Model Checking Human-Agent Collectives for Responsible AI

Humans and agents often need to work together and agree on collective decisions. Ensuring that autonomous systems work responsibly is complex especially when encountering dilemmas. This paper proposes a novel, systematic model checking approach to responsible decision making by a human-agent collective to ensure it is safe, controllable and ethical. Our approach, which is based on the MCMAS model checker, verifies the permissibility of an agent's actions by checking the decision-making behaviour against the logical formulae specified for safety, controllability and ethical behaviour. The verification results through counterexamples and simulation results can provide a judgement, and an explanation to the AI engineer of the reasons actions are refused or allowed.