From β-Blind to β-Aware AI for Preference-Sensitive Clinical Decisions: Achieving Non-Maleficence

Large language models (LLMs) increasingly support clinical decisions by synthesizing population-level evidence and estimating average biological treatment effects (α). However, they are not designed to represent the causal contribution of patient choice (β)-a limitation that becomes clinically decisive in preference-sensitive care, where biological differences between options are small. Using the Spine Patient Outcomes Research Trial (SPORT) as an orienting case-a uniquely designed trial in which randomization and patient choice coexisted-we show that preference-mediated effects were clinically meaningful under conditions of near biological equipoise. In such settings, α-centric decision-support systems systematically misrank options because they lack access to β. We formalize this limitation as "α-bias" and "β-blindness," and propose a regime-routed architecture-Detect → Elicit → Recommend → Learn-with explicit deferral rules, neutrality constraints, provenance, and auditable guardrails. When outcomes hinge on choice rather than biological superiority, elicitation is a precondition for recommendation, not an optional refinement. These principles extend beyond medicine to any domain in which outcomes depend on decision-contingent preferences rather than fixed parameters. When outcomes depend on choice rather than biology, improving choices is improving outcomes. A companion paper addresses how β-aware systems may permissibly improve concordance once these safety conditions are satisfied and thus AI decision support moves to βoptimization.