AbstractLeniency programs contribute to destabilizing collusion but can also be abused and generate perverse effects. This paper develops a simple model capturing this trade-off, which we use to relate the optimal leniency policy (the carrot) to the effectiveness of investigations (the stick). We show that it is always desirable to offer some leniency before an investigation is launched; it is also optimal to offer some leniency once an investigation is underway when investigations are unlikely to succeed in uncovering cartels, absent self-reporting. Our analysis also confirms the usefulness of restricting leniency to the first informant only; in contrast, it does not support prohibiting leniency for repeat offenders.