Synthetic data can be used in training models, but it must be created from the same distribution to maintain quality.