65 ms, SD = 54.37 ms) compared to an easily discriminable pwin pair (80/20, mean = 959.67 ms, SD = 42.51 ms) (F[1, 15] = 125.81, p < 0.0001, η2 = 0.89). There was also a linear effect of test number with participants becoming quicker with time (F[1, 15] = 35.65, p < 0.0001, η2 = 0.70). There were no effects of session (mean actor = 1038.63 ms, SD actor = 51.01 ms; mean observer = 1066.69, SD observer = 49.67 ms), showing that any difference

found between observational and operant learning was not explicable by GS-7340 mouse RT differences. The results from Experiment 1 show that, while value learning through trial-and-error is highly accurate, observational learning is associated with erroneous learning of low-value options (i.e. those with the lowest probability of reward). In essence, observational learners show a striking over-estimation of the likelihood of winning from the lower-value options, a fallacy leading to impaired accuracy when choosing between two low-value options. This learning difference was apparent even though monetary incentives and visual information were matched in actor and observer learning. A different number of test trials were paid for observers relative to actors and this might have had a general

effect on performance. However, it cannot explain observers’ asymmetrically poor accuracy when learn more choosing between the 40/20 gamble pairs, and financial incentives were matched across each learning session overall. It is important to note that over-estimation of the value of the 20% win option did not cause observers to perform significantly worse when choosing between the 80/20 pairs. This is likely to reflect the fact that the probability difference is

uniquely high for such pairs, allowing for lower uncertainty when determining the higher value choice. It is interesting to observe that individual choice accuracies do not asymptote to 100%, as might be expected from rational decision makers once they accurately learn the value of stimuli. This may partially reflect the phenomenon of probability matching, a common finding in learning experiments (Herrnstein, 1961, Lau and Glimcher, 2005 and Sugrue et al., 2004), arising from a matching of choice frequency to average reinforcement rate. Note that, in our data, choice Prostatic acid phosphatase frequencies do not simply match learnt probabilities of reward, moreover probability matching does not in itself predict any difference between acting and observational learning. Two potential design weaknesses can be identified in Experiment 1. First, by yoking the sequence of actor choices to participants’ subsequent observer session, to match actor and observer learning for information presented, we are not able to counterbalance session order. Since participants also learnt about novel stimuli in the second session, learning may be worse solely because the task has switched.