Learning from a black box

We introduce a learning model in which the decision maker does not know how recommendations are generated, called the contraction rule. We present behavioral postulates that characterize it. The contraction rule can be uniquely identified and reveals how the decision maker interprets and how much she trusts the recommendation. In a dynamic stationary setting, we show that the contraction rule is not dominated by completely following recommendations and is incompatible with a property called compliance with balanced recommendations. Following this negative result, we demonstrate that the contraction rule may generate and reinforce recency bias and disagreement.