A First Working Model of Responsibility

The responsibility problem in AI-mediated work becomes clearer when four questions are considered together: who appears autonomous, who has a meaningful choice, who understands the causal pathway, and who can actually prevent the harm. Perceived autonomy explains why an AI–human team can pull responsibility toward the human; meaningful choice, knowledge, and causal influence explain why a tightly constrained overseer can feel less responsible. The practical judgment is not to choose one of those findings as the general rule, but to use the four questions to diagnose the action system in front of us.

Responsibility has more than one dimension

A single score or verbal assignment of blame can hide several different judgments. When an observer says that a human is responsible for an AI-supported outcome, they may mean that the human occupied the formal role, that the human made the final selection, that the human could have prevented the harm, or simply that the human is the most recognizable moral agent in the situation. Those meanings overlap, but they are not identical.

The first dimension is perceived autonomy. In an AI–human team, the AI may be capable of producing influential recommendations yet be viewed as constrained by objectives, guardrails, training, and permissions. The human then appears to be the party with discretion: the one who can interpret the case, challenge the recommendation, and decide whether to proceed. That perception helps explain the AIHR finding that participants assigned more responsibility to the human when the teammate was AI than when the teammate was another human.

The second is meaningful choice. A person may have a formal option to approve or reject a recommendation, but that option matters only if it can change what happens. An interface that removes every outcome-relevant intervention can leave the person with a button but not a decision. Conversely, an interface can narrow a confusing set of choices while preserving a genuine decision among actions that lead to different outcomes.

The third is epistemic access, or the person’s understanding of the relevant situation. The overseer needs more than an alert that something is wrong. They need enough information about the system, the context, and the possible consequences to recognize the failure and select a proportionate response. A human cannot be expected to correct what the interface makes impossible to understand.

The fourth is causal influence. The person must have caused or contributed to the outcome, or possessed a real opportunity to prevent or alter it. Causal influence is not guaranteed by being the last person in the chain. A human who accepts the only action left by an interface may be temporally close to a crash while having little control over it. A developer, deployer, or organization may be farther from the moment of failure but may have shaped the system’s constraints and authority.

These dimensions can point in different directions. A human can be seen as autonomous without having enough information. An AI can be operationally influential without being seen as autonomous.