Informing Robot Wellbeing Coach Design through Longitudinal Analysis of Human-AI Dialogue

Social robots and conversational agents are being explored as supports for wellbeing, goal-setting, and everyday self-regulation. While prior work highlights their potential to motivate and guide users, much of the evidence relies on self-reported outcomes or short, researcher-mediated encounters. As a result, we know little about the interaction dynamics that unfold when people use such systems in real-world contexts, and how these dynamics should shape future robot wellbeing coaches. This paper addresses this gap through content analysis of 4352 messages exchanged longitudinally between 38 university students and an LLM-based wellbeing coach. Our results provide a fine-grained view into how users naturally shape, steer, and sometimes struggle within supportive human-AI dialogue, revealing patterns of user-led direction, guidance-seeking, and emotional expression. We discuss how these dynamics can inform the design of robot wellbeing coaches that support user autonomy, provide appropriate scaffolding, and uphold ethical boundaries in sustained wellbeing interactions.

💡 Research Summary

This paper investigates the interaction dynamics that emerge when university students engage with a large‑language‑model (LLM) based wellbeing coach over an extended period, and derives design implications for future robot coaches. Prior work on social robots and conversational agents for wellbeing has largely relied on self‑report questionnaires or short, researcher‑mediated sessions, leaving a gap in understanding how natural, multi‑session dialogues unfold in real‑world contexts. To fill this gap, the authors recruited 38 undergraduate students and gave them unrestricted access to a web‑based LLM coach for one week. The coach was powered by the OpenAI GPT API and operated under a fixed system prompt that instructed it to act as a supportive, non‑clinical coach: it asked open‑ended reflective questions, helped users refine and set SMART goals, and offered practical strategies inspired by Acceptance and Commitment Therapy (ACT). Conversation history was retained across sessions, allowing the coach to reference prior goals and follow‑up on earlier concerns.

The researchers performed a quantitative content analysis on the 4,352 user messages (average 116 dialogue turns per participant, counting both user and system turns). Guided by established concepts from psychology and human‑robot interaction (HRI), they defined six interaction dynamics: autonomy (expressing self‑determined choices or boundaries), agency (initiating actions or next steps), emotional self‑disclosure (revealing feelings or affective states), rumination (repetitive, self‑focused negative thinking), negotiation (modifying or reshaping system suggestions), and compliance (accepting and committing to suggestions without substantial modification). A codebook was created (see Table 1 in the paper), and each user turn could receive multiple codes. Two coders independently annotated the data, resolving disagreements through discussion.

Key findings:

• Autonomy appeared in 97 % of participants (37/38) and accounted for 10.5 % of all user messages (mean = 6.1 instances per participant who showed autonomy). Users frequently declared preferences or limits, positioning themselves as decision‑makers rather than passive recipients.
• Agency was also present in 97 % of participants, representing 20.8 % of messages (mean = 12.1 instances). Participants often announced independent actions (“I’ll start working now”) or managed task execution (e.g., scheduling goal check‑ins in a calendar).
• Emotional self‑disclosure occurred in 68 % of participants but comprised only 3.3 % of messages (mean = 2.8 instances). Users shared concise emotional states (“I feel unseen”) and sometimes interpreted the cause of those feelings.
• Rumination was rare, observed in 13 % of participants and less than 1 % of all messages (mean = 1.8 instances). When present, it manifested as looping self‑criticism about perceived failures.
• Compliance was found in 79 % of participants, making up 4.9 % of messages (mean = 3.5 instances). Users explicitly accepted the coach’s suggestions, ranging from brief confirmations to reflective endorsements of progress.
• Negotiation appeared in 79 % of participants, accounting for 5.8 % of messages (mean = 4.2 instances). Participants reshaped suggestions to fit personal constraints (“I aim for 180 min of intense workout a week, not 30”) or expressed discomfort with proposed deadlines.

Co‑occurrence analysis revealed that autonomy and agency overlapped in 95 % of participants, indicating that users who set or adjust goals also actively steered the conversation. Negotiation frequently co‑occurred with both autonomy (76 %) and agency (79 %), suggesting that users who guide the interaction also tend to question or modify system suggestions. Emotional disclosure and compliance similarly overlapped with autonomy and agency at high rates (≈ 70–78 %), showing that affective sharing and acceptance tend to happen within already engaged, self‑directed dialogues.

From these patterns, the authors draw several design implications for robot wellbeing coaches:

1. Support Autonomy Without Being Directive – Robots should provide scaffolding that respects user‑determined goals while offering optional adjustments. For example, after a user states a goal, the robot can acknowledge it, ask if any modifications are needed, and present alternatives without imposing a single path.
2. Facilitate Agency‑Driven Progress – Enable users to initiate actions and track their own progress. A robot could integrate with personal calendars or task managers, allowing users to log completed steps autonomously, reinforcing a sense of ownership.
3. Encourage Healthy Emotional Disclosure – Since emotional self‑disclosure is modest but beneficial, robots should invite affective sharing (e.g., “How are you feeling today?”) and then gently steer the conversation toward forward‑looking strategies, avoiding prolonged rumination loops. Techniques such as re‑framing or forward‑focused questioning can prevent negative spirals.
4. Balance Negotiation and Compliance – Design dialogue policies that accept user‑initiated modifications (negotiation) while still providing clear, actionable recommendations. This balance prevents the robot from becoming overly authoritarian or overly passive.
5. Maintain Ethical Boundaries – The system explicitly avoided diagnostic or crisis‑intervention functions, and participants were reminded of its non‑clinical nature. Future embodied robots should retain such transparency, offering referrals to professional services when needed and avoiding over‑reliance.

The study’s limitations include a homogeneous sample (all university students), a relatively short deployment (one week), and reliance on a single LLM and prompt design, which may affect generalizability. Nonetheless, the work offers one of the first large‑scale, longitudinal quantitative analyses of human‑AI wellbeing coaching, providing concrete metrics (frequency of interaction dynamics) that can inform the architecture, dialogue strategies, and ethical safeguards of future socially assistive robots. Future research directions suggested include expanding to diverse demographic groups, extending the observation period, and testing embodied robot implementations to examine how physical presence interacts with the identified dynamics.