Figure 3: Hypothetical and empirical results for Experiment 2.

(a) Hypothetical outcomes for the subset conditions, comparing participants’ performance during random subsampling of one object or during ensemble coding (an extreme version of ensemble coding in which 100% of the objects in the set are integrated). Left: If the participant randomly samples lifelikeness information from just one item in the set, the magnitude of the correlation should remain relatively constant, even when more information becomes available because the participant does not use the new information. Right: In contrast, if the participant integrates lifelikeness information from every item as it becomes available to them, the correlation between participants’ lifelikeness ratings and the predicted lifelikeness ratings of the crowd should increase as more information (more items) becomes available. (b) Experiment 2 Results. The Fisher z scores increase as the number of items displayed increases. This pattern indicates that participants integrated the available information and did not use a random subsampling technique to accomplish the task. Error bars represent s.e.m.