The Adaptive Opportunity Floor: Distributionally Robust Safeguards for Inclusive AI Tutoring

Abstract

Purpose: Generative artificial intelligence (AI) tutors can personalize learning, but improved average performance cannot establish equitable benefit for learners facing accessibility, language, connectivity, or time constraints. Methodology: This theoretical design-science study develops the Adaptive Opportunity Floor (AOF), combining causal potential-outcome comparisons, subgroup distributional robustness, accessible instructional alternatives, and educator-controlled fallback. We derive a conditional lower bound on subgroup benefit and distinguish direct empirical Wasserstein uncertainty from a separately calibrated population-shift model. Results: A formal bound and a hypothetical worked example demonstrate how average-gain optimization can reject a subgroup-safe policy. A finite-sample extension makes sampling uncertainty, candidate-policy selection, and probabilistic assumptions explicit. A comparative synthesis distinguishes AOF from predictive parity, Universal Design for Learning, conventional causal policy learning, and generic robust optimization. Implications: Institutions can predeclare accessible comparators, absolute attainment minima, subgroup margins, and evidence thresholds before deploying AI tutoring. The proposed safeguards are sufficient only under identified causal effects, defensible error and shift bounds, and valid outcome measurement. No learner data were collected and empirical effectiveness has not been demonstrated.