Humans do grammatical errors (non-native, for whatever expressive effect, typos, dialects, slangs etc.) => datasets contain a percentage of grammatical errors =

Humans do grammatical errors (non-native, for whatever expressive effect, typos, dialects, slangs etc.) => datasets contain a percentage of grammatical errors => LLMs does it. I mean it can be about frequency, and carry infrequent errors because of it. I don't see any contradiction.