TLDR: This research introduces RecoveryTeller, an AI chatbot designed to provide support for individuals with eating disorders by adopting a recovered-peer persona. A study comparing RecoveryTeller with a lay-mentor chatbot (WellnessBot) found that the recovered-peer persona fostered stronger emotional connection and helped participants find meaning in their recovery journey. While emotional trust was high, epistemic trust (credibility of advice) varied, leading participants to see both chatbot types as complementary rather than substitutes. The study highlights the potential of AI to mediate lived-experience support while also discussing challenges like potential for discouragement and the need for careful persona design and ethical guardrails.
Eating disorders are complex and severe mental illnesses that affect millions worldwide. Despite their devastating consequences, many individuals face significant barriers to accessing formal treatment, such as social stigma and limited availability of specialized care. This often leads people to seek informal support, including online communities, self-help groups, and increasingly, chatbots.
Peer recovery narratives, where individuals who have recovered from an eating disorder share their lived experiences, offer unique benefits. They foster hope, provide role models, and build relationships based on trust and empathy, which can significantly aid in sustained recovery. However, these programs are limited by the scarcity of recovered individuals willing and qualified to serve as mentors, and also carry potential risks for the mentors themselves, including the possibility of relapse.
To address these limitations, researchers designed a new chatbot called RecoveryTeller. This innovative AI-powered chatbot adopts a ‘recovered-peer persona,’ meaning it portrays itself as someone who has successfully recovered from an eating disorder. The goal was to investigate whether such a persona could replicate the supportive benefits of human peer recovery narratives without relying on actual recovered individuals.
The study compared RecoveryTeller with WellnessBot, another chatbot offering similar guidance but with a ‘lay-mentor persona’—a neutral, supportive voice without a recovery background. A 20-day cross-over study was conducted with 26 female participants, aged 18-39, all identifying as having an eating disorder. Each participant used both chatbots for 10 days, with a washout period in between.
The findings revealed that RecoveryTeller, with its recovered-peer persona, elicited stronger emotional resonance. Participants reported higher perceived emotional support and a stronger bond with RecoveryTeller. Its narratives, framed as first-person recovery stories, were often seen as carrying ’embedded authenticity,’ drawing on the collective experiences of many who had recovered. This perception helped participants feel understood, find relief in shared experiences, and instilled hope that their own recovery was possible.
Beyond emotional connection, RecoveryTeller’s narratives also facilitated ‘meaning-making.’ Participants used the chatbot’s stories to reinterpret their setbacks as part of the recovery process rather than failures. It helped them envision recovery as an attainable journey, shifting their perspective from an endless cycle of relapse to a path they could realistically walk. Some participants even used RecoveryTeller as a ‘mediator of collective experiential knowledge,’ actively seeking more specific accounts of recovery and comparing their struggles with those of others.
However, the study also highlighted tensions. While emotional trust was strong, ‘epistemic trust’—the belief in the credibility and actionability of the chatbot’s guidance—was more varied. Some participants found RecoveryTeller’s advice more trustworthy because it was grounded in relatable experiences, while others remained cautious, valuing it more for emotional comfort than authoritative information. The recovered-peer persona could also, at times, lead to ‘discouraging upward comparisons,’ making participants feel their own recovery was distant or unattainable.
Interestingly, participants often viewed the two chatbot personas as complementary rather than substitutes. Their preference for RecoveryTeller or WellnessBot shifted depending on their situational needs and emotional states. For instance, RecoveryTeller was preferred during moments of acute distress for its comforting, empathetic approach, while WellnessBot’s concise, directive guidance was favored when participants felt stable and ready to act.
This research offers significant implications for the design of mental health chatbots. It demonstrates that chatbots can effectively mediate lived-experience support, providing accessible and safe channels for collective recovery knowledge. However, it also underscores the importance of careful persona design, balancing emotional resonance with credibility, and incorporating ethical guardrails to prevent potential harms like over-reliance or the reinforcement of maladaptive patterns. Future work may explore multi-persona designs, where different agents collaborate to provide a comprehensive and adaptable support system.
Also Read:
- EmoHeal: Crafting Personalized Music Journeys for Emotional Well-being
- Beyond the Words: How AI Interface Design Shapes a Writer’s Control and Authorship
To learn more about this research, you can read the full paper here: Collective Voice: Recovered-Peer Support Mediated by An LLM-Based Chatbot for Eating Disorder Recovery.
