First draft of the General-Purpose AI Code of Practice

IRIS 2024-10:1/10

Amélie Lacourt

European Audiovisual Observatory

While the EU AI Act came into force on 1 August 2024, the rules on general-purpose AI will, for their part, become effective in August 2025. As part of the implementation process, and as mandated by Article 56 of the AI Act, the AI Office is to encourage and facilitate the drawing up of codes of practice at Union level in order to contribute to the proper application of the act’s provisions on general-purpose AI models. These notably include provisions related to transparency, copyright, systemic risk taxonomy, risk assessment, and mitigation measures.

The development process of this code notably involves:

- a multi-stakeholder consultation;

- four specialised working groups led by chairs and vice-chairs selected for their expertise, experience, and independence, as well as to ensure geographical and gender diversity;

- discussions and drafting sessions to be held between October 2024 and April 2025.

On 30 September 2024, the AI Office launched the inaugural event for the Code of Practice Plenary, with nearly 1 000 participants, including a high number of professional organisations. The four working groups, which began their work in October 2024, are each expected to meet three times to discuss various drafts. The four groups focus on:

- transparency and copyright-related rules;

- risk identification and assessment for systemic risk;

- technical risk mitigation for systemic risk;

- governance risk mitigation for systemic risk.

In addition, a workshop held on 23 October 2024 also allowed the main addressees of the code (providers of general-purpose AI models) to contribute to the process.

On 14 November 2024, independent experts unveiled a first draft of the General-Purpose AI Code of Practice. The chairs and vice-chairs presented this first draft as a foundation for further detailing and refinement, inviting feedback to help shape each iteration towards the final version of the code. They also outlined guiding principles and objectives for the code, aiming to provide stakeholders with a clear sense of the direction of the final code’s potential form and content. Open questions were included to highlight areas for further progress.

The drafting principles notably stress that measures, sub-measures, and Key Performance Indicators should be proportionate to the risks, take into account the size of the general-purpose AI model provider, and allow simplified compliance options for SMEs (small and medium enterprises) and start-ups. Following the AI Act, the code will also reflect notable exemptions for providers of open-source models. The principles also highlight the need for a balance between clear requirements and the flexibility to adapt as technology evolves. According to the chairs and vice-chairs, this draft seeks to provide a “future-proof” code that would also be appropriate for the next generation of models that will be developed and released in 2025 and thereafter.

This first draft concludes the first of four drafting rounds until April 2025. The final version of the code should be ready by 1 May 2025.