Given the amount of training data out there, LLMs should have been perfect by now.