Synthetic bootstrapping

How does training algorithms on synthetic data affect their interpretability?