Causality
Pearl showed that correlation and cause can be told apart formally, given a separate operator for intervening rather than only observing.
Why it matters
The question of what happens if one intervenes acquired a calculus, which statistics had avoided for a century.
The book introduces the do operator, the back-door criterion, and the means of deciding whether a causal effect is estimable from available data. It answers a long-standing prohibition on speaking of causes in statistics. For machine learning the consequence is direct: a model that has learned correlations does not know what happens when conditions change. Pearl received the 2011 Turing Award for his work on probabilistic and causal reasoning, Bayesian networks and causality together.