“Who Is to Blame?” and “What Is to Be Done?”

Abstract

AI systems are increasingly deployed in safety-critical domains, where they behave non-deterministically, act under uncertainty, and form part of larger multi-agent systems without central control. When such systems cause harm, two questions arise: who is to blame, and what could have been done to prevent it? This thesis develops formal frameworks that make responsibility and causation mathematically precise and algorithmically verifiable, both for attributing responsibility and for identifying the interventions that reduce risk. It is organised in three parts. Part I introduces logics of group responsibility in probabilistic multi-agent systems, based on the risk a group could jointly have avoided, with sound and complete axiomatisations, decidability, and efficient model checking. Part II proposes a new definition of actual cause, Dynamic Causality, studies temporal extension of causal models and their computational properties. Part III connects the two, inducing concurrent game structures from causal models to reason about responsibility in organisational structures. The thesis argues that once non-determinism, strategic ability, uncertainty, and time are made explicit, responsibility and causation admit precise and verifiable formal interpretations.