It sounds like you are trying to understand LLM behavior using a mental model that inaccurately personifies the stochastic parrot.

A more parsimonious explanation is that this term got more-or-less randomly boosted by the reinforcement learning loop because there was nothing in the training data to discourage its use.