Learning biases in Pavlovian aversive conditioning have been found in response to specific categories of threat-relevant stimuli, such as snakes or angry faces. This has been suggested to reflect a selective predisposition to preferentially learn to associate stimuli that provided threats to survival across evolution with aversive outcomes. Here, we contrast with this perspective by highlighting that both threatening (angry faces) and rewarding (happy faces) social stimuli can produce learning biases during Pavlovian aversive conditioning. Using a differential aversive conditioning paradigm, the present study (N = 107) showed that the conditioned response to angry and happy faces was more readily acquired and more resistant to extinction than the conditioned response to neutral faces. Strikingly, whereas the effects for angry faces were of moderate size, the conditioned response persistence to happy faces was of relatively small size and influenced by inter-individual differences in their affective evaluation, as indexed by a Go/No-go Association Task. Computational reinforcement learning analyses further suggested that angry faces were associated with a lower inhibitory learning rate than happy faces, thereby inducing a greater decrease in the impact of negative prediction errors signals that contributed to weakening extinction learning. Altogether, these findings provide further evidence that the occurrence of learning biases in Pavlovian aversive conditioning is not specific to threat-related stimuli and depends on the stimulus' affective relevance to the organism.