Three experiments were conducted to explore the effects of different interference treatments upon outcome-specific transfer from predictive learning to instrumental responding. A computer game was designed in which participants had to defend Andalusia from navy and air-force attacks. Participants learned the relationship between two instrumental responses (two keys in a standard keyboard) and two different outcomes (destruction of the ships or destruction of the planes). Then, they learned to predict which of two different cues predicted either outcome. Finally, participants were allowed to give either of the two instrumental responses in the presence of each cue. Outcome-specific transfer was shown as a preference for the response that shared the outcome with the current cue. Extinction of the cue before the transfer test had no effect upon transfer, regardless of the level of extinction (Experiments 1-3). However, pairing the cue with the alternative outcome (counterconditioning) reversed the outcome-based transfer effect (Experiment 3). The implications of these results for the contents of extinction in human predictive learning are discussed. (Contains 5 figures and 1 table.)