Computing the value of a choice in multi-step planning tasks is typically thought to require simulating its resultant sequence of task states. In this talk, I’ll present evidence that the brain simplifies this potentially intractable computation using the successor representation (SR), which forgoes simulation and instead stores and re-uses predictions about states that will occur multiple time-steps following the choice. Such predictions are learned from experience and used to provide a representation for choices under evaluation. I’ll first present a behavioral study demonstrating that, following task changes, humans make a pattern of errors, predicted by the SR, in altering choices. I’ll then discuss an FMRI study providing evidence that previously documented neural markers of predictive representations, observed in sensory cortex pre-activation, as well as in the representations of task stimuli in hippocampus and medial prefrontal cortex, relate to this pattern of errors, thus providing neural evidence for this approach. Overall, this work contributes towards a mechanistic understanding for how the brain accomplishes multi-step planning, as well as a refined understanding as to why behavior is often inflexible in the face of change.