What does double robustness actually guarantee?

Consider an augmented inverse-probability-weighted estimator of an average treatment effect. Under consistency, unconfoundedness and positivity, which statement best describes its double-robustness property?