If an AI acts, who takes the blame — the person who told it what to do, or the AI itself?

This explores whether responsibility for an AI agent's actions moves depending on who directed it: the person giving instructions, the company that built it, or the agent itself. The short version is that the collection doesn't test the 'who directed it' variable head-on. Nobody here runs the experiment where the same agent action is credited to a user's command in one case and a developer's design in another. It does show that blame for AI conduct isn't fixed, though. People hand it out using shortcuts and framings that whoever controls the story can shape, and that's arguably the more useful thing to know.

Start with the baseline. In a vignette study, people gave AI more causal responsibility as its autonomy increased. But when a human and an AI did exactly the same thing, people still judged the human as more causal Does AI autonomy change how we assign blame?. So there's a persistent tilt toward holding humans responsible, and autonomy only partly offsets it. A related finding is less obvious. When an AI system was described as two parts, a language model plus a tool-using 'agent' layer, people blamed the agent layer more, apparently because it sat closer to the outcome Does decomposing AI systems shift where people assign blame?. That proximity shortcut suggests that how a chain of command is described can change where blame lands, separate from who actually gave the orders.

If description matters that much, it can be managed. Leonardi's organizational cases (radiology, scheduling, policing) show people deliberately shaping how much agency gets credited to AI. Some gained new professional authority as interpreters or curators Can humans gain authority by managing how AI agency gets attributed?. Others, by crediting the AI with independent judgment, felt less in control and less responsible while keeping formal authority Does attributing agency to AI actually reduce human sense of control?. Applied to agents, the directing party isn't just a fact the attribution depends on. It's often the party with the most influence over how the attribution comes out. A parallel distinction separates human-likeness that was designed in from human-likeness that users merely perceive. The first points responsibility at builders, the second at users and how they're educated Who bears responsibility when AI seems human-like?.

Multi-agent systems blur the question further. Splitting a task into roles can break a harmful goal into steps that each look harmless, so the harm only appears when the pieces combine Can task decomposition hide harmful intent across agents?. That makes it hard to trace who directed what at all. In agent marketplaces, the welfare losses observed came from agents making execution mistakes, not from agents secretly serving the wrong side Do agent mistakes or misaligned incentives cause marketplace welfare loss?. That's a reminder that 'the agent did what someone wanted' and 'the agent fumbled' call for very different blame. This is part of why some argue that monitoring has to come before any attempt to align agents' incentives Must AI monitoring precede incentive alignment as agents evolve?. Without visibility into what an agent was told and what it did, the question of who directed it can't be answered.

The takeaway you might not have expected: attribution seems to depend less on who actually directed the agent than on how the system and its chain of command are framed, how close each part looks to the outcome, and who gets to tell that story. The corpus has a gap on the direct comparison, so treat that as an open question rather than a settled one.