Does Training on Synthetic Data Make Models Less Robust?