Abstract: Modern AI excels at learning statistical associations from large-scale observational data, but many important scientific and decision-making problems require reasoning about interventions and alternative conditions. This talk presents a statistical perspective on that gap through counterfactual learning. I will first discuss how modern causal inference, including orthogonal and semiparametric methods, can provide robust causal signals, and then show why learning the geometry of complex data is a natural next step. Two complementary examples will be considered: geometry-adaptive conditional treatment effect learning using graph structure and Ollivier–Ricci curvature, and diffusion-guided learning of high-dimensional counterfactual distributions with valid statistical inference. Together, these ideas point toward a broader framework that combines causal targets, adaptive geometry, and representation learning, with the longer-term goal of building AI systems that remain reliable under interventions, distribution shifts, and counterfactual scenarios.