PaperlayerView exact revision on arXiv

Source-linked AI summary

PeerPen: AI-Assisted Writing for Online Mental Health Peer Support

Jiwon Kim, Sherry Gong, Maya Ajit, Soorya Ram Shimgekar, Yunhao Yuan, Dong Whi Yoo, Eshwar Chandrasekharan, Koustuv Saha

arXiv:2609.14886v1cs.HCcs.AIcs.CLcs.CYcs.SI

TL;DR

Peer supporters may struggle to articulate helpful responses, and less is known about how they collaborate with AI while supporting others and preserving trust. The authors built PeerPen, a Reddit-like writing assistant with draft generation and user-text revision, and studied its use through semi-structured interviews with 15 participants. Participants reported lower effort composing responses, but anticipated that uncertainty about AI involvement could unsettle authorship and trust across the community.

  • Problem

    Peer supporters may struggle to articulate helpful responses, and less is known about how they collaborate with AI while supporting others and preserving trust.

    Source: §1 · p. 1

  • Method

    The authors built PeerPen, a Reddit-like writing assistant with draft generation and user-text revision, and studied its use through semi-structured interviews with 15 participants.

    Source: §1 · p. 2

  • Results

    Participants reported lower effort composing responses, but anticipated that uncertainty about AI involvement could unsettle authorship and trust across the community.

    Source: §6 · pp. 19–20, §7 · p. 25

  • Takeaways & Limitations

    Participants’ sense of ownership depended on remaining the source of communicative intent and retaining decision authority over how that intent was expressed.

    Source: §1 · p. 2

    The small sample and single one-hour session preclude generalizable or long-term effectiveness claims.

    Source: §6.6 · p. 24

Abstract

from arXiv · show

Online mental health communities thrive on peer support, yet those who volunteer to help often lack formal training and may struggle to articulate supportive responses. AI co-writing could lower this barrier; however, peer support derives much of its value from being perceived as personal, raising questions around authorship, ownership, and trust. We built PeerPen, a writing assistance tool embedded within a Reddit-like interface, supporting two main features: draft generation and revision of user-written responses. Through semi-structured interviews with 15 participants, we find that PeerPen reduced the burden of composing responses and increased confidence in offering support. Participants wanted AI to assist their writing without taking over authorship and anticipated tensions around authenticity and trust. Such assistance could make authorship uncertain even for responses written without it, weakening trust across the community. We contribute design implications for AI writing assistance that scaffolds supportive communication, preserves authorship, and accounts for community-level trust.

1 Introduction

The paper examines how AI writing assistance can support peer responses in online mental health communities while preserving human authorship and community trust. PeerPen pairs draft generation with revision of user-written text, and interviews reveal both reduced writing burden and concerns about authenticity and trust.

Motivation

  • Online mental health community members may struggle to find appropriate words, balance empathy with practical guidance, or avoid harm, and some refrain from responding despite wanting to help.Much of this support comes from non-expert peers.Source: §1 · p. 1

Research gap

  • Less is known about how people collaborate with AI while supporting others than about the quality, safety, and empathy of AI-generated responses.The paper treats assistance as part of how people formulate and refine their own supportive messages.Source: §1 · pp. 1–2

Research question

  • The study asks how peer supporters perceive and use AI-assisted writing, and what design opportunities and challenges it presents for support, trust, and community norms.The authors argue that assistance should fit supporters’ writing processes and account for trust and social norms.Source: §1 · p. 2

Approach

  • PeerPen offers AI-generated drafts and revisions of user-written responses, and 15 participants used it to respond to real community posts before semi-structured interviews.The two assistance modes let the study examine how different forms of AI involvement relate to authorship.Source: §1 · p. 2

Findings

  • Participants valued reduced composition effort but wanted to retain authorship; they anticipated that uncertainty about AI involvement could cast doubt even on wholly human-written responses.The authors call this anticipated community-wide uncertainty authorship spillover.Source: §1 · p. 2

Contributions

  • The paper contributes a tool, an account of process-dependent authorship, and design recommendations to scaffold rather than replace support while preserving authorship and attending to community norms.Participants’ ownership depended on whether they remained the source of communicative intent and retained decision authority over its expression.Source: §1 · pp. 2–3

2 Related Work

Prior work describes challenges for peer supporters, trade-offs in human–AI writing, and a growing body of mental-health AI research. This paper focuses on how supporters engage with writing assistance and negotiate authorship within peer-support communities.

2.1 Peer Support in Online Mental Health Communities

  • Peer support can be valuable when formal care is inaccessible or stigmatized, but replies may be generic or unsafe and non-expert supporters can be uncertain about responding.Prior work also identifies privacy and data-governance concerns when supporters move sensitive content to general-purpose AI tools.Source: §2.1 · p. 3

2.2 AI-Assisted Writing and Human–AI Collaboration

  • AI writing assistance spans generation, completion, localized suggestions, and revision, with writers balancing support against control, authorship, and ownership.Prior work describes scaffolding as potentially supporting learning and reflection, while higher levels of assistance may reduce perceived ownership.Source: §2.2 · pp. 3–4

2.3 AI in Emotionally Sensitive and Mental Health Contexts

  • Mental-health AI research often evaluates systems that provide support directly, while less is known about how peer supporters use AI writing assistance and negotiate authorship.Existing work includes tools for empathy feedback, counselor training, and supportive-response generation; this paper examines AI-generated and AI-revised writing within the supporter’s process.Source: §2.3 · p. 4

3 PeerPen: An AI-Powered Writing Assistance Tool for Peer Supporters

PeerPen embeds AI writing assistance in a Reddit-like peer-support workflow, pairing draft generation with localized revision. Its design goals emphasize reducing composition effort while keeping supporters active, reflective, and in control of what they share.

3.1.1 Lowering the Effort of Composing Supportive Responses.

  • PeerPen’s first design goal is to reduce the effort of beginning and composing supportive responses, a task that can challenge non-expert volunteers.The design aims to help users translate supportive intent into a response they can shape and stand behind.Source: §3.1.1 · p. 5

3.1.2 Preserving Ownership over the Supportive Response through Iterative Collaboration.

  • PeerPen’s iterative-collaboration goal preserves ownership by keeping users active contributors rather than shifting them into merely evaluating AI-written text.The design rationale notes that the amount and form of AI involvement can shape control and ownership.Source: §3.1.2 · p. 6

3.1.3 Guiding Responders Toward More Supportive Writing.

  • Its supportive-writing goal is to help users recognize and reflect on how to communicate support through localized, inspectable suggestions.The design rationale frames in-process guidance as a way to help responders consider empathy, validation, advice, and support.Source: §3.1.3 · p. 6

3.1.4 Fitting into Existing Community Environments.

  • Its platform-integration goal places assistance inside the community environment, where it can fit familiar workflows and community practices.The authors note that integration may give platforms greater control over data made available to models and how it is handled.Source: §3.1.4 · p. 6

3.2 Writing Assistance Features

  • Generation offers an editable initial draft, while Improvement revises user-written text through localized suggestions users can adopt or ignore individually.Users can move between the features, edit freely, and remain responsible for the response they submit.Source: Figure 1 · p. 7, Figure 2 · p. 7, §3.2.1 · p. 8

3.2.3 Backend Language Models.

  • The backend models, adapted from Mistral-7B using OMHC interactions and community preference signals, generate or revise supportive responses; prior evaluation reported strong ratings and preference for revised responses.The Generation Model averaged 4.41/5 for empathy and 4.49/5 for safety; participants preferred Improvement Model revisions over originals in 82.0% of comparisons.Source: §3.2.3 · p. 8

3.3 Interface Design

  • The Reddit-like interface places Generate and Improve in the composer and presents replacements and insertions as separate, visually legible suggestions.Users can apply or ignore recommendations individually instead of automatically rewriting the whole response.Source: §3.3 · p. 9, §3.3.1 · p. 9, §3.3.2 · p. 9

3.4 System Design and Architecture

  • A React frontend communicates with a Flask backend that routes requests to writing models and logs participant actions; revision feedback is structured into discrete Replace and Insert actions.Normalized matching keeps replacements applicable after surrounding text changes, while insertions enter at the current cursor position.Source: §3.4 · p. 9, Figure 3 · p. 10

4 User Study

The study used interviews and a simulated peer-support writing workflow to examine PeerPen’s use, perceived ownership, and fit with community norms. The team analyzed interview and interaction data qualitatively and used post-response ratings as a complement, not an independent model evaluation.

4 User Study

  • The user study examined how PeerPen shapes composition, ownership, and fit with online mental-health community norms through a simulated supportive-writing workflow.The study was IRB-approved and recruited participants experienced in online mental-health communities.Source: §4 · p. 10

4.1 Recruitment and Participants

  • The team recruited 15 participants with OMHC experience in the U.S.; eight primarily described themselves as observers who cited discomfort writing supportive responses.Recruitment was conducted through communities whose moderators granted permission and through researchers’ social-media accounts.Source: §4.1 · pp. 10–11

4.2 Study Procedure

  • Participants completed four to five writing tasks, first trying both assistance modes and then choosing how to use them while thinking aloud.They also completed a post-interaction survey after each submitted response and a semi-structured interview about their experience.Source: §4.2.1 · pp. 11–12

4.3.1 Inductive Coding and Thematic Analysis.

  • The researchers reviewed transcripts, think-aloud responses, and notes using inductive coding followed by reflexive thematic analysis.Three co-authors calibrated their coding on the first three transcripts; the team identified 315 open codes and consolidated them into five themes.Source: §4.3.1 · p. 12

4.3.2 Post Interaction Survey Analysis.

  • Survey means were reported separately for GF and IF responses across five shared quality dimensions and interpreted alongside qualitative accounts, not as an independent model-performance evaluation.The five-item survey covered readability, empathy, relational connection, actionable guidance, and avoidance of harmful or judgmental suggestions.Source: §4.3.2 · pp. 12–13, §4.2.1 · p. 12, Table 3 · p. 13

4.4 Safety, Privacy, and Reflexivity

  • The study used consent, voluntary participation, options to pause or withdraw, debriefing, and de-identification to address participant wellbeing and privacy.Participants were responding to posts involving suicide, depression, and emotional distress.Source: §4.4 · p. 13
  • The authors acknowledge that their team’s expertise and lived perspectives may have shaped how they interpreted participants’ viewpoints.They report reflecting on those perspectives during analysis to represent participants’ views faithfully.Source: §4.4 · p. 13

5 Findings

Participants rated both PeerPen features highly for readability and safety, while actionability favored revisions over generated drafts. Ratings for preserving content and style varied among participants.

5.1 Perceived Quality of AI-Assisted Responses

  • Readability and safety were rated at or near 4.8 for both generated and revised responses.Empathy and connection ratings fell between 4.24 and 4.63.Source: §5.1 · pp. 13–14
  • Actionability averaged 3.82 for generated responses and 4.24 for revised responses, with a third of generated-response ratings at 3 or below.Revised-response actionability ratings concentrated at the upper end of the scale.Source: §5.1 · p. 14
  • Content preservation averaged 4.33 and style preservation 4.22 for revisions, but these ratings had the table’s largest standard deviations.Ratings spanned the full scale, suggesting participants differed in whether revisions still felt like their own.Source: §5.1 · p. 14

5.2 Assistance Lowers the Barrier to Responding but can Shift Work into Editing

PeerPen eased the cognitive and emotional burden of starting supportive replies, especially for difficult posts, but its benefits depended on output that users could readily adapt. Slow, repetitive, generic, or mismatched suggestions sometimes added editing work.

5.2.1 Reducing the Burden of Getting Started.

  • PeerPen reduced the cognitive difficulty of finding words and the emotional labor of engaging with distressing posts.Participants described it as a way to begin and as lifting the emotional weight of writing from scratch.Source: §5.2.1 · p. 14
  • For difficult or high-stakes posts, participants said a starting point could ease panic about saying the wrong thing and reduce response time and energy.P02 described risky situations where a responder might make a suggestion that could be misinterpreted.Source: §5.2.1 · p. 14
  • Some participants, including observers who had not written supportive replies, said lower effort made them more willing and confident to respond.P01 reported more energy to respond and greater willingness to comment on emotionally heavy posts.Source: §5.2.1 · p. 14

5.2.2 When Assistance Introduced Additional Effort.

  • Slow responses and repetitive, lengthy suggestions interrupted writing and required participants to read overlapping recommendations.Generated text could also be superficial, generic, or cliché, and sometimes contained spelling or grammar errors participants had to correct.Source: §5.2.2 · p. 14
  • Assistance saved effort only when its output was close enough to users’ intent that editing cost less than composing anew.When the mismatch was wide, participants faced an additional editing task; the paper characterizes this as a condition of effort reduction.Source: §5.2.2 · p. 15

5.3 Authorship Depends on Who Supplies Intent and Directs Revision

Participants’ sense of ownership depended on how they contributed through the writing process, not simply on how much text came from them. User-led drafting and selective revision supported collaboration, while AI-led drafts could leave users feeling like proofreaders rather than authors.

5.3.1 Supporting Ownership through Iterative Collaboration.

  • Participants described revision as collaborative when they wrote first and then used AI to strengthen their response.P12 described the effort as roughly an even split while preferring to write something first.Source: §5.3.1 · p. 15
  • Discrete insert and replace suggestions let participants adopt some changes and reject others, supporting personalization and a sense of shared authorship.Participants valued choosing what to add or leave out rather than accepting a wholesale rewrite.Source: §5.3.1 · p. 15

5.3.2 When Assistance Displaced Authorship.

  • When the system supplied text that participants then edited, some described their role as cleaning up AI work rather than collaborating.P14 said the arrangement would have worked better if they had written something first for PeerPen to refine.Source: §5.3.2 · p. 15
  • For some participants, AI-written responses did not feel authentic or trustworthy even when the text was good, because they had not written the message themselves.One participant said they could not trust the tool to convey a genuine message because it felt like cheating rather than helping.Source: §5.3.2 · p. 16
  • Participants said AI could not draw on their personal experiences to relate to a poster, leaving its output general where theirs could be personal.This limitation compounded concerns that suggested responses did not sound like them and required further editing.Source: §5.3.2 · p. 16

5.4 Localized Revision Supports Reflection, but Inspectability is Not Explanation

Participants said revision exposed them to new ways of expressing support and approaching emotionally heavy posts. However, seeing proposed edits did not always explain their supportive rationale, and the markup could be hard to interpret.

5.4.1 Supporting Reflection through Revision.

  • Five participants said assistance surfaced phrasings and considerations they would not have produced on their own.P12 said it filled gaps and helped them be more considerate of the original poster.Source: §5.4.1 · p. 16
  • Participants described learning new phrases, considering different approaches, and changing how they might respond to emotionally heavy posts.Suggestions also included asking the poster questions and offering more concrete, action-oriented guidance.Source: §5.4.1 · p. 16

5.4.2 When Revision Guidance Fell Short.

  • The interface showed what text changed but did not explain why a revision might make a response more supportive, leaving participants to infer the rationale.The paper notes that inspectability alone did not necessarily explain the supportive principle behind a revision.Source: §5.4.2 · pp. 16–17
  • Participants sometimes struggled to identify which part of a longer response a suggestion targeted and requested clearer highlighting of the corresponding draft text.Lengthy or repetitive suggestions also made it harder to distinguish substantive guidance from stylistic rewriting.Source: §5.4.2 · p. 17

5.5 In-Situ Assistance Fits Familiar Workflows but Requires User Steering

Participants found PeerPen familiar and easy to use within a Reddit-like workflow, but unclear insertion placement made applying suggestions harder. They valued workflow fit while identifying interaction-level friction.

5.5.1 Fitting Assistance into Existing Workflows.

  • Participants found the interface intuitive and familiar, with assistance blending into Reddit and generated text matching the platform’s comment tone.Several could imagine using it regularly or integrating it directly into Reddit.Source: §5.5.1 · p. 17

5.5.2 Frictions in Platform Integration.

  • Unmarked cursor-based insertions made it difficult to anticipate where suggestions would land, requiring participants to track and verify placement.Participants located friction in the suggestion mechanics rather than the surrounding environment.Source: §5.5.2 · pp. 17–18

5.6 AI Assistance Creates Trust Tensions Beyond the Writer

Participants valued PeerPen as a writing aid but raised concerns about how AI assistance could affect supporters’ engagement, recipients’ sense of human care, and trust across the broader community. They anticipated that widespread use could make authorship uncertain even for comments written without AI.

5.6 AI Assistance Creates Trust Tensions Beyond the Writer

  • Even participants who found PeerPen helpful or supported the concept questioned whether AI assistance belonged in mental health peer support.Their concerns extended beyond usability and system performance to what widespread adoption could mean for the community.Source: §5.6 · p. 18, §5.6.1 · p. 18

5.6.1 Trust from the Perspective of Peer Supporters.

  • Some supporters rejected AI responses to emotionally heavy or suicidal posts and worried that easier composition could reduce the attention and empathy given to the poster.One participant would use the tool for an anonymous recipient but not for someone they knew personally.Source: §5.6.1 · p. 18

5.6.2 Trust from the Perspective of Support Seekers.

  • Support seekers said knowing a response was AI-generated could make it feel inauthentic, less serious, or like a failure of human care.Some feared that a person in crisis might feel worse receiving an AI-written response than receiving no response.Source: §5.6.2 · pp. 18–19

5.6.3 Trust from the Perspective of the Broader Community.

  • Participants anticipated that widespread use could make responses monotonous or spam-like, and that unreviewed mistakes could further make them feel artificial.These concerns were about consequences at community scale, not only the quality of an individual response.Source: §5.6.3 · p. 19
  • Widespread AI availability could create authorship spillover, making people doubt whether any comment was genuinely written by a person and weakening community trust.Participants worried that this ambiguity could persist even when individual AI-assisted responses were high quality and well intentioned.Source: §5.6.3 · p. 19

6 Discussion

PeerPen reduced the burden of composing supportive responses, but participants’ accounts show that response quality alone does not determine whether an interaction feels like peer support. The discussion argues for assistance that preserves user agency and attends to the relational and community-level meanings of authorship.

6.1 Beyond Response Quality: Preserving the Interpersonal Relational Meaning of Peer Support

  • Reducing composition effort can help people participate, but removing the effort of attending to a poster may change what a supportive response signifies.Participants distinguished the quality of the text from its relational meaning: a response can be empathetic and well written while still raising doubts about genuine engagement.Source: §6 · p. 19, §6.1 · p. 20

6.2 Authorship as a Process of Human–AI Collaboration in Peer Support

  • Participants’ ownership depended more on who initiated and directed the writing than on how much text the AI supplied.Localized suggestions let users evaluate changes against their own writing; AI-first drafts could leave them feeling more like proofreaders. Authorship also shaped whether supporters could stand behind a response and whether recipients read it as personal engagement.Source: §6.2 · pp. 20–21

6.3 From Individual Assistance to Community-Level Trust

  • At community scale, participants anticipated similar, unreviewed, or spam-like responses and uncertainty about whether responders had read the posts.Widespread availability could unsettle how people interpret all comments, including entirely human-written ones; communities may differ in which forms of assistance fit their norms.Source: §6.3 · p. 21

6.4.1 AI Assistance that Scaffolds rather than Supplies Peer Support

  • Future systems should vary assistance to match where a supporter is struggling, making full-response generation one option rather than the default.Outlines, example phrases, and localized revision offer different levels of help; progressive scaffolding may reduce burden while preserving room for users’ intent and experience, though repeated scaffolding may foster dependency for some.Source: §6.4.1 · pp. 21–22

6.4.2 Design Revision as In-Situ Guidance

  • Visible revisions can help supporters notice alternative ways to communicate, but guidance should also explain the supportive principle behind a change.The paper proposes lightweight explanations and frames revision as situated guidance; whether repeated exposure transfers to unaided writing remains a question for longitudinal study.Source: §6.4.2 · p. 22

6.4.3 Supporting User Steering without Ceding Authorship; 6.4.4 Personalizing Assistance while Preserving Granular User Control

  • Users wanted to express intent before assistance and to select, compare, and reverse localized changes while preserving their original writing.Personalization could use prior writing or specified tone, but should give users control over whether that writing is used and retained.Source: §6.4.3 · p. 23, §6.4.4 · p. 23

6.4.5 Making AI Involvement Transparent and Community-Governed; 6.5 Ethical Implications

  • The authors propose granular provenance and community-level choices about assistance and disclosure, while emphasizing that integration alone does not ensure privacy.Communities could set permitted assistance levels and rules for high-risk posts, disclosure, and personalization; deployment still needs explicit policies for data access, retention, and model use.Source: §6.4.5 · pp. 23–24, §6.5 · p. 24

6.6 Limitations and Future Directions

  • The small, single-session study and participant self-selection limit generalizable or long-term claims and may underrepresent stronger opposition to AI-assisted peer support.Participants were U.S.-based, reported English as their first language, and primarily considered the supporter perspective; recipient reactions were not directly established.Source: §6.6 · pp. 24–25

7 Conclusion

PeerPen reduced the burden of composing supportive responses and helped participants identify alternative ways to communicate support, but they wanted to retain control over what to say and how to say it. The study highlights designing assistance that scaffolds rather than replaces communication while accounting for authorship, trust, and community norms.

7 Conclusion

  • In a study with 15 participants, PeerPen reduced some composing burden, while participants emphasized user agency and concerns about authenticity, authorship, and trust.The paper concludes that response quality alone cannot capture the meaning of supportive communication; the writing process also matters.Source: §7 · p. 25

A Appendix

Table A1 summarizes participant-reported strengths and gaps alongside resulting design implications. It organizes them by design goal and concerns surrounding community trust.

A Appendix

Loading 2609.14886v1…