We measured variability of foot placement during gait to test whether lateral balance must be actively controlled against dynamic instability. The hypothesis was developed using a simple dynamical model that can walk down a slight incline with a periodic gait resembling that of humans. This gait is entirely passive except that it requires active control for a single unstable mode, confined mainly to lateral motion. An especially efficient means of controlling this instability is to adjust lateral foot placement. We hypothesized that similar active feedback control is performed by humans, with fore-aft dynamics stabilized either passively or by very low-level control. The model predicts that uncertainty within the active feedback loop should result in variability in foot placement that is larger laterally than fore-aft. In addition, loss of sensory information such as by closing the eyes should result in larger increases in lateral variability. The control model also predicts a slight coupling between step width and length. We tested 15 young normal human subjects and found that lateral variability was 79% larger than fore-aft variability with eyes open, and a larger increase in lateral variability (53% vs. 21%) with eyes closed, consistent with the model's predictions. We also found that the coupling between lateral and fore-aft foot placements was consistent with a value of 0.13 predicted by the control model. Our results imply that humans may harness passive dynamic properties of the limbs in the sagittal plane, but must provide significant active control in order to stabilize lateral motion.