The reinforcement paradox: Monetary incentives and Bayesian updating

We report the results of two pre-registered experiments designed to study the reinforcement paradox: increased incentives often fail to increase and sometimes even decrease performance in Bayesian updating tasks. We argue that, in the presence of win/loss cues, higher incentives have two countervailing effects: increased error rates for reinforcement behavior (win-stay, lose-shift) and increased performance for decisions resulting from Bayesian updating. We find some evidence that incentives increase performance when the win/loss cue is removed whereas when reinforcement is active the effects of incentives are mixed.