Recent work suggests that AlphaFold2 (AF2)-a deep learning-based model that can accurately infer protein structure from sequence-may discern important features of folded protein energy landscapes, defined by the diversity and frequency of different conformations in the folded state. Here, we test the limits of its predictive power on fold-switching proteins, which assume two structures with regions of distinct secondary and/or tertiary structure. Using several implementations of AF2, including two published enhanced sampling approaches, we generated >280,000 models of 93 fold-switching proteins whose experimentally determined conformations were likely in AF2's training set. Combining all models, AF2 predicted fold switching with a modest success rate of ~25%, indicating that it does not readily sample both experimentally characterized conformations of most fold switchers. Further, AF2's confidence metrics selected against models consistent with experimentally determined fold-switching conformations in favor of inconsistent models. Accordingly, these confidence metrics-though suggested to evaluate protein energetics reliably-did not discriminate between low and high energy states of fold-switching proteins. We then evaluated AF2's performance on seven fold-switching proteins outside of its training set, generating >159,000 models in total. Fold switching was accurately predicted in one of seven targets with moderate confidence. Further, AF2 demonstrated no ability to predict alternative conformations of two newly discovered targets without homologs in the set of 93 fold switchers. These results indicate that AF2 has more to learn about the underlying energetics of protein ensembles and highlight the need for further developments of methods that readily predict multiple protein conformations.