Meta-Predictions Enhance Social Learning

People learn from others to improve their judgments across a wide range of tasks, from financial forecasting to medical diagnosis. To facilitate social learning, people have historically either relied on simple aggregation methods, such as the average or majority opinion of a group, or on noisy signals of others’ expertise (e.g., confidence levels). However, these methods can be unreliable and inaccurate. Here, we test the theory that social learning can be enhanced through a signaling mechanism that indicates which group members are better at predicting others’ predictions, which we term “meta-prediction accuracy”. Although recently built algorithms leverage meta-predictions to improve the aggregation of people’s estimates, these algorithms are highly complex and difficult to communicate to individuals, and thus not well-suited for social learning. We develop a more intuitive yet equally accurate algorithm that harnesses the power of meta-predictions to improve real-time social learning among interacting decision-makers. Across two pre-registered experiments, we show that providing individual decision-makers with group members’ meta-prediction accuracy consistently improves individuals’ estimation accuracy more than providing other popular signaling mechanisms. This holds regardless of the complexity of the problem and whether estimates are binary or probabilistic. Together, these findings demonstrate that signaling meta-prediction accuracy in social groups reliably enhances social learning.