Understanding the General-Purpose AI Code of Practice
Today, the European Commission released the First Draft of the General-Purpose AI Code of Practice, marking another significant step in establishing guidelines for safe and responsible AI in the EU. This draft sets foundational principles for transparency, copyright, and risk assessment in general-purpose AI models. Open for feedback, it aims to guide AI providers toward compliance under the AI Act by 2025.
1. What is the Code of Practice?
The General-Purpose AI Code of Practice is a regulatory framework under the European Union’s AI Act, aimed at guiding the responsible development, deployment, and oversight of general-purpose AI (GPAI) models. It sets rules and standards to ensure that AI systems are transparent, safe, and aligned with EU principles, covering areas like transparency, copyright, and risk assessment. The Code focuses on managing systemic risks associated with advanced capabilities and widespread use, addressing the need for trustworthy and secure AI within the EU.
2. The Process of Creating the Code of Practice
The development of the Code of Practice follows an iterative, multi-stakeholder process. Led by the European AI Office and chaired by independent experts, the drafting involves input from a diverse range of contributors, including AI providers, civil society, academia, and international observers. More than a thousand stakeholders are involved, organized into four working groups focused on specific aspects like transparency, risk assessment, technical safety, and governance. These groups meet regularly to review and refine the draft, with the goal of finalizing the Code by April 2025.
3. The First Draft
The First Draft of the General-Purpose AI Code of Practice, now open for stakeholder review, establishes foundational guidelines for responsible AI usage. It introduces core principles, measures, and objectives, including transparency, copyright compliance, and structured risk assessment. While preliminary, this draft provides a broad framework of commitments, such as documentation standards, transparent usage policies, and protocols for identifying and mitigating systemic risks. It is designed to align with EU regulatory standards and includes open questions to guide feedback and further development.
领英推荐
4. The Next Steps
Following the release of the First Draft, stakeholders will provide feedback through a designated platform until late November. This feedback will help refine the document, adding specificity to the proposed measures and introducing key performance indicators (KPIs). Additional discussions in the upcoming Plenary will allow stakeholders to contribute further, guiding the draft’s evolution. The iterative process will continue, aiming to deliver a finalized, enforceable version of the Code by April 2025.
5. Taxonomy of Systemic Risks: A Preview of the First Draft
The First Draft presents a Taxonomy of Systemic Risks, which identifies and categorizes potential threats posed by general-purpose AI. This taxonomy provides a structured approach to assess and manage risks, classified by type, nature, and sources. Key risk types include:
The taxonomy also defines risk sources and characteristics, considering factors such as actors, motivations, and the likelihood and impact of these risks. This framework emphasizes the importance of continuous risk monitoring, transparency, and proactive measures to address high-stakes scenarios effectively.
The full draft of the Code of Practice is available for review here for anyone interested in exploring the document in greater detail.