Regulating ChatGPT and other Large Generative AI Models

Different actors need different rules: The paper identifies distinct roles in the AI value chain—developers, deployers, professional users, non-professional users, and recipients—and argues that regulatory duties should be tailored to each group rather than applying one-size-fits-all rules to everyone involved with generative AI systems.

Regulation should target high-risk applications, not the models themselves: Rather than restricting the development of large generative AI models like ChatGPT, the paper recommends focusing regulatory oversight on specific high-risk uses of these models, allowing innovation while protecting against harmful applications.

Three-layer regulatory approach is needed: The paper proposes minimum standards for all large generative AI models, additional obligations for high-risk use cases, and requirements for collaboration among different actors in the AI value chain to ensure responsible deployment.

Transparency and risk management are essential: The paper emphasizes that regulations must include obligations for developers and deployers to provide transparency about how these systems work and to implement robust risk management practices to identify and mitigate potential harms.

Current AI regulations weren't designed for generative models: Existing regulatory frameworks like the EU's AI Act were built for conventional AI systems and don't adequately address the unique capabilities and risks posed by large generative models, requiring tailored legal approaches.