About

Research in our lab spans multiple themes and disciplines, from reverse engineering the foundations of human cooperation to the engineering, design, and analysis of cooperative agents and institutions.

Understanding human cooperation and normativity

Human beings are some of the most flexible and intelligent cooperators on the planet. What are the cognitive capacities that enable such cooperation, and how can they be understood in rational terms? We investigate this question at multiple levels, from the modeling of basic social capacities like theory of mind and mentalistic language understanding, to the rational analysis of human norm learning and moral reasoning as tools for cooperating in a world of diverse and plural interests.

Building well-founded and rational cooperative machines

The dominant learning-based approach to AI suffers from weaknesses in reliabilty and coherence, and typically optimizes for single-agent performance rather than multi-agent interaction. In our lab, we believe it is possible to build machines that are instead safe and cooperative by design, either by constructing AI systems out of rational model-based components (e.g. automated planning systems and probabilistic programming languages), or training AI systems via rational incentives for cooperation. Through this approach, we aim to design AI agents that can reliably assist users under uncertainty, negotiate mutually-beneficial outcomes on behalf human principals, and align themselves with fair and cooperative norms.

Designing institutions and incentives for scalable cooperation

For scalable cooperation in a world where actors with divergent interests are increasingly empowered by AI, it is not enough to ensure that AI is capable of cooperating. Instead, to achieve mutual benefit and avoid destructive conflict, we also need to create the institutions and mechanisms that incentivize cooperation at scale. In our lab, we address this through approaches like decentralized norm learning and enforcement and agentic self-governance, with the aim of enabling safe and cooperative futures without centralizing power and control over AI.