The brain constructs an internal estimate of the gravitational vertical by integrating multiple sensory signals. In darkness, systematic head-roll dependent errors in verticality estimates, as measured by the subjective visual vertical (SVV), occur. We hypothesized that visual feedback after each trial results in increased accuracy, as physiological adjustment errors (A−/E-effect) are likely based on central computational mechanisms and investigated whether such improvements were related to adaptational shifts of perceived vertical or to a higher cognitive strategy. We asked 12 healthy human subjects to adjust a luminous arrow to vertical in various head-roll positions (0 to 120deg right-ear down, 15deg steps). After each adjustment visual feedback was provided (lights on, display of previous adjustment and of an earth-vertical cross). Control trials consisted of SVV adjustments without feedback. At head-roll angles with the largest A-effect (90, 105, and 120deg), errors were reduced significantly (p<0.001) by visual feedback, i.e. roll under-compensation decreased, while precision of SVV was not significantly (p>0.05) influenced. In seven subjects an additional session with two consecutive blocks (first with, then without visual feedback) was completed at 90, 105 and 120deg head-roll. In these positions the error-reduction by the previous visual feedback block remained significant over the consecutive 18–24 min (post-feedback block), i.e., was still significantly (p<0.002) different from the control trials. Eleven out of 12 subjects reported having consciously added a bias to their perceived vertical based on visual feedback in order to minimize errors. We conclude that improvements of SVV accuracy by visual feedback, which remained effective after removal of feedback for ≥18 min, rather resulted from a cognitive strategy than by adapting the internal estimate of the gravitational vertical. The mechanisms behind the SVV therefore, remained stable, which is also supported by the fact that SVV precision – depending mostly on otolith input - was not affected by visual feedback.