Source: Carnegie Endowment for International Peace

Carnegie Endowment for International Peace presents a comprehensive analysis advocating for a shift in AI regulation focus from individual AI models or their uses to the entities developing frontier AI systems. The authors argue that entity-based regulation, common in sectors like finance and insurance, is better suited to address the complex, rapidly evolving risks posed by advanced AI technologies. Model-based regulation, such as California’s SB 1047, relies on compute thresholds that may quickly become outdated due to innovations like reinforcement learning and multi-agent systems, and faces challenges in defining and measuring training compute accurately.

Entity-based regulation would focus on firm characteristics, such as annual R&D spending, to trigger regulatory oversight, reducing burdens on smaller companies and concentrating resources on major AI developers. This approach allows for transparency requirements, risk management protocols, and governance mechanisms tailored to the entity’s activities, including those unrelated to specific AI models. The paper acknowledges challenges such as potential evasion through corporate structuring and underinclusivity regarding proliferation of dangerous AI capabilities beyond frontier firms. Nonetheless, the authors emphasize that entity-based regulation is essential for managing systemic risks and ensuring responsible AI development, while also highlighting the need for adaptable, well-designed regulatory frameworks that can evolve with technological progress.