Human walking has previously been described as "controlled falling." Some computational models, however, suggest that gait may also have self-stabilizing aspects requiring little central nervous system control. The fore-aft component of walking may even be passively stable from step to step, whereas lateral motion may be unstable and require motor control for balance, as through active foot placement. If this is the case, walking humans might rely less on integrative sensory feedback, such as vision, for anterio-posterior (AP) balance than medio-lateral (ML). We tested whether healthy humans (N = 10) exhibit such direction-dependent control, by applying low-frequency perturbations to the visual field (a projected virtual hallway) and measuring foot placement during treadmill walking. We found step variability to be nearly ten times more sensitive to ML perturbations than AP, as quantified by the increase in root-mean-square step variability per unit change in perturbation amplitude. This is not simply due to poorer physiological sensitivity of vision in the AP direction: Similar perturbations applied to quiet standing produced reversed direction dependence, with an AP sensitivity 2.3 times greater than ML. Tandem (heel-to-toe) standing yielded ML sensitivity 3.0 times greater than AP, suggesting that the base of support influences the stability of standing. Postural balance nevertheless appears to require continuous, integrative motor control for balance in all directions. In contrast, walking balance requires step-by-step, integrative control for balance, but mainly in the lateral direction. In the fore-aft direction, balance may be maintained through an "uncontrolled," yet passively stabilized, series of falls.