Abstract
Determining responsibility in multi-agent systems is a complex task that requires accounting for agency, knowledge, and temporal dynamics. This paper introduces a formal framework based on extensive form games of imperfect information to allocate responsibility. By distinguishing between forward-looking capability and backward-looking accountability, and separating strategic from causal responsibility, the authors provide a mathematically rigorous and computationally efficient method for quantifying individual blame using the Shapley value.
Introduction: The Challenge of Responsibility
In any system where multiple agents—human or artificial—interact to produce an outcome, a fundamental question arises: who is responsible, and to what degree? Existing frameworks, such as the Halpern-Pearl (HP) model of actual causality, often rely on structural equations that can be too "thin" to capture the nuances of agency. Specifically, they may struggle with epistemic knowledge (what a player knew at the time), agency (the ability to plan), and temporal sequencing.
The authors propose a bridge between qualitative "folk ethics" and rigorous mathematical modeling. Their framework ensures that responsibility attribution is both philosophically sound and computationally feasible, moving beyond simple physical causality to a more nuanced account of informed decision-making.
Modeling Agency with Extensive Form Games
The primary modeling tool used in this framework is the Extensive Form Game of Imperfect Information. This allows for the representation of moves in a sequence and the use of "information sets" to model what a player knows (or doesn't know) about the state of the game at any given moment.
In the figure above, the dashed line between nodes $s_1$ and $s_2$ indicates that Player 2 cannot distinguish between these states, effectively modeling uncertainty about Player 1's preceding action.
Three Notions of Responsibility
The framework defines three distinct types of responsibility:
- Forward (Prospective): Can a group of agents guarantee a different outcome across all possible future plays?
- Strategic Backward (Retrospective): Given a specific play and the agents' knowledge during it, could they have acted differently to change the outcome?
- Causal Backward: A "but-for" condition. If we freeze everyone else's strategy, would a change in this agent's action have changed the result, regardless of their knowledge?
The distinction between strategic and causal responsibility is a key insight. An agent might be the "cause" of an outcome (causal responsibility), but if they could not have known their action would lead to it based on their information set, they are not strategically responsible.
Quantifying Blame via the Shapley Value
To move from qualitative responsibility to a quantitative score, the authors treat the set of responsible agents as a coalition in a cooperative game. They then apply the Shapley value—a classic power index from cooperative game theory—to distribute a "responsibility score" among individuals. This method naturally handles "responsibility voids" and "over-determination," such as in democratic voting where only the majority voters carry causal responsibility.
Computational Efficiency and Benchmarks
A significant result of the paper is the proof that deciding whether a coalition is responsible is decidable in polynomial time (Theorem 3.14). This offers a major advantage over logic-based frameworks which are often PSPACE-hard. The framework was evaluated against a benchmark of classic thought experiments, including the Bystander Effect, Democratic Voting, and the Firing Squad, demonstrating its alignment with intuitive notions of blame.
Conclusion
This game-theoretic account provides a robust, descriptive model for responsibility allocation. By focusing on the formal relationship between actions, information, and outcomes, it avoids normative judgments while remaining computationally practical. Future work may extend this framework to probabilistic outcomes or games with infinite durations, further broadening its applicability to complex AI-human interactions.
Summary updated