That's because LLMs don't actually think, they pattern-match. Since all the existing estimations out there are made assuming that a human is going to perform the task, the estimation that the LLM provides has the same inherent assumption. The LLM doesn't have a corpus of LLM-led estimations so it cannot take that into account.