EU Drafts Code of Practice for General-Purpose AI
Independent experts has prepared a draft Code of Practice for providers of general-purpose AI models, with feedback open until November 28. The Code aims to guide compliance with the AI Act, focusing on transparency, risk assessment, and mitigation, particularly for powerful AI models deemed to have systemic risk. The draft is currently a high-level document, and future iterations are expected to expand significantly as industry and civil society feedback is incorporated. The Code emphasizes the importance of proportionality, tailoring requirements to the size and capacity of providers, and considering different distribution strategies.
The draft Code outlines transparency requirements, including the disclosure of web crawlers used in AI development and handling copyrighted material. This is crucial as AI developers face legal challenges over data acquisition methods. Additionally, GPAI providers must establish a single point of contact for copyright-related grievances. The Code also proposes documentation standards for data sources used in AI training, testing, and validation.
Systemic risk management is a key focus, with specific rules for GPAIs using over 10^25 FLOPs. The Code identifies various systemic risks, such as cybersecurity threats, misinformation, and discrimination. It encourages providers to recognize additional risks, like privacy breaches and public health threats. The draft also suggests a Safety and Security Framework for ongoing risk assessment and management, requiring providers to forecast potential systemic risks and notify the AI Office of any substantial risks.
The draft Code, influenced by international standards and extensive stakeholder consultation, remains provisional. It seeks further input to refine measures, sub-measures, and KPIs, aiming for a comprehensive final version by May 1, 2025. This initiative reflects the EU’s commitment to regulating AI development responsibly, balancing innovation with public safety and ethical considerations.
