Previous research on human balance recovery suggests that, prior to an externally triggered postural perturbation, healthy subjects can pre-select their postural response based on the environmental context, but it is unclear whether this pre-selection includes the selection of a stepping leg when performing compensatory steps. We sought to determine how pre-selecting a stepping limb affects the compensatory steps and stability of young, healthy subjects when responding to postural perturbations. Nine healthy subjects (24–37 years of age) stepped in response to backward translations of a platform under their feet when, prior to the perturbations, the subjects either knew whether they were to step with their left or right leg to a visual target (the Predictable condition) or did not know whether to step with their left or right leg until one of two targets appeared at perturbation onset (the Unpredictable condition). The Unpredictable condition also included randomly inserted trials of toes-up rotations and backward translations without targets (catch trials). The results showed that, in the Predictable condition, the subjects consistently exhibited one anticipatory postural adjustment (APA; a lateral weight shift toward the stance limb) before stepping accurately to the target with the correct leg. In the Unpredictable condition, the subjects either (1) exhibited multiple APAs, late step onsets, and forward center-of-mass (CoM) displacements that were farther beyond their base of support, or (2) exhibited an early step with only one APA and kept their CoM closer to the base of support, but also stepped more often with the incorrect leg. Thus, when the subjects had to select a stepping leg at perturbation onset, they either became more unstable and used multiple APAs to delay stepping in order to provide enough time to select the correct stepping leg, or they stepped earlier to remain stable but often stepped with the incorrect leg. In addition, responses to catch trials in the Unpredictable condition included distorted step placements that resembled steps to anticipated targets, despite allowing the subjects to step with a leg of their choice and to a location of their choice. Lastly, the subjects’ voluntary stepping latencies to visual targets presented without perturbations were twice as long as their stepping latencies to the backward platform translations. Therefore, healthy subjects appear to pre-select their stepping limb, even when the perturbation characteristics are unpredictable, because relying on visual input provided at perturbation onset requires a delayed response that leads to greater instability.