Learning typically shows a negativity bias. Yet, in our re-analysis of meta-analytic evaluative conditioning (EC) data, we found no evidence of such bias. We provide a cognitive-ecological explanation for this lack of negativity bias. Ecologically, negative stimuli are rare, while positive stimuli are frequent. However, EC experiments typically present a balanced distribution of unconditioned stimuli (USs). In a first set of pre-registered studies (N = 785), we manipulated the frequency distribution of USs. Our results revealed symmetrical valence asymmetries: both rare positive and rare negative USs led to stronger EC effects. In a second set of pre-registered studies (N = 1200), we examined how participants evaluated groups when either one member was paired with a negative/positive US or four members were paired with positive/negative USs. If participants rated the group, pairing one member was as effective as pairing four members; yet, the rare pairing did not lead to a valence asymmetry. We discuss possible explanations for these results and suggest several paths for future research.