AIs should learn human preferences, not biases — LessWrong