Cooperation and Control in Delegation Games @ Intelligent Cooperation Workshop

Lewis Hammond explores delegation games in AI safety, where humans delegate tasks to AI systems in a context of multi-agent problems. He sees both control problems involving aligning preferences and capabilities, and cooperation problems which are aiming for high joint welfare. He emphasizes cooperation problems, as AI systems will likely interact more frequently with one another. On the other hand, he discusses measuring cooperative capabilities using concepts like the price of anarchy and equilibrium selection. He recognizes that understanding and measuring cooperative capabilities will help manage dynamics among multiple AI systems. Given that preventing collusion between machine learning agents is a challenge, he sees the importance of exploring detection and mechanism design to ensure ethical behavior and trust in AI systems.