Answering questions honestly instead of predicting human answers: lots of problems and some solutions — LessWrong