This study explores how human users respond to feedback and evaluation from a robot. A between-subjects experiment was conducted using the Wizard of Oz method, with 63 participants randomly assigned to one of three evaluations (good evaluation vs. neutral evaluation vs. bad evaluation) following a training session. When participants attempted to reproduce the physical motion taught by the robot, they were given a verbal evaluation of their performance by the robot. They showed a strong negative response to the robot when it gave a bad evaluation, while showing positive attraction when it gave a good or neutral evaluation. Participants tended to dismiss criticism from the robot and attributed blame to the robot, while claiming credit to themselves when their performance was rated positively. These results have theoretical implications for the psychology of self-serving bias and practical implications for designing and deploying trainer robots as well as conducting user studies of such robots.