LLM inference does not just regurgitate the training corpus; RLHF is almost certainly to blame for this. There’s probably some Google n-gram graph to prove it.