If an AI acts entirely on its own, does that make it easier — or harder — to pin down who's responsible?

This explores whether an AI system has to act on its own before law or society can hold it, or the people behind it, responsible for what it does. The collection can't give you a legal ruling. It has no notes on court decisions or statutes about AI liability. What it does have is three related lines of work. One measures how people actually assign blame. One borrows from agency law, the body of law that covers people who act on someone else's behalf. One treats autonomy as a setting that designers choose. Together they suggest the question is less about whether AI is autonomous enough and more about who decided how autonomous it would be.

Start with how people judge blame, which is where law often begins. In vignette studies, where people react to short written scenarios, participants gave AI more causal responsibility as its independence increased. But when a human and an AI did exactly the same thing, people still saw the human as more of the cause Does AI autonomy change how we assign blame?. So autonomy affects blame without fully deciding it. A follow-up finding complicates things further. When an AI system was described as two parts, a language model plus a tool-using 'agent' layer, people put more blame on the agent part. It wasn't more autonomous. It was just closer to the outcome Does decomposing AI systems shift where people assign blame?. That means intuitions about responsibility depend partly on how a system is described, which could matter a lot if those intuitions shape legal arguments.

The closest the collection gets to legal doctrine is agency law and principal-agent economics, which study the problems that arise when one party acts for another. They point to three core problems: the agent knows things the principal doesn't, the agent has room to make its own decisions, and the agent's loyalty can't be guaranteed. The usual fixes are incentives, monitoring, and enforcement. The note argues those fixes break down when an AI agent makes decisions nobody can interpret, at a speed and scale no one can follow Can traditional governance tools actually control AI agents?. In that framework, the accountability problem appears as soon as an agent has some discretion, well before it is fully autonomous. Autonomy doesn't create the responsibility question. It makes the normal tools for answering it weaker.

The idea you may not expect to find useful is that autonomy is chosen rather than inherent. One line of research separates autonomy from capability. A very capable agent can be set to check with the user before every action, across five levels that range from the user as operator to the user as observer Can we design AI autonomy separately from capability?. Other work splits this choice into three places: the model's rules for when to escalate to a human, the software around the model that enforces those rules, and how delegation is shown to the user How should AI agents balance autonomy with user oversight?. If autonomy is a setting, then responsibility can be traced back to whoever chose the setting. The same reasoning appears in work on human-like AI. Features a designer deliberately builds in and qualities a user merely perceives lead to different responsible parties Who bears responsibility when AI seems human-like?. Deloitte describes the practical version: humans placed at planned decision points, with cryptographic receipts and action logs that can't be altered, so there is always a record of who approved what How should organizations supervise autonomous AI agents?.

There is a catch. Several notes warn that more autonomy wears down the human oversight that accountability depends on. Users understand less of what the agent is doing, and their judgment weakens with long use Does granting agents more autonomy undermine human oversight?. Others argue that risk grows steadily with autonomy and that fully autonomous agents shouldn't be built at all Does AI risk increase with the autonomy we give it? Should AI systems stay collaborative rather than fully autonomous?. Against that, Ball argues that self-directed 'self-sovereign' agents will appear anyway, driven by capability and economic pressure. If no human principal remains, responsibility frameworks built on a principal have nothing to attach to Will self-sovereign AI agents inevitably emerge despite policy efforts?. That is why some argue binding government oversight is needed and companies can't police themselves Can companies alone manage the risks of AI systems?. The open question the corpus leaves you with is not whether AI is autonomous enough to blame. It is what happens to accountability once nobody is left who chose the level of autonomy.