Linguistic Blind Spots of Large Language Models