PhD defence: Computational Theory of Mind for Effective Hybrid Intelligence

PLEASE NOTE: The candidate gives a layman's talk, therefore the livestream will start fifteen minutes earlier.

Artificial Intelligence (AI) is becoming part of everyday life, increasingly working alongside humans in complex environments. While AI is highly effective at processing data, there is a growing opportunity to enhance its social understanding, such as interpreting intentions, preferences, and trust.

This thesis explores how to strengthen AI as a collaborative partner by equipping it with a form of social intelligence. The core idea is to design AI agents with computational Theory of Mind: the ability to reason about what and how others think and want. Instead of tracking every possible available information, such as beliefs or intentions of others, the thesis introduces an abstraction-based approach. Here, the abstractions represent human-like concepts such as trust or respect and serve to simplify decision-making, making their reasoning efficient and easy to interpret.

This thesis develops and evaluates the abstraction-based approach in four stages. First, it presents a framework that allows AI agents to reason using high-level abstractions in collaborative settings where humans and AI work together. Second, it shows how AI agents can incorporate human values (especially privacy) into decision-making to improve outcomes. Third, it proposes a modular architecture that keeps an AI agent’s reasoning consistent and adaptable over time. Finally, it introduces methods to model relationships between abstractions like trust and suspicion, enabling richer social interactions. Overall, this research advances Hybrid Intelligence by realizing human-inspired reasoning through computational models. It lays the foundation for AI agents that can collaborate naturally and flexibly with humans in domains like healthcare and social media.