We investigate whether experimental participants follow their private information and contradict herds in situations where it is empirically optimal to do so. We consider two sequences of players, an observed and an unobserved sequence. Observed players sequentially predict which of two options has been randomly chosen with the help of a medium quality private signal. Unobserved players predict which of the two options has been randomly chosen knowing previous choices of observed and with the help of a low, medium or high quality signal. We use preprogrammed computers as observed players in half the experimental sessions. Our new evidence suggests that participants are prone to a 'social-confirmation' bias and it gives support to the argument that they naively believe that each observable choice reveals a substantial amount of that person's private information. Though both the 'overweighting-of-private-information' and the 'social-con firmation' bias coexist in our data, participants forgo much larger parts of earnings when herding naively than when relying too much on their private information. Unobserved participants make the empirically optimal choice in 77 and 84 percent of the cases in the human-human and computer-human treatment which suggests that social learning improves in the presence of lower behavioral uncertainty.