How AI Chatbots Choose Medical Solutions: A Discrete Choice Experiment (Preprint)

Read full text Summarise with VIRAT
Authors: Yuyang Yi, Yan He, Lai YIM, Richard Mohnish SHAH, Anyu Gao, Hei Hang Edmund Yiu, Wai‐Kit Ming
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

BACKGROUND The proliferation of artificial intelligence chatbots within the contemporary healthcare ecosystems represents a paradigmatic shift toward enhanced accessibility and operational efficiency. Nevertheless, extant human-computer interaction demonstrates substantial deficiencies in addressing users' core needs and value propositions. A critical lacuna emerges in understanding the heterogenous prioritization mechanisms employed by disparate AI chatbots attributes when synthesizing recommendations – a phenomenon conceptualized herein "AI response preference." OBJECTIVE This study aims to conduct a comprehensive quantitative assessment of response predispositions exhibited by six mainstream AI chatbots within medical communication contexts, while systematically elucidating their intrinsic preference architectures and underlying algorithmic internal weights. METHODS The Discrete Choice Experiment was systematically constructed utilizing seven key communication attributes as foundational parameters. Each experimental model underwent iterative implementation across 150 replications to ensure statistical robustness. The data employed mixed logit modeling frameworks to derive correlation coefficient (β), relative importance (RI), odds ratios (OR), and utility values, thereby facilitating comprehensive preference structure elucidation. RESULTS The AI chatbots demonstrated pronounced positive preference orientations, exhibiting particularly robust predilections toward dimensions encompassing elevated empathetic and psychosocial support provision, diagnostic precision and information veracity, and the clarity and feasibility of treatment suggestions. Conversely, preference coefficients for response latency and privacy cognizance manifested comparatively attenuated magnitudes. Significant inter-model heterogeneity in attribute-specific relative importance weightings was empirically observed across the different models: ChatGPT emphasized on monetary valuation parameters, Gemini prioritized primacy toward diagnostic accuracy optimization, Grok and DeepSeek manifested concentrated focus on communicative excellence and affective resonance, Doubao emphasized the clarity of treatment suggestions, whereas Wenxinyiyan displayed bifurcated emphasis encompassing both monetary considerations and treatment precision metrics. Subsequent willingness-to-pay analytical decomposition corroborated these differentiated value hierarchies and preference stratifications. CONCLUSIONS Current AI chatbots exhibit significant systematic biases in their response generation within the medical field, overemphasizing disproportionate valorization of technical proficiency and actionable attributes while concurrently neglecting humanistic dimensions particularly privacy preservation and communicative fluency. The conceptual framework of algorithmic response bias emerges as a pivotal evaluative metric for assessing human-machine value congruence, thereby catalyzing the advancement of medical AI systems characterized by diagnostic precision, empathetic resonance, and epistemic trustworthiness.

Fields of Study & Subjects

Artificial Intelligence in Healthcare and Education AI in Service Interactions Health Technology Consumer Behavior

Publication details

Publication date
25 March 2026
Language
EN
Document type
Preprint