Individuals, Crowds, and the Network Dynamics of Belief Accuracy (2407.00199v3)

Abstract: Does talking to others make people more accurate or less accurate on numeric estimates such as quantitative evaluations or probabilistic forecasts? Research on peer-to-peer communication suggests that discussion between people will usually improve belief accuracy, while research on social networks suggests that error can percolate through groups and reduce accuracy. One challenge to interpreting empirical literature is that some studies measure accuracy at the group level, while others measure individual accuracy. We explain how social influence impacts belief accuracy by analyzing a formal model of opinion formation to identify the relationship between individual accuracy, group accuracy, and the network dynamics of belief formation. When opinions become more similar over time, change in individual error is always strictly better than change in group error, by a value equal to the change in variance. We show that change in group error can be decomposed into the influence network centralization, the accuracy/influence correlation ("calibration"), and the averageness/influence correlation ("herding"). Because group dynamics both theoretically and empirically lead people to become more similar over time, one might intuitively expect that the same factors which reduce group accuracy will also reduce individual accuracy. Instead, we find that individuals reliably improve under nearly all conditions, even when groups get worse. We support this analysis with data from six previously published experiments.