Laboratory-reared predator-naive three-spined sticklebacks from two sites, one with abundant predatory fish (the high-risk site) and the other essentially predator-free (the low-risk site), were given a passive avoidance conditioning task in which they received a simulated attack from a model avian predator whenever they entered a previously-favoured feeding patch. 15/16 fish learned to avoid the dangerous patch within 15 days, but those from the high-risk site did so significantly faster and received fewer attacks in the process. The two categories of fish did not differ either in active avoidance of the attack stimulus or in the rate at which they started to re-exploit the dangerous patch once negative reinforcement ceased. It is argued that fish from high- and low-risk sites differ in the negatively reinforcing properties of the same, standardised attack.