Using a model with n>2 agents with random abilities I show that: a winner takes all tournament increases agents incentives to take costly risks; the incentive to take risks decreases in ability at the higher end of the ability distribution; using a tournament reduces the information output gives about ability. These predictions hold, despite the presence of piece-rate incentives and thus a direct cost of taking risks. Based on experimental data generated for Blinder and Morgan (2005) I confirm these predictions within an experimental setting. Tournaments significantly and adversely affected performance of subjects because they took in-efficient amounts of costly risks. The controlled environment of the lab also allows me to high-light some of the behavioral channels that caused the difference in performance.