Third Draft of EU AI Code of Practice Released for Stakeholder Feedback

The third draft of the Code of Practice for general-purpose AI models has been published, incorporating feedback from previous drafts and refining its commitments and measures. The draft introduces a streamlined structure, focusing on transparency and copyright obligations for all providers, alongside specific safety and security commitments for models classified as posing systemic risks. A new Model Documentation Form simplifies transparency compliance, while copyright measures have been clarified. Exemptions for certain open-source models are maintained, aligning with the AI Act.

For providers of advanced AI models with systemic risks, the draft outlines measures such as risk assessments, incident reporting, and cybersecurity obligations. These measures aim to balance clear commitments with flexibility to adapt to technological advancements. The Code is designed to serve as a compliance tool for the AI Act, with finalization expected in May 2025. Stakeholders are encouraged to provide feedback through written submissions, working groups, and workshops.

Complementary to the Code, the AI Office is independently developing a template for public training data summaries as required under Article 53(1)d of the AI Act. It also plans to release guidance clarifying key definitions, responsibilities, and exemptions, such as those for open-source AI models. This guidance aims to ensure a consistent understanding of AI Act rules and their application to general-purpose AI models.

The Code of Practice is a collaborative effort involving over 1,000 stakeholders, including EU Member States and international observers. The Chairs and Vice-Chairs will continue engaging stakeholders through discussions and workshops, with a focus on civil society organizations and downstream industries. Stakeholders have until 30 March 2025 to submit written feedback, contributing to the finalization of this critical framework for AI governance in the EU.