Action-outcome contingencies can be learnt either by active trial-and-error, or vicariously, by observing the outcomes of actions performed by others. The extant literature is ambiguous as to which of these modes of learning is more effective, as controlled comparisons of operant and observational learning are rare. Here, we contrasted human operant and observational value learning, assessing implicit and explicit measures of learning from positive and negative reinforcement. Compared to direct operant learning, we show observational learning is associated with an optimistic over-valuation of low-value options, a pattern apparent both in participants' choice preferences and their explicit post-hoc estimates of value. Learning of higher value options showed no such bias. We suggest that such a bias can be explained as a tendency for optimistic underestimation of the chance of experiencing negative events, an optimism repressed when information is gathered through direct operant learning. (Contains 3 figures.)