Questioning the Survey Responses of Large Language Models