Fig 1.
Subjects were randomly assigned to a peer network that judged the veracity of news by exchanging either (i) binary or (ii) probabilistic judgments. Fewer trials were needed in the binary condition with equivalent statistical power (“Materials and Methods”). Each group in each condition consisted of 20 unique individuals. All group-level observations are independent.
Fig 2.
Comparing conditions in terms of the benefits of social learning for individual and collective judgments of news veracity.
(A) The fraction of subjects signaling belief improvements (i.e., belief revisions in the correct direction), first to final round, split by partisanship and condition (averaged separately for each partisan group in each peer network). (B) The fraction of groups revising their collective judgments of news veracity in the correct direction, first to final round (measured at the question level in terms of the fraction of questions for which each group improved). (C) The probability of individuals and groups improving in their classification accuracy from first to final round (measured as the fraction of individuals and groups who were initially incorrect in their veracity classification but who became correct in their final veracity classification as a result of the communication process). (D) The gain in classification accuracy as a result of communication in the probabilistic as opposed to the binary condition, measured as the probability of an increase in classification accuracy in the probabilistic condition minus the probability of an increase in classification accuracy in the binary condition (positive values indicate that improvements were greater in the probabilistic condition). (A, N = 270; B, N = 45; C, N = 90; D, N = 90). Error bars indicate 95% confidence intervals. Prob., Probabilistic Condition.
Fig 3.
Comparing the binary and probabilistic condition in terms of the likelihood of groups increasing in the accuracy of their collective judgments, split by the topic area of news content.
The probability of groups increasing in accuracy is measured as the probability of providing the correct veracity judgment by the final round, conditional on the majority being initially incorrect at the first round. We calculate the fraction of groups that increased in accuracy for each question in each condition, and then we average this fraction by topic area. There were 9 questions in each condition with groups that were initially inaccurate, producing 18 question-level observations. The politics and vaccines topic each contained 3 questions; all other topic areas contained a single question.
Fig 4.
Partisan differences in trust toward online content, averaged across questions at the trial-level.
Trust is measured by the rate at which subjects in each condition evaluated questions as more likely to be true. Density distributions indicate the fraction of questions that each trial collectively evaluated as true according to the communication style of each condition (binary vs. probabilistic). The rate at which questions were evaluated as true was averaged separately for both partisan groups in each trial. The data display the fraction of questions that Democrats and Republicans in each trial evaluated as true for the binary condition at the first (Panel A) and final round (Panel C), and for the probabilistic condition at the first (Panel B) and the final round (Panel D). Since each group in each condition encountered two true and two false news items, the appropriate fraction of questions that each trial should evaluate as true is 50%. Panel A & C, N = 30; C & D, N = 60. *p<0.1; **p<0.01; ***p<0.001; ns., not significant.