Crowd simulations are used extensively to study the dynamics of human collectives. Such studies are underpinned by specific movement models, which encode rules and assumptions about how people navigate a space and handle interactions with others. These models often give rise to macroscopic simulated crowd behaviours that are statistically valid, but which lack the noisy microscopic behaviours that are the signature of believable real crowds. In this article, we use an existing Turing test for crowds to identify realistic features of real crowds that are generally omitted from simulation models. Our previous study using this test established that untrained individuals have difficulty in classifying movies of crowds as real or simulated, and that such people often have an idealised view of how crowds move. In this follow-up study (with new participants) we perform a second trial, which now includes a training phase (showing participants movies of real crowds). We find that classification performance significantly improves after training, confirming the existence of features that allow participants to identify real crowds. High-performing individuals are able to identify the features of real crowds that should be incorporated into future simulations if they are to be considered realistic.