Plugging ML predictions directly into a causal contrast smuggles in regularization bias: shrinkage leaves confounding signal in the residuals, and it leaks into the effect estimate. DML fixes this with orthogonalization, the Frisch-Waugh-Lovell move: predict the outcome from confounders, predict the treatment from confounders, and regress the first residual on the second, so nuisance-model errors only enter at second order.
Cross-fitting closes the remaining gap, with each residual formed by a model that never trained on that row, letting arbitrarily flexible learners serve as the nuisance models without overfitting bias. EconML’s DML estimators package the recipe and extend it to effect heterogeneity.