Explanation in Artificial Intelligence: Insights from the Social Sciences

Tim Miller was an associate professor (now professor) in the School of Computing and Information Systems at the University of Melbourne when he wrote this, and it is the paper he is best known for. His background spans agent-based AI and human–computer interaction, which positions him to argue that a subfield dominated by machine-learning engineers had overlooked decades of work in the human sciences on what explanation is and does.

The paper appeared in 2019 in Artificial Intelligence, one of the field's senior peer-reviewed journals, after a 2017 arXiv preprint; the published version runs thirty-eight pages and draws on a very large reference list spanning philosophy, psychology, and cognitive science. It is a survey with an argument: Miller assembles findings on how people actually explain — that explanations are contrastive, picking out why one outcome rather than another; that people select a few causes from many; that explanation is a social act, a transfer of understanding between explainer and audience — and presses the claim that explainable AI should be designed to those findings instead of to researchers' intuitions about what looks like a good explanation.

The piece sits as a secondary, synthesising work, and a foundational citation for the human-facing wing of explainable AI. It extends the human-grounded strand of Doshi-Velez and Kim's evaluation scheme, and it imports into the machine debate exactly the literature the human-black-box sources belong to — the psychology of how people account for their own minds, which in Nisbett and Wilson turns out to be unreliable. Where Lipton uses human opacity to deflate the demand for interpretable machines, Miller uses the social science of human explanation to raise the standard for machine explanation; the two import the same field for opposite purposes. The question of what it is to explain an agent at all connects to Dennett's intentional stance.

Miller's stake is academic and agenda-setting, with no commercial dimension. The paper effectively defined a research programme — "XAI should learn from the social sciences" — and the very large citation count that followed is itself the reputational return, which a reader should weigh as an incentive shaping how confidently the case is put rather than as a conflict of interest.