I don’t know if small amounts of text could do that, but I could imagine if LLMs keeps get trained on data generated by itself and other LLMs (which is likely to become a major source of content on the internet in the future), the quality of output can decrease significantly.