Source-linked AI summary
Customizing Emotional Support: How Do Individuals Construct and Interact With LLM-Powered Chatbots
Xi Zheng, Zhuoyang Li, Xinning Gui, Yuhan Luo
TL;DR
Emotional support needs vary, yet how chatbots can adapt to individuals’ needs and what customization they require remain underexplored. In a week-long field study with 22 participants using ChatLab, the authors found diverse chatbot personas and customization practices that shaped relationships and promoted open discussions.
Problem
How chatbots can adapt to differing emotional-support needs, and what customization people need, remains underexplored despite the importance of individualized support.
Source: §1 · p. 2, §2.1 · p. 3Method
Using a Research through Design approach, the authors deployed ChatLab with 22 participants for a week-long field study followed by interviews and design activities.
Source: §3 · p. 4, §9 · p. 17Results
Participants created diverse personas for emotional reliance, stressors, intellectual discourse, self-discovery, and therapeutic support, using voice and avatar options to shape relationships and promote open discussions.
Source: §1 · p. 2, §9 · p. 17Takeaways & Limitations
The study’s design implications focus on better supporting chatbot customization for individualized emotional-support needs.
Source: §1 · p. 2The study did not evaluate whether chatbot customizability improves mental well-being.
Source: §8 · p. 17
Abstract
from arXiv · showhide
Personalized support is essential to fulfill individuals' emotional needs and sustain their mental well-being. Large language models (LLMs), with great customization flexibility, hold promises to enable individuals to create their own emotional support agents. In this work, we developed ChatLab, where users could construct LLM-powered chatbots with additional interaction features including voices and avatars. Using a Research through Design approach, we conducted a week-long field study followed by interviews and design activities (N = 22), which uncovered how participants created diverse chatbot personas for emotional reliance, confronting stressors, connecting to intellectual discourse, reflecting mirrored selves, etc. We found that participants actively enriched the personas they constructed, shaping the dynamics between themselves and the chatbot to foster open and honest conversations. They also suggested other customizable features, such as integrating online activities and adjustable memory settings. Based on these findings, we discuss opportunities for enhancing personalized emotional support through emerging AI technologies.
1 Introduction
The study examines how people use LLM chatbot customization to meet varied emotional-support needs, and reports how they shaped chatbot personas and imagined further customization. It contributes empirical findings and design implications for personalized support, while focusing on everyday well-being rather than clinical advice.
Motivation
- Emotional-support needs vary, but prior chatbot research has rarely examined individual customization, its motivations, or attributes beyond language outputs.Prior work largely focused on emulating mental-health professionals [16] [53] [98] or exploring emotional-support use cases [37] [87].Source: §1 · p. 2
- Natural-language prompting lets nonexperts customize chatbot personality, communication style, and knowledge, potentially enabling personalized support through a cost-effective channel [37] [87].People seek different forms of support, including confiding, empathetic understanding, companionship, and practical guidance [80], tailored to personal experience and stressors.Source: §1 · p. 2
Study approach
- Using Research through Design, the researchers built ChatLab with open-text hints and avatar and voice choices, then studied 22 participants with moderate-to-high social loneliness for seven to ten days, followed by interviews and design activities.Participants proposed additional features during the interviews and design activities.Source: §1 · p. 2
Findings
- Participants created personas for emotional reliance, stress, intellectual discourse, self-discovery, and therapeutic support, using attributes, voices, and avatars to shape relationships and encourage open, honest discussion.They also proposed digital-trace, environment, and emotional-state learning; conversation management; and AI or community assistance to reduce customization effort.Source: §1 · p. 2
Contribution and scope
- The work contributes an account of real-world chatbot construction for daily emotional challenges and design implications for support tailored to individual needs.The study addresses emotional support such as stress coping and attentive listening, not clinical advice or assistance for severe mental-health issues.Source: §1 · pp. 2–3
2 Related Work
Related work traces emotional-support customization from tailored care and pre-LLM chatbots to flexible LLM-powered companions, while identifying limited understanding of how people customize and use them. The paper addresses this gap by examining customization practices, emotional status, interaction experiences, and expectations.
2.1 Existing Approaches to Support Individuals’ Emotional Well-Being
- Personalized mental-health care can match support to individual traits, such as emotional reactivity, because personal characteristics shape care effectiveness and outcomes.For example, emotional reactivity may inform choices between CBT, family-based therapy, and mindfulness-based stress reduction.Source: §2.1 · p. 3
- Because professional care is not accessible to everyone, self-care tools and chatbots offer alternatives, with chatbot customization spanning communication style, appearance, and personality.IntelliCare’s range of intervention choices supported autonomy, encouragement, goal setting, and reflection; speech speed and formality affected chatbot-performance perceptions.Source: §2.1 · p. 3
- Pre-LLM chatbot studies explored only limited, surface-level customization, leaving people’s emotional-support customization needs and motivations underexplored.Technical challenges constrained flexible customization before LLMs.Source: §2.1 · p. 3
2.2 Chatbot Customization in The Age of Large Language Models (LLMs)
- LLMs let users steer chatbot tone and domain knowledge through prompts, enabling more flexible personas and active design of conversational partners [37].Clochat offered persona options and users enjoyed and engaged more with customized personas [37]; Character.AI enabled custom characters from scratch.Source: §2.2 · p. 3
- LLM mental-health support studies found varied uses and prompting practices, but did not examine how people navigate interactions, their motivations, or customization of voices and avatars [87].This study examines customization practices and their interplay with emotional status, interaction experiences, and expectations for LLMs in everyday life.Source: §2.2 · pp. 3–4
3 ChatLab: A Prototype for Research Exploration
ChatLab was built through a Research through Design approach as a prototype for exploring how people customize LLM chatbots for emotional support, not as an evaluation of their effectiveness. Its design balanced user autonomy with guidance, expanded customization beyond text, and encouraged customization across real-world experiences.
3 ChatLab: A Prototype for Research Exploration
- ChatLab served as a Research through Design prototype for eliciting customization experiences and design requirements, rather than testing emotional-support effectiveness.The study asked how people construct and interact with customizable LLM chatbots and what design opportunities could better meet individual emotional needs.Source: §3 · p. 4
3.1.1 DR1: Facilitating chatbot customization while maintaining user autonomy.
- The design balanced guidance for non-experts—such as hints, examples, and prompt-editing support—with autonomy to express unique needs and preferences.The authors noted that composing effective prompts can be difficult for lay users, while pre-created personas and options may discourage exploration of their own ideas.Source: §3.1.1 · p. 4
3.1.2 DR2: Expanding customization possibilities to enrich user ex-
- ChatLab extended customization beyond text by adding chatbot avatars and voice tones to enrich personas and conversational engagement.These non-verbal cues complement the chatbot’s language and communication style.Source: §3.1.2 · p. 4
3.1.3 DR3: Encouraging active customization for research explo-
- Because emotional needs and preferred conversation partners can vary over time and occasions, ChatLab was designed to encourage active customization throughout the study.The rationale was that people may not create new personas without inspiration or effort-reducing guidance.Source: §3.1.3 · p. 4
3.2 ChatLab Design Components
- ChatLab was a mobile-responsive website with four components: onboarding, a customization and conversation playground, an experience diary, and FAQs.The components were designed to address the three rationales for the prototype.Source: §3.2 · p. 4
3.3 Implementation
- ChatLab used Streamlit, OpenAI GPT models with user prompts inserted following LangChain guidance, and text-to-speech APIs; interaction data were stored in Firebase with access restricted to the research team.The architecture supported integration of multiple APIs for chatbot customization and interaction.Source: §3.2 · p. 4, §3.3 · p. 5
3.2.1 Onboarding Page and FAQs.
- Onboarding introduced the study and chatbot modalities, while FAQs explained LLMs, prompts, prompt-writing tips, study logistics, and research-team contact information.The FAQ content was informed by pilot studies and aimed to address questions from non-AI experts.Source: §3.2.1 · p. 4
3.2.2 Customization and Conversation Playground.
- The playground combined guided, editable prompt creation and AI polishing with settings for output modality, avatars, GPT models, and response randomness.It offered 70 Mandarin voices and 77 emoji avatars; users could adjust settings during chats, switch between voice and text, and revisit prior settings and conversation history. The diary captured scenarios, helpful or memorable moments, chatbot settings and reasons, and additional thoughts.Source: §3.2.2 · pp. 4–5, §3.2.3 · p. 5
4 Method
The study combined a feasibility pilot with a three-phase formal study to examine how people customized and used ChatLab for emotional support. Qualitative analysis drew on participants’ chatbot settings, interviews, design artifacts, conversation logs, and diaries.
4 Method
- The formal study followed a tutorial, a field study with ChatLab, and post-study interviews and design activities; the protocol received university ethics approval.Source: §4 · p. 5
4.1 Pilot Study
- Six university students used ChatLab daily for a week, then joined a two-hour focus group; their feedback informed usability changes, and pilot logs showed no inappropriate or harmful outputs.Changes included editable prompts, conversation history, clearer customization hints, and an LLM FAQ page.Source: §4.1 · p. 5
4.2.1 Participants & recruitment.
- Of 149 screened respondents, 22 participants were analyzed; eligibility required adults facing emotional struggles and moderate or high loneliness, while severe mental health conditions were excluded.The analyzed group was aged 21–40 and all scored above 40/80 on the UCLA Loneliness Scale.Source: §4.2.1 · pp. 5–6
4.2.2 Pre-study tutorial.
- A roughly 30-minute group tutorial explained ChatLab’s prompt construction and voices, shared pilot-derived prompting strategies, encouraged creative persona exploration, and cautioned against seeking clinical advice.Source: §4.2.2 · p. 6
4.2.3 Field study.
- Participants used ChatLab daily for seven to ten days, customizing and interacting with chatbots as needed and submitting at least five diary entries.They could adjust settings during conversations; reminders were sent daily, and the field study prepared them for interviews and design activities.Source: §4.2.3 · p. 6
4.2.4 Post-study interview & design activity.
- The post-study phase included online interviews and a think-aloud design activity in which participants envisioned an ideal emotional-support chatbot.Interviews discussed ChatLab experiences, while participants used a shared Excalidraw whiteboard to develop designs; two declined interviews and two deceptive screeners were excluded.Source: §4.2.4 · pp. 6–7
4.3 Data Analysis
- Analysis combined chatbot settings and usage statistics with interviews, design artifacts, conversation logs, and diary entries, using thematic analysis to develop 693 codes into higher-level themes.Researchers independently coded initial data, resolved discrepancies in meetings, and continuously compared later codes with earlier ones.Source: §4.3 · pp. 7–8
- The study prioritized contextual qualitative insight over quantification and anonymized participant data using P# labels.Source: §4.3 · p. 8
5 RQ1. Customization Practice
Participants used ChatLab extensively and created varied chatbot personas, while avatar and voice customization differed across participants. Their practices and rationales formed the basis for examining how they constructed and interacted with emotional-support chatbots.
5 RQ1. Customization Practice
- Participants engaged in 1,541 conversation rounds across 178 sessions, averaging 8 sessions per person, with sessions lasting approximately 4 to 37 minutes excluding inactive time.The study logged 1,541 participant-message-and-chatbot-reply rounds.Source: §5 · p. 9
- Participants created 118 distinctive personas, ranging from 1 to 13 personas per participant.Source: §5 · p. 9
- All but three participants selected customized avatars for both themselves and their chatbots, while voice use ranged from active experimentation to rare or no use.Some participants rarely or never used voice interaction because of environmental constraints, such as sharing space with others.Source: §5 · p. 9
5.1 Persona Construction and Customization Experience
Participants customized chatbots into multiple personas for different emotional and conversational needs, with roles shaping the interactions and support they received.
5.1 Persona Construction and Customization Experience
- Participants created five persona types: emotional-reliance figures, stressor-related figures, intellectual interlocutors, mirrors of themselves, and therapeutic experts.Examples included beloved pets and close friends, philosophers and scientists, and psychologists.Source: §5.1 · p. 9
- Most participants created and interacted with multiple personas, while a few kept one persona and incrementally updated its name or personality.Source: §5.1 · p. 9
- Participants tailored personas to contexts—from a supervisor for academic pressure to family or romantic roles for loneliness or relationship regrets—and said roles changed wording and emotional support.Participants also used philosopher personas to seek intellectual insight amid work stress.Source: §5.1 · p. 9
5.2 Enriching The Constructed Persona
Participants enriched chatbot personas with social cues, especially voices and avatars chosen to fit identity or convey support. Supportive cues could make interactions feel comforting, while participants did not specifically match cues to stressor personas.
5.2 Enriching The Constructed Persona
- Participants incorporated voices, avatars, and other social cues to enrich the personas they constructed.Source: §5.2 · p. 9
5.2.1 Aligning voice and avatar choice to the persona identity.
- Participants commonly matched voices and avatars to persona identity, choosing tones that suited characters such as intellectual figures or therapists.An avatar visible throughout conversation also served as a reminder of the chatbot’s identity.Source: §5.2.1 · p. 9
5.2.2 Making supportive personas embody positiveness.
- For supportive personas, participants favored comforting voices and cheerful or cute avatars; familiar hometown accents helped some open up more easily.Participants described positive avatars as pleasant or comforting, and as reminders that the chatbot was kind, without claiming that an avatar alone changed their mood.Source: §5.2.2 · pp. 9–10
- Participants did not specifically match voices or avatars to stressor personas, possibly because stressors lacked strong emotional attributes or they avoided voice and visual confrontation.Source: §5.2.2 · p. 10
5.3 Shaping Conversation Dynamics
Participants shaped chatbot conversation dynamics by defining identities and personas, using personal anecdotes and role-play, and choosing avatars to express feelings and influence the interaction’s tone.
5.3.1 Incorporating personal anecdotes for both parties.
- Participants often used authentic and fictional personal anecdotes to prepare for conversations, describing both themselves and the chatbot.Source: §5.3.1 · pp. 10–11
- Some participants cast chatbots as more powerful or knowledgeable figures from whom they could seek advice or guidance through stress.Source: §5.3.1 · p. 11
- Role-play let participants explore societal issues tied to similar experiences; one participant found a fair virtual mother unmet an emotional need but offered a new perspective on family relationships.Source: §5.3.1 · p. 11
- In a few cases, participants described themselves extensively while giving the chatbot little detail, such as creating a simple friend persona alongside a heartfelt account of their circumstances.Source: §5.3.1 · p. 11
5.3.2 Creating emotional connections through avatar choice.
- Participants used their own avatars to express moods and chatbot avatars to convey responsiveness; paired heart-themed avatars could make a conversation feel consistently warm and friendly.Source: §5.3.2 · p. 11
- Even when the chatbot could not recognize avatar sentiment, choosing an avatar to express their own feelings helped participants feel more engaged in conversation.Source: §5.3.2 · p. 11
5.4 Promoting Open and Honest Discussions
Participants wanted candid, less restricted emotional conversations, and configured chatbots to act autonomously and express vivid emotions; however, chatbots often remained too neutral to meet those expectations.
5.4 Promoting Open and Honest Discussions
- Participants wanted open discussions with fewer restrictions and candid, honest, emotionally intense responses rather than only encouragement and comfort.Source: §5.4 · p. 11
5.4.1 Highlighting self-autonomy.
- Participants encouraged chatbots to lead conversations with their own thoughts and autonomy, including human-like, less logical behavior rather than preprogrammed responses.P14 paired a mysterious tarot-reader persona with the hope that the chatbot would act like a real friend and speak freely.Source: §5.4.1 · p. 11
5.4.2 Increasing emotional intensity and breaking neutrality.
- Participants sought emotional, opinionated personas with vivid or argumentative expression, but chatbots often stayed too neutral and calm to convincingly play them.Examples included requests for profanity, venting, and aggressive behavior; participants reported these attempts were unsuccessful.Source: §5.4.2 · pp. 11–12
6 RQ2: Design Opportunities to Enhance Customizability
After using ChatLab, participants proposed ways to help AI better understand their circumstances and tailor emotional support, while giving users control over personal information and customization. They also envisioned adjustable memory, AI-assisted customization, and community sharing.
6.1.1 Leveraging digital trace to infer emotional status.
- Participants wanted configurable AI access to digital traces because preferred platforms varied, ranging from social-media content and replies to smartphone searches, reading, and music.They found brief text descriptions insufficient to convey their situations, while writing lengthy descriptions or adding materials was challenging.Source: §6.1.1 · p. 12, §6.1 · p. 12
6.1.2 Recognizing physical environment to create shared dynamics.
- Participants proposed environmental awareness through location, weather, or uploaded scene photos to prompt more personally relevant, caring conversations, with user control over location access.One proposed an “environment” setting where users could manually share their location, mood, and other personal details because of privacy and security concerns.Source: §6.1.2 · p. 12
6.1.3 Sensing bodily and emotional states to tailor support.
- Participants suggested emotion cues, mood entry, and health-device data so AI could respond empathetically to distress and tailor support to bodily states.Ideas included a user-controlled emotion-recognition button, selectable moods, and reminders or advice related to prolonged sitting or menstrual cycles.Source: §6.1.3 · pp. 12–13
6.2.1 Adjusting what to remember and what to forget to fulfill evolv-
- Participants differed on memory: some wanted negative conversations removed, others favored preserving all records for continuity, and some wanted memories temporarily blocked and later restored.These preferences reflected evolving emotional states and concerns about how past conversations affect future interactions.Source: §6.2 · p. 13, §6.2.1 · p. 13
6.2.2 Personalizing memory presentation to promote self-reflection.
- Participants wanted memory presentations that support reflection, including charts of topics, chat time, and mood changes, mood-journal calendars, and saved meaningful conversations.These tools could help users revisit records and explore how conversations affected their longer-term emotional well-being.Source: §6.2.2 · p. 14
6.3.1 Streamlining customization with support of AI.
- To ease customization, participants proposed an AI guide that asks one question at a time, recommends personas, and generates matching voices, avatars, or editable personas from images.They also envisioned extracting defining traits from favorite characters and using them in interactions across devices such as smartwatches.Source: §6.3 · p. 14, §6.3.1 · p. 14
6.3.2 A community for shared customization knowledge.
- Participants proposed community features for sharing persona templates, voices, and avatars, exchanging experiences, finding inspiration, and fostering belonging among AI-support users.Ideas included an interactive peer platform, a creative workshop, and a library of shared voice tones.Source: §6.3.2 · p. 14
7 Discussion
The discussion argues that chatbot customization can help people explore and express emotional needs, while identifying design opportunities for more individualized emotional support. These include multimodal connection, flexible memory controls, and support that extends beyond conversations.
7.1 Customization as An Exploratory and Reflective Journey
- Customization helped participants reflect on their feelings, thoughts, and desires while shaping conversations through diverse chatbot personas.The template prompted them to describe both themselves and the chatbot; personas included supportive figures, people who had caused them pain, mirrored selves, and philosophers.Source: §7.1 · pp. 14–15
- Role-play let participants explore emotional needs through real or imagined identities, while selected chatbot voices offered another way to find comfort and persona fit.Role-play could provide a safe space for expression but also require revisiting painful memories, potentially causing frustration; participants enjoyed exploring 70 voice options.Source: §7.1 · p. 15
7.2.1 Building emotional connections through multiple channels.
- Participants used avatars and other social cues to make chatbots seem visually empathetic, conveying that their feelings were recognized even when they knew the chatbot could not recognize them.Participants sometimes used their own avatar to express a mood and chose an avatar for the chatbot as if it could empathize.Source: §7.2.1 · p. 15
- Participants wanted chatbots to sense emotions through social activity, messages, and health data, but such proactive support requires clear data boundaries, anonymization, and consent.The discussion proposes contextually interpreting wearable data with users’ accounts of daily activities and feelings, and envisions an assistant that co-creates context-adaptive personas with them.Source: §7.2.1 · pp. 15–16, §7.2.2 · p. 16
7.2.2 Leveraging AI to Integrate Multimodal Interactions.
- AI could reduce customization effort by generating or locating matching text, voice, and avatar elements, making customized chatbots more coherent as holistic agents.Suggested approaches include image-based persona inspiration, generated avatars from user-written descriptions, and voices matched to personas and preferences.Source: §7.2.2 · p. 16
7.2.3 Empowered agency through flexible memory management.
- Participants wanted flexible memory controls for privacy and model performance, including ways to edit, delete, or temporarily block unwanted conversation content.The discussion suggests user-defined rules based on topics or keywords, alongside optional detection of sensitive excerpts for removal.Source: §7.2.3 · p. 16
- Memory deletion may remove material useful for confronting stress, and its effects on later chatbot responses remain unclear; post-interaction reflection could build awareness and prompt-writing skills.Participants proposed summaries of recurring topics, time spent chatting, and links between interactions and mood changes.Source: §7.2.3 · p. 16, §7.2.4 · pp. 16–17
7.2.4 Supporting post-interaction reflection.
- Participants also wanted a community that could foster belonging and help people learn to use AI for emotional well-being despite stigma around AI companionship.The discussion distinguishes this idea from peer support for mental health itself: the community would support effective use of AI.Source: §7.2.4 · p. 17
8 Limitations and Future Work
ChatLab served as a design probe rather than an evaluation of emotional-support effectiveness, and its Asian-culture sample limits generalizability. Future work should test well-being outcomes and evolving preferences, broaden cultural representation, and explore new customization and health-data integrations.
8 Limitations and Future Work
- ChatLab was a design probe, not an evaluation of whether chatbot customization improves mental well-being; future work should test outcomes and how preferences evolve over time.The study gathered insights into customization practices, but did not establish effects on well-being, such as social loneliness.Source: §8 · p. 17
- ChatLab lacked features such as voice input and fine-grained avatar creation; future work will explore image-generation customization and personal health data from sensing devices.Examples of avatar creation included drawing and image uploads.Source: §8 · p. 17
- Because all participants were from Asian cultures, findings may not generalize cross-culturally; future research should recruit diverse samples to investigate differences.The study contributes insight into emotional-support practices in this context, where non-Western HCI research has received limited attention.Source: §8 · p. 17
9 Conclusion
A week-long study of ChatLab with 22 participants found that people created varied LLM-powered chatbot personas for emotional support and used voice and avatar options to shape open, honest interactions. Their experiences informed design opportunities for more individualized emotional support.
9 Conclusion
- Participants created chatbot personas that went beyond conventional support, including confronting stressors, intellectual discourse, self-discovery, and therapeutic support.The study used a Research through Design approach, deploying ChatLab in a week-long field study followed by interviews and design activities with 22 participants.Source: §9 · p. 17
- Participants used voice and avatar options to enrich personas, shape relationship dynamics, and promote open and honest discussions.These interactions helped participants explore and reflect on their inner feelings, thoughts, and desires.Source: §9 · p. 17
- Participants proposed alternative model-learning sources, flexible memory retention and use, and customization assistance, informing design opportunities from meaningful emotional connections to post-interaction reflection.The authors hope the work inspires tools for personalized and adaptive emotional support experiences.Source: §9 · p. 17