This seems to go against everything a LLM is, how come ingesting more data will make it worse? It wouldn’t have got to where it is. There might be a limit, but rarely more data will affect it negatively.
You seem to assume that it would be particularly confused about its own content