How can we make AI explainable and trustworthy for healthcare?

Imagine getting health advice or therapy from an Artificial Intelligence (AI). While Large Language Models (LLMs) —the powerful AI behind tools — are becoming incredibly smart, would you trust them with your health?

Publication date
10 nov 2025
Share this page

This question is at the core of the PhD thesis "From Aligned Models to Trusted Interfaces: Explainable Health Intervention and Transparent Health Information Seeking." that Xin Sun will defend on 20 November 2025. His dissertation was supervised by Dr. Abdallah El Ali (CWI). In his research, Xin Sun explores how we can make these AI tools not just intelligent, but also explainable and trustworthy in sensitive health contexts.

Improving explainability and control of LLM's

Xin’s work bridges human–computer interaction, natural language processing, and psychology to address two core themes. The first theme concerns improving explainability and control in situations where AI acts as a digital therapist. Current LLMs often generate responses as a black box, without showing their reasoning. To solve this problem, Xin introduces the Script-Strategy Aligned Generation (SSAG) approach. This method forces the LLM to first predict on a therapeutic strategy (like "reflection" or "advice") before it generates the actual dialogue. This way he actions of LLMs are more controllable and explainable to both researchers and users, which is vital for building confidence in digital psychotherapy tools.

Fostering trust in LLM-powered health information

The second theme of Xin’s research focuses on fostering trust in LLM-powered health information through mixed-methods study and behavioral and physiological sensing. Surprisingly, the study shows that people trust LLM-powered conversational search more than traditional search engines, yet they perceive human-labeled information as more trustworthy than AI-labeled content.

By connecting technical model alignment with empirical assessments of human trust, Xin provides actionable insights for designing explainable and transparent LLM systems in digital healthcare. His work bridges how AI operates with how people experience and trust it, offering valuable guidance for developers and designers seeking to make LLMs both effective and trustworthy in real-world health applications.

More Information About the PhD Defense