Causal inference in artificial intelligence

In recent years, artificial intelligence researchers have paid a great deal of attention to the statistical literature on recursive structural equation models and graphical models, especially to the literature on directed acyclic independence graphs. Some researchers have argued that graphical models are useful in studying causal relationships among variables, and that causal relations can be read off the conditional independence graph. Some authors have even attempted to use the data to derive both a causal ordering and causal relationships. A number of others have refrained from imparting a causal relationship to graphical models. This paper discusses the subject of causation and causal inference in general, and then applies the discussion to the case of graphical models. Specifically, a distinction between causation and causal inference is made; the inferential activity consists of making statements about causation. Using a counterfactual account of the causal relation that derives from work in statistics, I show that the usual types of inferences made in the structural equations literature and the graphical models literature do not square with this account; the results are relevant because a number of authors have claimed otherwise