“Training a model to say "I don't know" based on human labelers who don't know the answer can inadvertently train the model to withhold information that it actually possesses.”