Physicians' Experiences of Generative AI-Assisted Diagnostic Reasoning and Perceived Applicability in the Kenyan Healthcare Context: A Qualitative Study from a Tertiary Teaching Hospital

Background: Generative artificial intelligence (AI), particularly large language model (LLM)-based tools, offers new opportunities to support clinical reasoning and diagnostic decision-making. However, direct experiential evidence on how frontline physicians in low- and middle-income countries engage with these tools in practice remains largely absent from Literature.

Methods: Embedded in a wider multi-site parallel group randomized controlled trial study, we conducted a qualitative descriptive study guided by a constructivist-interpretivist paradigm anchored on the Normalisation Process Theory(May & Finch, 2009) framework. Semi-structured exit interviews were conducted with 19 physicians across diverse departments at a Kenyan tertiary teaching hospital immediately following structured diagnostic simulations with a generative AI chatbot (ChatGPT-4o API access). Data was analysed inductively through reflexive thematic analysis.

Findings: Ten sub-themes emerged within two overarching domains. Perceived benefits included broadened differential diagnoses beyond initial formulation, more structured history-taking and clinical reasoning, targeted investigation planning, enriched patient education, and consolidated workflow efficiency. Contextual barriers included workflow incompatibility with high-volume clinical settings, systematic misalignment between AI outputs and local epidemiology and resource availability, physician concern that visible AI use during consultation would undermine clinical authority and patient trust — a finding rooted in the relational architecture of the African clinical encounter — interface usability limitations, and risk of overreliance with reduced critical engagement.

Interpretation: Kenyan physicians showed strong readiness to engage with generative AI as a clinical thinking partner, but structural, epidemiological, and relational barriers prevented routine clinical enactment. These findings reframe AI implementation in African health systems from a technical deployment problem to a sociotechnical integration challenge requiring localised model training, workflow-sensitive interface design, and institutional renegotiation of AI's role in the clinical encounter.