"Words are not enough": Examining Emotional Support by Conversational AI for Caregivers

Paper Title

'Words are not enough': Examining Emotional Support by Conversational AI for Caregivers

Publication Info

• Topic area: Emotional support for caregivers using conversational AI.
• Keywords: Conversational AI, caregivers, emotional regulation, empathic AI, mental health, trust, accessibility, cultural sensitivity, co-design, therapeutic alliance.

Background and Problem

• Problem / challenge: Caregivers face emotional difficulties, including stress, depression, and isolation, but existing conversational AI systems lack emotional nuance and fail to address caregivers' complex lived realities.
• Significance: Addressing caregivers' emotional needs is critical to improving their well-being and sustaining their caregiving roles, which are central to healthcare systems.
• Motivation and related work: Prior work on conversational AI, such as ERICA and SERMO, has shown potential for emotional support but faces limitations like shallow emotional nuance, limited context sensitivity, and ethical risks. Current systems are not designed specifically for caregivers, leaving a gap in addressing their unique emotional realities.

Solution

• Proposed approach: Development of empathic AI—text-based conversational systems designed to support caregivers' emotion regulation and relational needs through attentive presence, trust-building, equitable accessibility, and worldview negotiation.
• Novelty:
  1. Empirical insights into caregivers' values, expectations, and concerns regarding conversational AI for emotion regulation.
  2. Design implications across five empathic dimensions: active listening, non-verbal communication, trust, accessibility, and worldview sensitivity.
  3. Introduction of "words are not enough" as a metaphor for designing empathic AI, emphasizing non-verbal cues and relational depth.
• Procedure and key techniques:
  • Two-phase qualitative study: co-design focus groups with caregivers and mental health professionals, followed by semi-structured interviews.
  • Use of ChatGPT-3.5 for interactive prompting sessions to explore AI's emotional support capabilities.
  • Thematic analysis of transcripts to identify challenges and design opportunities.

Results

• Concrete findings:
  • Caregivers valued AI's accessibility and anonymity but noted gaps in active listening, adaptive tone, and cultural sensitivity.
  • Five themes emerged: active listening dynamics, communication beyond words, trust and self-disclosure, access and ethics, and worldview negotiation.
  • Participants highlighted the need for reflective feedback, symbolic silences, adaptive tone modulation, and cultural responsiveness.
• Advantage over baselines: Empathic AI design addresses limitations in current systems by incorporating relational depth, dynamic trust calibration, and cultural sensitivity, tailored specifically to caregivers' realities.
• Experiments / evaluation:
  • Focus groups with 17 participants (caregivers and mental health professionals) across dementia, neurodiversity, and grief contexts.
  • Semi-structured interviews with 16 participants to deepen insights.
  • Thematic analysis using NVivo 14 to identify design implications.
• Limitations and future work:
  • Small sample size limits generalizability; broader representation needed.
  • Short-term interactions may not capture long-term trust and engagement dynamics.
  • Focus on text-based AI; future work should explore multimodal systems like voice and avatars.
  • Ethical guardrails for disclosure and autonomy need further exploration.

Summary

This study investigates how conversational AI can support caregivers' emotional regulation through empathic design. Findings reveal that while caregivers appreciate AI's accessibility and anonymity, current systems lack relational depth, cultural sensitivity, and adaptive communication. The paper proposes design dimensions for empathic AI, including active listening, trust calibration, equitable accessibility, and worldview negotiation, tailored to caregivers' unique needs. Future work should expand sample diversity, explore multimodal systems, and refine ethical safeguards to ensure safe and meaningful AI-mediated emotional support.