second draft
The Code is a guiding document for providers of general-purpose AI models in demonstrating compliance with the AI Act along the full life cycle of the models.
The second draft builds upon previous work while aiming to provide a “future-proof” Code; it will be particularly relevant for models which will be released after 2 August 2025 when the new rules on general-purpose AI models start to apply.
The first part of the draft Code details transparency and copyright obligations for all providers of general-purpose AI models, with notable exemptions for providers of certain open-source models in line with the AI Act. Regarding the template for an adequate summary of training data content, a proposal can be expected from the AI Office early next year. The AI Office aims to take into account comments from the Code of Practice participants before the template is officially adopted by the Commission.
The second part of the Code is only relevant for a small number of providers of most advanced general-purpose AI models that could pose systemic risks, in accordance with the classification criteria in Article 51 of the AI Act. Here the Code outlines measures for systemic risk assessment and mitigation, including model evaluations, incident reporting, and cybersecurity obligations. The Code structure was adapted to outline objectives, commitment, and measures, (replacing the previous draft’s structure consisting of goals, measures, and sub-measures), and include preliminary examples of Key Performance Indicators (KPIs).
An attempt was also made to clearly point out the obligations stemming from the AI Act, which the Code aims to further specify through concrete measures that the Signatories commit to undertaking in order to effectively fulfil these obligations.
In light of the evolving state of the art, the need to balance clear commitments with the flexibility to adapt as technology evolves has been stressed. The need for further development of ecosystems for AI governance and risk management has been stressed. The Chairs emphasised that the second draft remains a work in progress. They focused primarily on providing clarifications, adding essential details, and aligning to the principle of proportionality, such as the size of the general-purpose AI model provider. In subsequent iterations, the Chairs will work towards ensuring that all parts of the Code fit together seamlessly and are easy to understand, alongside other refinements and provision of further detail.
Next steps
In addition to the written feedback received by 15 January 2025, verbal discussions on the second draft with Chairs and Vice-Chairs are planned in line with the tentative timeline published by the AI Office.
The next Working Group (WG) meetings will take place in the week of 13 January 2025, concretely:
- Monday, 13 January: WG3 Technical risk mitigation for systemic risk
- Tuesday, 14 January: WG1 Transparency
- Wednesday, 15 January: WG2 Risk assessment for systemic risk
- Thursday, 16 January: WG4 Governance risk mitigation for systemic risk
- Friday, 17 January: WG1 Copyright-related rules.
Workshops with general-purpose AI model providers and Member State representatives in the AI Board Steering Group are planned for the weeks of 20 and 27 January respectively.
The third draft of the Code of Practice can be expected for the week of 17 February 2025.
This Q&A, written by the AI Office, provides insights into the regulatory approach to general-purpose AI in the AI Act.
Read the 2nd Draft
Further information
Source – EU Commission