Rethinking AI in clinical decision support: a framework for reciprocal human-AI interaction.

Clinical AI systems increasingly match or exceed clinicians on some diagnostic benchmarks. Yet this reveals little about how AI output functions within clinical reasoning or how repeated reliance affects clinicians' independent capability. This paper proposes Bounded Reciprocal Adaptation for Clinician Engagement (BRACE), a framework for AI-enabled clinical decision support centred on the clinician-AI interaction rather than the model output. BRACE links three levels of support. Within a case, it makes uncertainty visible, preserves reasoning state and selectively prompts verification. Across encounters, it supports the development and maintenance of independent reasoning through bounded adaptation and formative review. Institutionally, it separates developmental interaction from evaluative use and routes clinician-initiated reports of AI error or contextual mismatch through governed review. Its governing principle is to provide sufficient support for safe care while preserving the work required for continued independent development. Adaptation is bounded in what the system may infer or modify and how interactional data may be used. These bounds are essential because a reasoning record has developmental value only if clinicians can reason freely, without withholding uncertainty, doubt or provisional judgements out of concern that they may later be used evaluatively. These provisions address design inversion, in which support intended to reduce cognitive vulnerability instead amplifies it: acutely through overreliance and cumulatively through loss or distortion of independent capability. Repeated cognitive and metacognitive offloading to AI may disrupt illness-script formation, maintenance or accuracy, resulting respectively in script-level forms of never-skilling, deskilling or mis-skilling. When AI displaces the discrimination needed to represent an unfamiliar presentation, the corresponding illness script may not form; this risk may be concentrated in ambiguous cases that also attract greater support. The central hypothesis is that the amount and type of cognitive and metacognitive work preserved during a BRACE-mediated encounter should predict independent capability subsequently demonstrated without AI. Evaluation pairs Tier 1 interactional measures derived from ordinary use with Tier 2 external anchors of later AI-off performance. Within-clinician longitudinal analyses test whether preserved cognitive and metacognitive work predicts those outcomes, while inference-drift audits test whether adaptation remains within its declared bounds.