Towards Causal Foundations of Safe AI - Tutorial

Explore a comprehensive tutorial from the Uncertainty in Artificial Intelligence conference that delves into the causal foundations of safe AI. Learn how Pearlian causality provides a formal framework for reasoning about AI risks and mitigation strategies. Discover causal models of agents and methods for uncovering them. Examine causal definitions of key concepts like fairness, intent, harm, and incentives. Investigate potential risks associated with AI systems, including misgeneralization and preference manipulation. Gain insights into mitigation techniques such as impact measures, interpretability, and path-specific objectives. Presented by James Fox and Tom Everitt, this 1 hour 36 minute session offers valuable knowledge for ensuring ethical and beneficial AI development as its capabilities and societal impact continue to grow.