Teaching ML to answer questions honestly instead of predicting human answers — LessWrong