This paper presents the results of a laboratory experiment in which workers perform a real-effort task and supervisors report the workers' performance to the experimenter. The report is non verifiable and determines the earnings of both the supervisor and the worker. We find that not all the supervisors, but at least one third of them bias their report. Both selfish black lies (increasing the supervisor's earnings while decreasing the worker's payoff) and Pareto white lies (increasing the earnings of both) according to Erat and Gneezy (2009)'s terminology are frequent. In contrast, spiteful black lies (decreasing the earnings of both) and altruistic white lies (increasing the earnings of workers but decreasing those of the supervisor) are almost non-existent. The supervisors' second-order beliefs and their decision to lie are highly correlated, suggesting that guilt aversion plays a role.