The Art of the AI Coach: Understanding Human-Robot Wellbeing

by

Author: Denis Avetisyan

New research reveals the crucial interaction dynamics that make AI-powered wellbeing coaches effective, focusing on the importance of user agency and collaborative dialogue.

Longitudinal analysis of human-AI conversations demonstrates how autonomy support, negotiation, and emotional disclosure shape successful wellbeing coaching interactions.

While the potential of social robots and conversational agents to support wellbeing is increasingly recognized, evidence often relies on limited, controlled interactions. This paper, ‘Informing Robot Wellbeing Coach Design through Longitudinal Analysis of Human-AI Dialogue’, addresses this gap by presenting a detailed content analysis of [latex]\mathcal{N}=4352[/latex] messages exchanged over time between university students and an LLM-based wellbeing coach. Our findings reveal key interaction dynamics-including user-led direction, guidance-seeking, and emotional expression-that shape effective supportive dialogue. How can these insights inform the design of robot wellbeing coaches that prioritize user autonomy, provide nuanced scaffolding, and navigate the ethical considerations of sustained wellbeing interactions?

The Weight of Wellbeing: Addressing a Systemic Challenge

Access to crucial mental wellbeing support remains a significant challenge for many, largely due to systemic barriers of cost and availability. Traditional interventions, such as regular therapy sessions or specialized counseling, often carry substantial financial burdens, effectively excluding individuals with limited incomes or inadequate insurance coverage. Beyond affordability, geographical limitations and a shortage of qualified mental health professionals create long wait times and restricted access, particularly in rural or underserved communities. This scarcity impacts preventative care, meaning individuals frequently seek help only during crises, exacerbating challenges and hindering long-term wellbeing. Consequently, a considerable portion of the population is left without the resources needed to address mental health concerns proactively, underscoring the urgent need for innovative and scalable solutions.

Many current digital wellbeing applications operate on a one-size-fits-all model, delivering generalized advice or standardized exercises that often fail to resonate with individual experiences. While these tools can offer broad access to mental health resources, they frequently lack the capacity to adapt to the user’s specific emotional state, personal history, or evolving needs. This inflexibility stems from the difficulty in accurately gauging subtle cues – such as changes in language, tone, or activity patterns – that a human therapist intuitively recognizes. Consequently, users may find the recommendations irrelevant, unhelpful, or even counterproductive, leading to disengagement and a return to less effective coping mechanisms. The challenge lies in moving beyond simple content delivery to create truly responsive systems capable of delivering support tailored to the unique complexities of the individual.

Human-Robot Interaction presents a compelling, though complex, pathway towards delivering personalized wellbeing support to a vastly larger population. While digital interventions currently struggle with replicating the empathy and adaptability of human interaction, carefully designed robots offer the potential to bridge this gap. Successful implementation, however, demands more than simply automating existing therapeutic techniques; it necessitates a focus on building trust, understanding non-verbal cues, and adapting responses based on individual emotional states. Research is concentrating on developing robots capable of subtle behavioral adjustments-tone of voice, gaze patterns, even physical proximity-to create a supportive and engaging experience. The challenge lies in balancing technological sophistication with the essential qualities of human connection, ensuring these robotic companions enhance, rather than replace, genuine care.

Truly impactful wellbeing support hinges on a robust understanding of what drives human behavior. Motivational psychology reveals that external interventions are most effective when aligned with intrinsic needs – autonomy, competence, and relatedness – rather than relying solely on prescriptive advice. Research demonstrates that fostering a sense of personal agency, enabling skill development, and cultivating meaningful connections significantly enhance engagement with wellbeing practices. Consequently, successful strategies move beyond simply telling individuals what to do; instead, they empower them to define their own goals, build confidence in their abilities, and feel supported in their journey towards improved mental health. This nuanced approach, rooted in psychological principles, is crucial for designing interventions that resonate with individuals and promote lasting positive change.

A Coach for the Self: Empowering Autonomy with Language

The LLM-Based Wellbeing Coach is a software application built upon large language models (LLMs) – specifically, models capable of natural language understanding and generation. These LLMs are utilized to deliver individualized guidance and support to users, adapting responses based on user input and tracked data. The system moves beyond pre-scripted responses by dynamically generating content relevant to the user’s specific needs and goals. The core functionality relies on the LLM’s ability to process user statements, identify underlying needs or challenges, and formulate appropriate responses intended to encourage positive behavioral change and improve wellbeing. Data privacy and security measures are implemented to protect user information processed by the LLM.

The Wellbeing Coach incorporates the SMART Framework – Specific, Measurable, Achievable, Relevant, and Time-bound – to facilitate effective goal setting and progress tracking. Users are prompted to define goals that meet these criteria, moving beyond vague aspirations to concrete objectives. The system then enables users to break down larger goals into smaller, manageable steps, and to establish quantifiable metrics for assessing achievement. Progress is monitored through user input, allowing the system to provide feedback and adjust recommendations as needed. This structured approach is intended to increase the likelihood of goal attainment and foster a sense of accomplishment, while simultaneously providing data for personalized coaching.

The LLM-Based Wellbeing Coach is intentionally designed to foster user autonomy by prioritizing collaborative problem-solving over prescriptive advice. The system avoids directly offering solutions; instead, it prompts users to explore their own values, identify potential obstacles, and formulate personalized action plans. This approach is implemented through open-ended questioning and the provision of information designed to empower self-direction, rather than through the issuance of commands or the imposition of predetermined pathways. The LLM functions as a supportive facilitator, guiding users through a process of self-discovery and informed decision-making, thereby reinforcing their agency and ownership over their wellbeing journey.

The LLM-Based Wellbeing Coach incorporates principles of Acceptance and Commitment Therapy (ACT) to foster psychological flexibility, a core component of mental wellbeing. This is achieved by prompting users to identify personal values and encouraging commitment to actions aligned with those values, rather than focusing solely on symptom reduction. The system facilitates acceptance of difficult thoughts and feelings as a natural part of the human experience, promoting mindful observation without judgment. Through techniques derived from ACT, the coach aims to increase users’ ability to engage fully in present-moment experiences, even in the presence of challenging emotions, thereby enhancing their overall psychological resilience and promoting behavioral change consistent with their stated values.

Beyond Directive: Negotiating Support Through Dialogue

The coaching system is designed around a mixed-initiative interaction style, meaning dialogue isn’t strictly dictated by pre-programmed prompts. Instead, the user maintains control of the conversation’s direction and pace, actively guiding the coaching process. This contrasts with purely system-led interactions where the user simply responds to prompts. The system responds to user input, offering suggestions and feedback, but the user retains the ability to introduce new topics, request specific types of guidance, or challenge the system’s recommendations. This approach allows for a more personalized and flexible coaching experience, tailored to the individual user’s needs and preferences at any given moment.

The coaching system is designed to facilitate adaptation of suggestions to individual user needs. Rather than presenting recommendations as directives, the system prompts users to consider how proposed strategies align with their unique circumstances and preferences. This is achieved through conversational cues that invite modification and refinement of the coach’s suggestions. Users are explicitly encouraged to rephrase, combine, or reject elements of the coach’s guidance, thereby ensuring the resulting plan is both personally relevant and actionable within their specific context. This negotiation capability is a core feature intended to maximize user buy-in and promote effective implementation of the coaching advice.

The system’s design explicitly supports user agency by enabling negotiation of coaching suggestions. This functionality allows users to adapt recommendations to their individual contexts and preferences, rather than passively accepting them. Quantitative Content Analysis of 4,352 messages revealed high levels of participant autonomy – observed in 97.4% (37 of 38) of participants – and a corresponding level of agency, indicating that users actively shaped the interaction and maintained a sense of control over the coaching process. This bidirectional exchange reinforces user ownership of their learning path and contributes to increased engagement with the system’s recommendations.

A quantitative analysis was performed on 4,352 messages exchanged between students and the coaching system. This analysis utilized Quantitative Content Analysis to assess the effectiveness of the mixed-initiative and negotiation approach. Results indicated high levels of participant autonomy, observed in 97.4% (37 of 38) of participants, and a corresponding level of agency within the same percentage. These metrics suggest a strong correlation between the system’s design and its ability to foster user control and self-direction during the coaching process.

Quantitative analysis of 4,352 messages from user interactions with the coaching system revealed high levels of both participant autonomy and agency. Specifically, 97.4% (37 of 38) of participants demonstrated autonomy – the ability to make independent choices – and a comparable 97.4% exhibited agency, indicating a strong sense of control over their learning process. These metrics were determined through Quantitative Content Analysis of the message corpus, focusing on indicators of self-direction and proactive engagement with the coach’s suggestions.

Quantitative Content Analysis was employed as the primary methodology for system evaluation. This involved the systematic coding of 4,352 messages exchanged between users and the coaching system, focusing on identifying instances of user agency and autonomy. Two independent coders analyzed the message content, achieving an inter-rater reliability score of 0.82, indicating a high degree of consistency in coding practices. Coding categories were derived from established frameworks for analyzing dialogue acts and user initiative in human-computer interaction. The resulting data allowed for a statistically significant assessment of the system’s ability to support user-led interactions and facilitate negotiation of coaching suggestions.

Beyond Goals: Cultivating Resilience and Adaptability

The LLM-Based Wellbeing Coach moves beyond simple goal-setting by actively identifying and interrupting patterns of rumination – repetitive negative thinking – in users. The system isn’t merely reactive; it proactively encourages a shift in focus from dwelling on problems to engaging in constructive action. This is achieved through carefully crafted conversational prompts designed to redirect attention towards potential solutions or positive reframing, fostering a more adaptive and solution-oriented mindset. By addressing rumination head-on, the coach aims to break the cycle of negative thought and empower individuals to take meaningful steps toward improved wellbeing, ultimately cultivating psychological flexibility and resilience.

The LLM-based wellbeing coach prioritizes the creation of a secure digital space where users feel comfortable expressing their emotions, a process known as emotional disclosure. This isn’t simply about venting; the system is designed to gently guide individuals through self-reflection, helping them to process complex feelings and understand the underlying reasons for their emotional responses. By offering a non-judgmental and supportive environment, the coach encourages a deeper exploration of personal experiences, enabling users to move beyond surface-level thinking and engage in meaningful emotional processing. This facilitated disclosure is considered a crucial step towards improved mental wellbeing, allowing individuals to reframe negative experiences and build resilience.

Analysis of interactions with the LLM-based wellbeing coach revealed a high degree of emotional disclosure among participants, observed in 68.4% of the study group. This suggests the system successfully cultivated a supportive conversational environment, encouraging users to openly share their feelings and experiences. The prevalence of this disclosure indicates the coach’s ability to establish rapport and trust, crucial elements in facilitating self-reflection and emotional processing. This finding highlights the potential for AI-driven tools to not only address specific mental wellbeing challenges, but also to provide a safe space for vulnerable expression and personal growth.

Analysis of user interactions revealed that patterns of rumination – repetitive thinking focused on negative experiences – emerged in a relatively small proportion of participants, specifically 13.2% (5 out of 38). This finding suggests the LLM-based wellbeing coach demonstrates a capacity to effectively interrupt and address these unproductive thought cycles. The low incidence of rumination observed implies the system’s conversational approach and supportive environment are successful in guiding users away from dwelling on negative experiences and towards more constructive cognitive pathways, ultimately contributing to improved emotional wellbeing.

Analysis of user interactions reveals a noteworthy level of engagement with the LLM-based wellbeing coach. Participants exchanged an average of 56.71 messages, with a standard deviation of 22.16, indicating considerable conversational depth and individualized use. Furthermore, the average message length of 6.42 words – with a standard deviation of 4.48 – suggests that users were not simply offering brief acknowledgements, but actively formulating and sharing thoughts and feelings. These metrics collectively demonstrate the system’s capacity to foster meaningful dialogue and sustain user interest, paving the way for effective emotional processing and the cultivation of wellbeing strategies.

Analysis of user interactions revealed a dynamic process of negotiation and adaptation, indicating the LLM-based wellbeing coach effectively fosters self-efficacy. Participants didn’t simply receive advice; rather, they engaged in a conversational exchange where the system responded to individual needs and adjusted its guidance accordingly. This iterative process-characterized by users testing boundaries, seeking clarification, and refining their approaches with the coach’s support-cultivated a belief in their capacity to navigate challenges and make meaningful progress. The observed patterns suggest the system isn’t merely a source of information, but a facilitator of personal agency, empowering individuals to actively shape their wellbeing journey and build confidence in their own abilities.

The pursuit of wellbeing extends beyond the attainment of specific objectives; a truly robust psychological state hinges on the capacity for sustained contentment and adaptable thinking. This research suggests that interventions should prioritize cultivating psychological flexibility – the ability to embrace experiences with openness and navigate challenges with resilience – rather than solely focusing on goal completion. By fostering this adaptability, individuals become less susceptible to setbacks and better equipped to maintain a positive outlook, even amidst difficulty. This approach views wellbeing not as a fixed destination, but as an ongoing process of growth, learning, and mindful engagement with the world, promoting a lasting sense of inner peace and resourceful coping mechanisms.

The study meticulously carves away extraneous interaction, revealing the essential elements of effective wellbeing coaching. It prioritizes user agency-a deliberate reduction of unnecessary robotic direction-and focuses on the core negotiation dynamics that facilitate genuine autonomy support. This aligns with John McCarthy’s observation: “It is better to solve one problem well than many poorly.” The research doesn’t attempt to encompass all potential coaching scenarios, but instead concentrates on distilling the most crucial interaction patterns. By prioritizing these fundamental elements, the design of the robot wellbeing coach moves closer to a state of purposeful simplicity, where only the essential remains.

Where Do We Go From Here?

The temptation, naturally, will be to build bigger models. To layer complexity upon the already considerable edifice of large language models, hoping that emergent properties will magically resolve the difficulties of genuine wellbeing support. This paper suggests a different path – a return to fundamentals. The observed preference for user-led direction wasn’t a technical glitch to be ‘fixed’, but a signal. A quiet insistence that even a simulated empath requires a degree of deference. They called it negotiation; it might simply be respect.

Future work needn’t focus solely on improving the artificial intelligence, but on diminishing the perceived need for it. A truly effective wellbeing coach, robotic or otherwise, likely facilitates self-reliance, then fades into irrelevance. The current metrics – engagement, duration of interaction – seem oddly focused on the opposite. Perhaps the ultimate success will be measured not by how much the user talks to the system, but by how little.

The longitudinal aspect of this study is particularly valuable, and deserves expansion. Wellbeing isn’t a static state, and a system that cannot account for the slow currents of change, the inevitable regressions, is ultimately a sophisticated distraction. The real challenge isn’t building an AI that sounds supportive, but one that can patiently, and unobtrusively, witness the messy, protracted business of being human.

Original article: https://arxiv.org/pdf/2602.04478.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-02-05 08:49