Social dilemmas are ubiquitous in the world of human social interactions. The human ability to solve complex spatially and temporally extended social dilemmas is a key hallmark of our intelligence. Conversely, the cumulative nature of social tasks may provide an automatic path towards increasingly powerful cognitive abilities, under appropriate constraints. In this talk, I will argue for the ubiquity of social dilemmas from some simple principles, and show how complex spatially and temporally extended dilemmas may be solved by including appropriate inductive biases in state-of-the-art deep reinforcement learning agents. This approach simultaneously advances the fields of behavioural economics and artificial intelligence. I shall end with an outlook on how these fields might interact more closely in the future, both to improve our understanding of multi-agent dynamics in human society, and to construct generally intelligent artificial systems.