Deep learning methods of predicting protein structures have reached an accuracy comparable to that of high-resolution experimental methods. It is thus possible to generate accurate models of the native states of hundreds of millions of proteins. An open question, however, concerns whether these advances can be translated to disordered proteins, which should be represented as structural ensembles because of their heterogeneous and dynamical nature. Here we show that the inter-residue distances predicted by AlphaFold for disordered proteins are accurate, and describe how they can be used to construct structural ensembles. These results illustrate the possibility of making structural predictions for disordered proteins using deep learning methods trained on the large structural databases available for folded proteins.