Large Language Models in Peer-Run Community Behavioral Health Services: Understanding Peer Specialists and Service Users' Perspectives on Opportunities, Risks, and Mitigation Strategies

Peer-run organizations (PROs) provide critical, recovery-based behavioral health support rooted in lived experience. As large language models (LLMs) enter this domain, their scale, conversationality, and opacity introduce new challenges for situatedness, trust, and autonomy. Partnering with Collaborative Support Programs of New Jersey (CSPNJ), a statewide PRO in the Northeastern United States, we used comicboarding, a co-design method, to conduct workshops with 16 peer specialists and 10 service users exploring perceptions of integrating an LLM-based recommendation system into peer support. Findings show that depending on how LLMs are introduced, constrained, and co-used, they can reconfigure in-room dynamics by sustaining, undermining, or amplifying the relational authority that grounds peer support. We identify opportunities, risks, and mitigation strategies across three tensions: bridging scale and locality, protecting trust and relational dynamics, and preserving peer autonomy amid efficiency gains. We contribute design implications that center lived-experience-in-the-loop, reframe trust as co-constructed, and position LLMs not as clinical tools but as relational collaborators in high-stakes, community-led care.

💡 Research Summary

This paper investigates how large language models (LLMs) might be integrated into peer‑run community behavioral health services, focusing on the perspectives of frontline stakeholders—peer specialists and service users. Partnering with Collaborative Support Programs of New Jersey (CSPNJ), the authors conducted co‑design workshops using a method called comic‑boarding with 16 peer specialists and 10 service users. Participants examined a prototype LLM‑based recommendation system that could suggest resources, summarize conversations, and flag potential crises.

The central finding is that LLMs can reconfigure the relational authority that underpins peer support. Depending on how the technology is introduced, constrained, and co‑used, it can (1) sustain existing trust by acting as an information‑supplement, (2) undermine trust by presenting over‑confident or culturally mismatched advice, or (3) amplify relational authority when it transparently augments the peer specialist’s judgment. These dynamics crystallize into three tensions:

1. Scale ↔ Locality – LLMs draw on massive, generic corpora, but peer support relies on local knowledge, cultural nuance, and community‑specific resources. Participants reported that generic model outputs often missed or mis‑represented regional services, suggesting the need for hybrid models that combine LLM capabilities with locally curated datasets.

2. Trust ↔ Relational Dynamics – Peer support is built on mutual empowerment, trauma‑informed disclosure, and shared lived experience. An opaque or overly authoritative LLM can make users feel judged by a machine, eroding the co‑constructed trust essential to recovery. Participants called for transparent output provenance, uncertainty indicators, and a “human‑in‑the‑loop” decision point to preserve relational safety.

3. Autonomy ↔ Efficiency Gains – Automation promises to reduce caseload pressure, yet there is a risk of deskilling peer specialists and displacing their lived‑experience authority. The authors propose a design principle called “Lived‑Experience‑in‑the‑Loop,” where the LLM’s suggestions are always reviewed, edited, or rejected by a peer specialist before reaching the service user.

To mitigate the identified risks, the paper outlines concrete strategies: (a) build and continuously update region‑specific resource databases; (b) surface confidence scores and source citations alongside each recommendation; (c) design interfaces that make the peer specialist the final decision maker; (d) implement ongoing feedback loops and training sessions for both specialists and users; and (e) create a fail‑safe mechanism that alerts a human supervisor when the model produces hallucinations or contradictory advice.

The authors contribute three scholarly advances: (1) empirical insight into how peer‑run organizations—often under‑resourced and technologically marginal—perceive LLM opportunities and threats; (2) the introduction of “experience‑in‑the‑loop” as a guiding principle for AI design in high‑stakes, community‑led care, reframing judgment, trust, and authority as relational rather than purely algorithmic; and (3) actionable design implications that position LLMs not as clinical decision‑support tools but as relational collaborators that augment, rather than replace, human expertise.

Overall, the study demonstrates that responsible LLM integration in peer‑run behavioral health requires careful balancing of scale with locality, transparent co‑construction of trust, and preservation of peer autonomy, offering a roadmap for designers, policymakers, and technology vendors seeking to deploy AI in similar community‑centered contexts.