Can we estimate treatment effects better by accepting more bias?

When using black-box ML models to estimate nuisance functions in causal inference, errors propagate multiplicatively under standard double machine learning—a foundational approach. This work proves you can do better by deliberately under-smoothing: accepting more approximation bias while reducing stochastic error lets first-order noise vanish entirely. The new estimator achieves rate n⁻¹/² + approximation error + (stochastic error)², instead of the usual linear multiplicative dependence. The insight extends to treatment effect and policy learning, suggesting popular orthogonal score methods leave substantial performance on the table.