Selectivity towards certain kinds of information, such as danger and threat, has been observed in different areas of human cognition such as learning and attention. Error Management Theory (EMT) argues that this kind of selectivity, also referred to as bias, reflects the asymmetry in costs of making mistakes with this kind of information relative to other kinds of information (Haselton & Nettle, 2006). Based on the principles of EMT, we reasoned that such selectivity may also exist in the generalization of information where generalization mistakes are differentially costly. Specifically, we predicted that, for novel foods, edibility information would be generalized to a lesser degree while toxicity information would be generalized to a greater degree than neutral information. Experiment 1 tested this in adults, and found that, as predicted, participants overgeneralized toxicity information and undergeneralized edibility information, relative to neutral control information. However, Experiment 2 did not find this same selectivity in 4- to 5-year-old children. Experiments 3 and 4 further explored selective generalization in adults. The results suggest that (i) selective generalizations did not rely solely on categorization judgments (Experiment 3) and (ii) there was some indication of a broader negativity/positivity effect in generalization (Experiment 4), although this appears less pronounced than the selectivity we found in Experiment 1. Taken together these results show selectivity in generalization systems that reflect the relative costs of errors in the food domain.