This is a toy example of how it could happen, in an artificial setting where you train entirely on generated outputs many times in a row.

It does not say that it is happening in production LLMs. It is a theoretical concern right now.