Perceptual learning is reward-based. A recent mathematical analysis showed that any reward-based learning system can learn two tasks only when the mean reward is identical for both tasks [Frémaux, Sprekeler and Gerstner, 2010, The Journal of Neuroscience, 30(40): 13326-13337]. This explains why perceptual learning fails when two differing stimulus types are presented randomly interleaved from trial to trial (i.e. roving), even though learning occurs efficaciously when the stimuli are presented in separate sessions. Hence, the unsupervised bias hypothesis makes the surprising prediction that no perceptual learning occurs when a very easy and a hard task are roved because of their different rewards. To test this prediction, we presented bisection stimuli with outer-line-distances of either 20’ or 30’. In both tasks, observers judged whether the central vertical line was closer to the left- or right-outer line. Task difficulty was adjusted by manipulating the center line’s offset. Easy and difficult discriminations corresponded to 70 and 87 percent correct respectively. In accordance with theoretical predictions, subjects failed to learn in this roving task for both bisection-stimulus types. Hence, surprisingly, perceptual learning of a hard task can be disturbed by performing a simple, undemanding task.