Delegation Changes the System, Not the Duty

The answer to the opening question is precise: delegating a decision to AI does not automatically delegate moral responsibility because responsibility does not follow the location of an output. It follows the conditions under which an action was produced—who could understand the situation, who had a meaningful choice, who could influence or prevent the outcome, and who designed the roles and constraints around the decision. This final synthesis brings those conditions together so that an AI-supported workflow can be judged for actual control rather than for the reassuring presence of a model or a human checkpoint.

The two supplied research papers reveal why a simpler answer fails. In one set of studies, adding AI to a human team makes the human appear more responsible after an ambiguous moral failure. In the other, restricting the human overseer’s choices makes that person feel less responsible even as performance improves. Those movements are not contradictory. They show that “AI delegation” is not one intervention. It can change who appears autonomous, who can choose, and who can understand and alter what happens next.

Responsibility does not travel with the output

A model can generate the recommendation. A human can approve it. An organization can require the workflow. A developer can shape the model and interface. A deployer can decide where the system is used. The outcome can emerge from all of those links. Saying that the decision was “delegated to AI” describes one transfer of work, but it does not answer which links in the action system remain morally or institutionally significant.

This is why three forms of responsibility must remain distinct.

Perceived responsibility is what participants or observers assign. It is the judgment that one actor looks more answerable than another. The AIHR studies show that this perception can move toward the human when the AI is framed as a teammate and is treated as structurally constrained. The human becomes the apparent bearer of discretion even when the scenario gives no factual reason to say that the human caused more of the harm.

Actual causal contribution concerns what each actor could do and did do. The human-oversight study shows why a person’s presence at the final intervention point is not enough. If the interface has removed the actions that could prevent the outcome, the person’s causal connection may be weak. The location of a click is not the same as control over the decision.

Institutional responsibility concerns the people and organizations that establish the conditions under which decisions are made. Who chose the model? Who set the objective? Who approved the data and deployment context? Who decided that a recommendation was mandatory, that an override was available, or that escalation was unnecessary? The supplied experiments do not provide a complete allocation across these parties, but they make clear why such an allocation cannot be inferred from participant blame alone.

These categories can diverge without any of them being imaginary. A frontline worker may genuinely feel responsible because an AI–human team frame makes their discretion salient. That feeling may shape whether they report a problem or accept a sanction. Yet the organization may still need to investigate whether the worker could understand the model’s failure or prevent the outcome. Conversely, a person may feel less responsible because an interface eliminated meaningful choice even though the organization formally assigned them oversight. Their feeling does not settle every question, but it is evidence that the work design may have weakened the basis for the assignment.