Two experiments explored retroactive interference in human predictive learning. The name of a food was paired first with a gastric illness (A+), and then paired with a different gastric illness that was incompatible with the first one (A*). Experiment 1 presented three additional cues. C was followed by no outcome (C-). B was followed by * during the first phase, and then it was not presented during the second phase. Finally, D was presented only during the second phase, and it was followed by +. Under these conditions, retroactive interference was found as participants judging that A was followed by the second outcome, rather than by the first one. However, this treatment was generalized to B. This generalization was eliminated in the second experiment when the number of cues was increased, so that participants had the opportunity to learn that some cues may have not changed their meaning across phases. These results suggest that to find a clear effect of retroactive interference is needed to give participants the opportunity to learn that the meaning of different cues is independent of one another.