This paper explores the use of deep reinforcement learning to enhance cooperation in human groups. A 'social planner' AI agent was trained using simulation to make recommendations for connecting or disconnecting individuals within a network cooperation game. The trained agent significantly improved cooperation rates in groups of human participants playing for real monetary stakes, achieving a 77.7% cooperation rate compared to 42.8% in control groups. Unlike previous methods that separate defectors, the social planner employed a conciliatory approach, guiding defectors towards smaller, cooperative groups.

Kevin R. McKee, Andrea Tacchetti, Michiel A. Bakker, Jan Balaguer, Lucy Campbell-Gillingham, Richard Everett, Matthew Botvinick