Aligning with Whom? Large Language Models Have Gender and Racial Biases in Subjective NLP Tasks