Machine-learning models trained to classify human actions using synthetic data can outperform models trained using real data in certain situations. This could help scientists identify when it’s better to use synthetic data for training, which could eliminate bias, privacy, security, and copyright issues that often impact real datasets.