Causal inference promises to answer “what if” questions from observational data. But the standard approach — specify a parametric model, fit it, report the estimate — breaks down under misspecification. And misspecification is the norm, not the exception.