Children have repeatedly been shown to selectively prefer to learn from previously accurate informants rather than those who have been inaccurate in the past. However, the stability of individual differences in performance on such tasks has yet to be studied. We presented preschoolers with two identical selective learning tasks conducted one week apart (Study 1; N = 53) or two parallel selective learning tasks conducted one immediately after the other (Study 2; N = 54). Results on both studies indicate that preschoolers tend to rely on the previously accurate informant more often than chance, replicating prior findings, but also showed weak correlations between performance across trials and administrations. In Study 1, performance on the first administration weakly, but significantly predicted performance on the second, identical, administration; in Study 2, performance on the first task only predicted performance on the second (non-identical) task when informant order was kept constant. Furthermore, in both studies, changes to the order of presentation of informants between trials within a task disrupted performance. Neither age, sex, theory of mind nor vocabulary predicted individual performance. These results confirm that there is some stability in individual tendencies to selectively learn from an accurate informant, but this stability can be fragile and easily disrupted by superficial task characteristics, raising several questions regarding the suitability of existing selective learning measures for the testing of individual differences.