Think of it like making a xerox copy of a xerox copy. The copy of the copy is always shittier.
Using synthetic data can escalate model collapse, as a model is only as good as its training data, which is partly why these LLM models "hallucinate", having been trained on a wealth of garbage from Reddit.