Reclaiming human clinical reasoning in generative AI-enabled healthcare: a perspective on cognitive agency, skill...

Generative artificial intelligence (GenAI) and large language models (LLMs) are transforming clinical decision support systems (CDSS) from traditional discrete predictions into fluent, context-aware clinical narrative generation. Although these capabilities substantially improve the efficiency of clinical workflows, they also introduce distinctive cognitive challenges: highly fluent and seemingly plausible medical narratives may reshape the way physicians construct clinical meaning and make analytic judgments. This perspective paper introduces the concept of Narrative Cognitive Substitution (NCS), proposed as a testable mid-level theoretical construct, which describes the hypothesized shift in human cognition from active hypothesis generation to passive retrospective validation when GenAI outputs pre-organize diagnostic reasoning and therapeutic pathways. We further propose a Skill Evolution Framework that characterizes the potential effects of GenAI on physician competence across five trajectories: Augmentation, Restructuring, Deskilling, Never-skilling, and Mis-skilling. In addition, we analyze the resulting Accountability Gap: although algorithmic outputs are increasingly embedded in clinical decision-making, decision influence is distributed across clinicians, institutions, developers, and deployers, yet professional accountability and final decision authority remain formally concentrated on the individual physician. To address these systemic challenges, we propose a human-centered governance framework -presented as design hypotheses requiring prospective evaluation- encompassing user interface design (embedding risk-adaptive cognitive friction), medical education (establishing targeted, protected AI-free reasoning exercises ("Reasoning Cleanrooms")), and institutional policies (implementing substantive human oversight and algorithmovigilance). We argue that the goal of digital health governance should not be to maximize automation, but rather to safeguard human cognitive agency, diagnostic vigilance, and layered, distributed accountability within AI-enabled healthcare systems.