Sometimes You Need Facts, and Sometimes a Hug: Understanding Older Adults’ Preferences for Explanations in LLM-Based Conversational AI Systems

Paper Title

Sometimes You Need Facts, and Sometimes a Hug: Understanding Older Adults' Preferences for Explanations in LLM-Based Conversational AI Systems

Publication Info

• Topic area: Older adults' preferences for AI-generated explanations in assistive contexts.
• Keywords: Conversational AI, explainability, older adults, LLMs, aging in place, human-centered AI, smart home, emergency alerts, routine reminders, user preferences.

Background and Problem

• Problem / challenge: Existing explainable AI (XAI) approaches prioritize technical transparency for developers rather than end-users, leaving gaps in understanding older adults' preferences for AI-generated explanations in everyday contexts.
• Significance: As older adults increasingly adopt Conversational AI systems for assistive support, designing context-sensitive and user-centered explanations is critical to fostering trust, usability, and accountability.
• Motivation and related work: Prior research highlights the importance of explainability for trust but lacks empirical insights into older adults' preferences. Emerging capabilities of LLMs offer opportunities for personalized and conversational explanations, but challenges like opacity, hallucination, and lack of contextual grounding persist.

Solution

• Proposed approach: An exploratory Speed Dating study to investigate older adults' preferences for AI-generated explanations across two assistive contexts: routine reminders and emergency alerts.
• Novelty:
  1. Empirical insights into how older adults' preferences for explanations shift across assistive contexts.
  2. Examination of the influence of different information sources (e.g., conversational history, environmental data) on perceptions of trustworthiness and usefulness.
  3. Identification of explanations as part of interactive, multi-turn conversational exchanges that calibrate urgency, guide actionability, and enhance awareness of shared routines.
• Procedure and key techniques:
  • Conducted Speed Dating sessions with 23 older adults (11 dyads and 1 individual) using storyboard-based vignettes.
  • Explored four explanation types: conversational history, environmental data, activity-based inferences, and internal system logic.
  • Participants rated explanations and engaged in semi-structured interviews to discuss preferences and perceptions.

Results

• Concrete findings:
  • Routine reminders: Explanations referencing conversational history were rated most positively.
  • Emergency alerts: Explanations using environmental data were rated highest.
  • Explanations based on activity-based inferences were moderately rated, while internal system logic (confidence scores) received the lowest ratings.
• Advantage over baselines:
  • Context-sensitive explanations (e.g., personalized and evidence-based) were perceived as more trustworthy and actionable compared to generic or numerical confidence-based explanations.
• Experiments / evaluation:
  • Participants rated explanations on a 5-point Likert scale and provided qualitative feedback.
  • Two assistive contexts (routine vs. emergency) were explored using storyboards, with explanations differentiated by information source.
• Limitations and future work:
  • Low-fidelity storyboards limit ecological realism; future work should involve interactive prototypes to study real-time explanation breakdowns.
  • Focused on older adults with early-stage cognitive changes; future studies could examine evolving preferences as cognitive conditions progress.
  • Further exploration needed into other explanatory methods (e.g., contrastive, counterfactuals) and multi-user settings.

Summary

This study explores older adults' preferences for AI-generated explanations in Conversational AI systems across routine and emergency contexts. Findings reveal that preferences are highly context-dependent, with conversational history valued for routine tasks and environmental data preferred for emergencies. Explanations were seen as interactive and multi-functional, aiding urgency calibration, actionability, and shared routine awareness. The study highlights the need for context-sensitive, personalized, and emotionally attuned explanations, while raising important considerations for privacy and multi-user settings. Future work should focus on higher-fidelity prototypes, longitudinal studies, and adaptive explanation systems.