Abstract

We build models to make better decisions. We build explanations so we can trust those decisions. But what if the explanation is confidently pointing at the wrong thing? In this lecture I want to show that interpretability is not just a tool for transparency. It is a lens for catching when your model has learned something it should not have, and a bridge to causal reasoning that actually holds up in the real world. I’ll start with a provocation: concept-based models are one of our most powerful tools for interpretability, but they carry a hidden assumption that is almost always violated in healthcare data. The representations underlying those concepts are riddled with shortcuts, spurious correlations that look predictive during training and collapse at deployment. I’ll show why regularization, our default fix, quietly makes things worse, and how intervening directly in gradient space can rescue the concepts we thought we had. Then I’ll show what becomes possible when concepts are clean: I’ll share how they can be used to evaluate clinical policies we never actually tested, letting us ask counterfactual questions about patient care using structure a clinician would recognize and trust. If interpretability has ever felt like a box-ticking exercise to you, I hope this lecture changes that.