When an AI system causes harm, does responsibility disappear or move?

The responsibility gap is one of the founding worries of AI ethics. A system that learns behaves in ways its designers did not specify and could not foresee in detail. When it causes harm, the designers lacked the control that blame usually requires, the users lacked the knowledge, and the system is not a moral agent. So, the worry runs, responsibility falls into a gap between them.

Why I think it moves rather than vanishes

The worry is sometimes framed as if responsibility were a quantity that must be found in one place. Our ordinary practice is less tidy and more accurate. Nissenbaum showed how accountability is spread across the ‘many hands’ that build and run complex systems. We already hold the makers of certified instruments answerable, regulate how instruments are deployed, and expect users to rely on them competently. None of those parties controls every outcome, and none is excused for that reason.

The same structure applies here. Designers answer for the systems they build and the claims they make about them; deployers, for where and how they put them to use; overseers, for the checks they do or do not run. The system itself can be criticised for failing at its function, but that criticism describes a failure; it does not create a new bearer of blame.

What is genuinely hard

It does not follow from responsibility’s moving that its new distribution is easy to draw. Learning systems redistribute control and knowledge in ways that our practices of blame were not designed to track, and some harms arise from interactions no single party could have foreseen. The task is to describe the distribution accurately enough that it can be acted on, by law, by design and by the people who rely on these systems. Read in this way, the gap is a map of where control and knowledge actually sit rather than a hole into which responsibility falls.