Key point from your linked paper:

> We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models

No one is training on indiscriminate synthetic data. It's very much discriminated, but still synthetic.