Figure 1.

Simulations predict strong reliance on social information where rewards are highly variable, and low levels where rewards do not vary (a). When bees were trained and tested with bee models (social cues), we observed significantly more bees landing on a flower with a cue when they previously experienced high-variance distributions compared with the no-variance distributions (b). There was no difference between individuals trained and tested with the non-social cues. Dashed line indicates random choice (at 33.3%, as there were four out of 12 flowers with a cue). Error bars indicate standard errors (a) and 90% confidence intervals (b, adjusted Wald interval). ***, p < 0.001; n.s., non-significant.