Should AI Agents Obey Human Laws? (with Cullen O'Keefe)

Finn interviews Cullen O'Keefe about his co-authored paper, "Law-Following AI: Designing AI Agents to Obey Human Laws." O'Keefe argues for aligning AI agents with existing laws, enabling them to refuse illegal actions and orders, distinguishing this approach from intent and value alignment in AI safety. The discussion covers the concept of AI henchmen, potential misuse scenarios in both criminal and governmental contexts, and the challenges of holding AI accountable under current legal frameworks. They explore the complexities of defining "obedience" for AI, the need for a distinct legal category for AI agents, and the practical implications of implementing law-following AI, including government procurement processes and consumer contexts, while also addressing potential resistance and trade-offs.