JMIR Publications

ABSTRACT

Background:

AI-powered conversational agents offer accessible, culturally responsive, and personalized support for screening decisions. However, concerns about accuracy, privacy, and safety may undermine patient trust and adoption. AI risks and mitigation strategies are typically defined by researchers and clinicians, with limited patient input. Consequently, less is known about which risks patients and end-users prioritize and how mitigating them may foster trust and adoption.

Objective:

Using lung cancer screening (LCS) as an example, we explore the concerns, priorities, and needs of screening-eligible individuals to inform a AI Risk Mitigation Framework to guide the development of AI agents for supporting screening decisions.

Methods:

We conducted 60-minute semi-structured online focus groups with LCS-eligible community members from September to November 2025. Participants underwent at least one LCS LDCT scan. Discussions explored decision-making, provider support, familiarity with AI agents, trust, perceived risks, and preferences for AI-supported shared decision-making. Focus groups were recorded, transcribed verbatim, quality-checked, and supplemented with researcher notes. Reporting followed the COREQ checklist. The analysis was conducted through the lens of Self-Determination Theory. The study results, multidisciplinary team input, and current literature on risk and risk mitigation strategies informed the User-Informed AI Risk Mitigation Framework.

Results:

A total of 14 participants attended three focus groups. The reported mean age was 62.1 years; 57% were women, 57% Black/African American, 43% White, 36% had a bachelor’s degree, and 71% reported annual incomes below $50,000. In addition to the themes outlined by Self-Determination Theory, competence, autonomy, and relatedness, two additional themes were included: AI trust and preference for chatbot delivery. The resulting User-Informed AI Risk Mitigation Framework links users’ key concerns to corresponding risk mitigation strategies. The framework's novel contribution lies in prioritizing strategies and specific disclosures that LCS-eligible users perceive as necessary to foster trust and acceptance of AI conversational agents.

Conclusions:

The User-Informed AI Risk Mitigation Framework integrates patient perspectives, Self-Determination Theory, and prior evidence on AI risk mitigation. It provides a user-centered approach to designing AI agents that are likely to be trusted and accepted for screening decisions.