We consider the problem of explaining a data generating process (DGP) to a decision maker (DM) who cannot understand it. Explanations are information, not approximations: they are useful because they rule out possible DGPs. If the DM maximizes her average payoff, explanations using OLS are robustly useful, and augmenting them with summary statistics makes them more so. Sampling error can destroy these guarantees, but theoretical assumptions linking that error to the DGP can restore them. If the DM is sufficiently ambiguity averse, explanations that are affine in outcomes are not robustly useful, but certificates reporting lower bounds on outcomes are.
