Large language models don’t behave like people, even though we may expect them to