Leading models of human planning posit that planning relies on learned forward predictions, from a given state to the outcomes that typically follow it. Here, however, we show that in many situations planning can be made more efficient by relying on backward predictions, from a given outcome to the states that typically precede it. This holds specifically in environments where the number of states an agent may occupy increases with time (i.e., in diverging environments), because in such environments, backward predictions can be more compactly represented than forward predictions. Correspondingly, in three preregistered experiments, we find that humans engage in backward learning and planning in a diverging environment, and in forward learning and planning in a converging environment. We validate the applicability of these findings to real-life learning in a large-scale real-world prediction task. Thus, we establish that humans adaptively deploy forward and backward learning in the service of efficient planning, and this changes how they plan and what decisions they reach.