We study optimal income taxation in a framework where one's willingness to report his income truthfully is positively correlated with his type. We show that allowing low-productivity types to cheat leads to Pareto-superior outcomes as compared to deterring them, even if audits can be performed costlessly. When there is no cheating, redistribution takes place on first- and second-best frontiers and can never make low-ability types more well-off than high-ability types. Letting low-ability types cheat allows first-best redistribution up to a limit at which low-ability types are better off than high-ability types.