We report a controlled laboratory experiment examining risk-taking and information aggregation in groups facing a common risk. The experiment allows us to examine how subjects respond to new information, in the form of both privately observed signals and signals reported from others. We find that a considerable number of subjects exhibit ‘reverse confirmation bias’: they place less weight on information from others that agrees with their private signal and more weight on conflicting information. We also find a striking degree of consensus when subjects make decisions on behalf of the group under a random dictatorship procedure. Reverse confirmation bias and the incidence of consensus are considerably reduced when group members can share signals but not communicate.