They’re not instructed to, it’s an inescapable artifact of their training.

Their training selects on completions that humans prefer, not completions that are most accurate. Sometimes accurate completions are what humans prefer, but the accuracy is not the goal, so accuracy in the completion is a happy accident.