Imagine a committee of eleven voters deciding on a controversial policy. The policy passes unanimously, 11 to 0. Now imagine the same policy passing by a narrow 6 to 5 margin. In which scenario does an individual "Yes" voter bear more responsibility for the outcome?
Intuitively, a voter in the 6–5 split feels pivotal; if they had changed their mind, the outcome would have flipped. In the 11–0 landslide, no single individual's flip changes anything. Yet, if the policy causes harm, we still feel the voters in the landslide are collectively at fault.
This tension highlights a fundamental challenge in computer science, law, and philosophy: how do we formally define and calculate responsibility and blame? As autonomous systems increasingly operate in teams and make decisions under uncertainty, we cannot rely on simple "but-for" counterfactuals (i.e., "the outcome wouldn't have happened but for your action"). Instead, a growing body of research focuses on epistemic blameworthiness—disentangling what an agent physically caused from what they could have reasonably foreseen or prevented given their state of knowledge.
The Structural Foundations of Responsibility and Blame
To move beyond binary definitions of causality, researchers utilize structural causal models to quantify degrees of responsibility. As formalized by Chockler et al., responsibility can be treated as a fraction determined by how close an agent was to being a decisive factor. In an 11–0 vote, an individual's responsibility is low because you would need to change five other votes (a large counterfactual shift) before that individual's vote becomes pivotal. In a 6–5 vote, the individual is immediately pivotal, yielding maximum responsibility.
However, Chockler et al. and later Halpern recognize that actual responsibility is purely retrospective—it depends on the true state of the world after the fact. Blame, by contrast, depends on an agent's ex ante epistemic state (their beliefs and information before taking action).
┌──────────────────────────┐
│ True World State │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Actual Responsibility │ (Retrospective, Fact-Based)
└────────────┬─────────────┘
│
Weighted by Agent's Epistemic Uncertainty
│
▼
┌──────────────────────────┐
│ Expected Blame │ (Prospective, Belief-Based)
└──────────────────────────┘
Halpern defines the degree of blame as the expected degree of responsibility across all possible worlds the agent thought they could be in. If an agent takes an action that accidentally triggers a disaster, but they had no way of knowing that the environment was in a highly volatile state, their actual responsibility is high, but their blame is zero.
This perspective is supported by empirical cognitive science. Lagnado et al. demonstrate that human judgments of responsibility do not just look backward at "but-for" counterfactuals; they incorporate prospective assessments of an agent's criticality—how important the agent was perceived to be before the outcome occurred.
Strategic Capability vs. Epistemic Awareness
When moving from static scenarios to dynamic, multi-agent environments, assigning blame requires analyzing what strategies were available to the agents.
In game-theoretic models, a coalition of agents is typically considered blamable for an unwanted outcome if the outcome occurred but the group had a collective strategy to prevent it. However, Naumov et al. identify a major flaw in assuming perfect information: agents cannot be fairly blamed for failing to prevent an outcome if they were completely unaware that a preventive strategy existed.
To capture this, Naumov et al. distinguish between two forms of responsibility in strategic games with imperfect information:
- Strategic Responsibility: The physical capability to guarantee or prevent an outcome.
- Epistemic Blameworthiness: Knowing how to execute that preventive strategy given the agent's current information state.
For example, if a robot can prevent a system failure by pressing a button, it has the strategic capability. But if the robot's sensors are broken and it cannot distinguish between a safe state and a failure state, it does not know when or how to deploy that strategy. In this case, the robot may be causally responsible for the failure, but it lacks epistemic blameworthiness.
This distinction is crucial for understanding "responsibility gaps" in collective decision-making. As Naumov et al. demonstrate, when agents operate under imperfect information, mechanisms that distribute power can inadvertently create scenarios where an undesirable outcome occurs, yet no individual or coalition can be held blameworthy because no group possessed both the information and the power to avert it.
Apportioning Blame in Teams
When a multi-agent system fails, how do we distribute blame among the participants? One common approach borrows the Shapley value from cooperative game theory to fairly divide credit or blame based on an agent's marginal contribution to the group's failure (Halpern et al.; Baier et al.).
However, applying standard cooperative game theory to sequential, uncertain environments introduces severe incentive problems. Triantafyllou et al. show that standard Shapley-based blame attribution can actually penalize high-performing agents or fail to account for what an agent could observe during execution. They propose an uncertainty-aware blame reduction principle: if the system has imperfect observations of an agent's policy or internal state, it should decrease the assigned blame to avoid false accusations.
Similarly, Yazdanpanah et al. leverage temporal logic and concurrent game structures to formalize how providing agents with better sensor data or communication channels shifts them from being merely causally responsible to being legally or morally blameworthy. By explicitly tracking an agent's epistemic horizon over time, their framework prevents the "problem of many hands," where individual accountability dissolves within a complex system.
Active vs. Passive Responsibility
A robust accountability framework must also distinguish between actions of commission (doing something that caused harm) and omission (failing to do something that would have prevented harm).
Mu et al. extend strategic logic to reason about causal active responsibility versus causal passive responsibility in probabilistic environments.
| Responsibility Type | Definition | Example |
|---|---|---|
| Active (Commission) | An agent's explicit choice directly forces a negative outcome. | An autonomous vehicle decides to change lanes into an occupied space. |
| Passive (Omission) | An agent fails to take a known action that would have prevented a negative outcome. | A monitoring system fails to trigger an emergency brake when an obstacle appears. |
By formalizing both dimensions, the authors allow autonomous agents to treat responsibility as a strategic objective during path planning, synthesizing joint plans that minimize their exposure to passive blame in the event of a system-wide failure.
Why This Matters for Trustworthy AI
Building accountable cyber-physical systems requires moving away from ad-hoc logging and toward formal verification of intent and capability. As Pretschner notes, a complete accountability infrastructure must integrate system logs with formal counterfactual causality models to determine why an unwanted state occurred.
Furthermore, we must distinguish between harmless rule-breaking and actual causal deviance. As Datta et al. demonstrate in security protocol verification, an agent might violate a minor protocol rule without that specific deviation being the actual cause of a security breach. Algorithmic accountability must isolate the true causal levers.
Ultimately, frameworks like the one proposed by Cano et al.—which combines Markov Decision Processes with probabilistic model checking—allow us to calculate an agent's "intention-quotient." By quantifying the exact degree of control an agent has over its environment alongside its epistemic state, we can finally distinguish between unavoidable physical accidents and genuine, blameworthy negligence in autonomous systems (Basilico et al.).
Key Foundational Work
Causal Responsibility and Counterfactuals (2013) — contributed 3 concept elements to this direction.
Responsibility and Blame: A Structural-Model Approach (2003) — contributed 2 concept elements to this direction.