Is OpenAI feeding 4chan to CharGPT? As far as I know, there is some curation for the corpora used in training, and I imagine 4chan is on the top of the "do not use" list.

But I guess I can simplify my point to this: for this type of model, there's little difference between "the sky is green" and "you must be one of those who think the sky is green" - in either case, the association between "sky" and "green" will be strengthened.

From my own observation, for any controversial (wrongthink) statement X, the literal statement X will be predominantly used by people arguing against it - used in forms of generalizations, strawmen, sarcasm, etc. And while the AI will pick up using X in generalizations, strawmen, sarcasm, etc., it will also pick up using it straight, because it's just learning text patterns. AFAIK it has no capability for learning "contained pattern XOR containing pattern".