Between Knowledge and Care Evaluating Generative AI-Based IUI in Type 2 Diabetes Management

Jul 16, 2026

Upload by Logic

Generative AI systems increasingly deliver everyday health guidance, yet their suitability for chronic care remains uncertain. Centering on Type 2 Diabetes Mellitus (T2DM), this mixed-methods study examines how patients interpret AI-generated information and how physicians evaluate it in China. Using formative patient grounding and a dimension-based physician rubric, it assesses AI responses across Accuracy, Safety, Clarity, Integrity, and Action Orientation. Results show strong performance in factual explanation and general lifestyle advice, but frequent failures in safety signaling, contextual judgment, and responsibility boundaries, especially when fluent outputs trigger overtrust. The work reframes quality dimensions as an interpretive lens to inform safe long-term care IUI design.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Introduction

Related Work

Between Knowledge and Care: Evaluating Generative AI-Based IUI in Type 2 Diabetes Management Through Patient and Physician Perspectives Yibo Meng∗ Tsinghua University Beijing, China [mengyb22@mails.tsinghua.edu.cn](mengyb22@mails.tsinghua.edu.cn) Ruiqi Chen∗ University of Washington Seattle, Washington, United States [ruiqich@uw.edu](ruiqich@uw.edu) Bingyi Liu University of Michigan, Ann Arbor Ann Arbor, Michigan, United State [bingyi@umich.edu](bingyi@umich.edu) arXiv :2510 . 10048v2 [ cs .HC] 2 Jan 2026 Yan Guan Tsinghua University Beijing, China [guany@tsinghua.edu.cn](guany@tsinghua.edu.cn) Abstract Generative AI systems are increasingly used by patients seeking everyday health guidance, yet their appropriateness in chronic care contexts remains unclear. Focusing on Type 2 Diabetes Mellitus (T2DM), this paper presents a mixed-methods investigation into how AI-generated health information is interpreted by patients and evaluated by physicians in China. Drawing on formative patient grounding and a dimension-based physician evaluation, we examine AI responses along five quality dimensions: Accuracy, Safety, Clarity, Integrity, and Action Orientation. Our findings reveal that while current systems perform well in factual explanation and general lifestyle guidance, they frequently break down in safety signaling, contextual judgment, and responsibility boundaries—particularly when fluent responses invite overtrust. By treating quality dimensions as an interpretive lens rather than a fixed framework, this work highlights the need for intelligent user interfaces that actively mediate AI outputs in chronic disease management, supporting calibrated trust and responsible boundary-setting in long-term care. Keywords Generative AI, Health IUI, Chronic Care Support, T2DM Management, Human-AI Collaboration 1 INTRODUCTION Generative AI systems are rapidly entering consumer-facing health contexts, raising renewed interest in how large language models (LLMs) might support chronic disease self-management. While recent work demonstrates that LLMs can synthesize biomedical knowledge and produce fluent explanations [22, 43], fluency and coverage alone do not guarantee that health information is safe, actionable, or interactionally appropriate for real users [3] . These concerns are particularly salient in chronic care, where patients repeatedly seek guidance under uncertainty, and where small omissions or poorly framed advice can accumulate into meaningful risk over time [11, 32, 33] . ∗ Both authors contributed equally to this research. Xiaolan Ding North China University of Science and Technology Health Science Center Tangshan, China In this work, we focus on Type 2 Diabetes Mellitus (T2DM) as a representative chronic condition to examine the role boundaries and reliability of generative AI in everyday self-management. T2DM requires sustained decisions around medication, diet, physical activity, and psychosocial regulation [20] . In China, where clinical workloads are high and specialist access is uneven, patients increasingly turn to always-available digital sources—including generative AI tools—to interpret symptoms, translate medical terminology, and explore self-care options outside the clinic [21, 22, 28] . This usage creates a tension central to HCI and IUI: AI can lower barriers to health information access, yet it can also invite overreliance when its outputs appear confident but fail to reflect contextual constraints and responsibility boundaries. A core limitation in prior evaluations is that health information quality is often operationalized as a largely one-dimensional property—typically correctness on benchmark-style question answering—despite mounting evidence that patients and clinicians attend to additional dimensions such as safety signaling, coherence, and the practical framing of next steps [3, 22] . More broadly, HCI research has shown that interface fluency can inflate perceive

FAQs

What does the paper evaluate about generative AI systems for T2DM care?

The paper evaluates how patients interpret AI-generated health information and how physicians judge those responses using five quality dimensions: Accuracy, Safety, Clarity, Integrity, and Action Orientation.

What problems do the authors find in current AI systems for chronic care?

The findings indicate breakdowns in safety signaling, contextual judgment, and responsibility boundaries, with overtrust becoming more likely when responses sound fluent and confident.

How does the paper’s dimension-based evaluation support better design of health IUIs?

Treating quality dimensions as an interpretive lens, the study supports designing intelligent interfaces that actively mediate model outputs, calibrating user trust and aligning actionability with clinically grounded risk boundaries.

Jul 16, 2026

Upload by Logic