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
The paper asks how individuals can customize LLM-powered chatbots to address varied emotional-support needs, an area left underexplored by prior work. Through ChatLab, a Research through Design prototype, the authors studied 22 participants in a week-long field study followed by interviews and design activities. Participants created diverse personas and used voices and avatars to shape relationship dynamics and encourage open, honest discussion, while also proposing further customization opportunities.
Problem
Prior work left underexplored how individuals use LLM customizability to meet unique emotional-support needs, particularly beyond language outputs.
Method
The authors used Research through Design to deploy ChatLab in a week-long field study with 22 participants, followed by interviews and design activities.
Results
Participants created diverse personas for emotional reliance, stress confrontation, intellectual discourse, self-discovery, and therapeutic support, using voices and avatars to shape interactions and promote open, honest discussions.
Takeaways & Limitations
Personalized emotional-support chatbots can be designed around users’ varied needs, relationship dynamics, and preferences for more open interaction.
Takeaways & Limitations
The study did not evaluate whether chatbot customizability improves mental well-being, and its participants were all from Asian cultures.
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 paper examines how people can customize LLM-powered chatbots for varied emotional-support needs, addressing limited understanding of individualized customization beyond language outputs. Using ChatLab and Research through Design, it studies how participants constructed personas and shaped interactions to support open, honest conversations.
- Emotional support needs vary from confiding and empathy to companionship and practical guidance, motivating personalized chatbot support.
- Prompt-based LLM customization lowers barriers for non-experts to specify chatbot personality, communication style, and domain knowledge.
- Prior research underexplored how individuals use LLM customizability to meet unique needs and how voices and avatars augment emotional-support experiences.
- Using Research through Design, the authors developed ChatLab as a probe for constructing and interacting with purpose-specific customized chatbots.
- Participants created personas for emotional reliance, stress confrontation, intellectual discourse, self-discovery, and therapeutic support, using voices and avatars to shape relationship dynamics and promote openness.
- The study contributes empirical insight into everyday LLM-chatbot customization and design implications for emotional-support tools tailored to individualized needs.
- The paper treats emotional support as broader and less formal than mental-health intervention, without endorsing LLMs for severe mental-health issues or clinical advice.
2 Related Work
Prior emotional-support systems offered personalization but largely relied on limited surface-level options, leaving individualized customization needs underexplored. LLMs expand customization possibilities, motivating this study of how people shape chatbots for emotional support.
- Personalized professional care can be difficult to access because of financial, time, and location constraints, encouraging peer-support networks and self-care applications.
- Pre-LLM customizable chatbots offered limited options such as avatars and personalities, leaving users’ intricate emotional needs and customization motives underexplored.
- LLMs differ from rigid pre-LLM conversation flows through stronger language understanding, reasoning, and content-generation capabilities.
- Prior studies found LLM chatbots supporting venting, comfort, routine conversation, and lifestyle advice, while highlighting tensions between user agency and therapeutic growth.
- This paper examines customization practices alongside participants’ emotional states, interaction experiences, and expectations for LLMs in everyday life.
3 ChatLab: A Prototype for Research Exploration
ChatLab was designed as a Research through Design probe to let people create and interact with customizable LLM chatbots for emotional support. Its interface combined prompt templates with adjustable modalities, avatars, models, temperature, conversation history, and diary-based reflection.
- ChatLab served as a medium for eliciting experiences and design requirements around customizable emotional-support chatbots, rather than as an effectiveness evaluation.
- The design expanded customization beyond text because visual appearance and voice can complement verbal persona cues and enrich conversational engagement.
- ChatLab encouraged repeated customization because desired conversation partners and communication styles may change with users’ feelings and events.
- The prototype supported chatbot customization, additional interaction settings, chatting, conversation history, onboarding, FAQs, and an experience diary.
- A structured template helped users describe their backgrounds and chatbot expectations, while an LLM organized their inputs into coherent prompts without prescribing specific customization behaviors.
- Users could customize output modality, avatars, voices, GPT models, and temperature, including a library of 70 Mandarin voices.
- The system used GPT models, text-to-speech APIs, and secure Firebase storage to support flexible chatbot interaction.
4 Method
The study combined a pilot, tutorial, week-long ChatLab field deployment, interviews, and design activities, analyzing participants’ chatbot configurations and qualitative accounts. The final analysis included 22 participants and used thematic analysis of interviews, artifacts, logs, and diaries.
- The formal study comprised an online tutorial, field study, post-study interviews, and design activities after an initial feasibility pilot.
- Six pilot participants used ChatLab daily for one week and then joined a two-hour focus group, leading to usability improvements before the formal study.
- 149 people completed screening, 30 met inclusion criteria, 26 completed the field study, and 22 participants entered the final analysis.
- Participants were adults experiencing emotional struggles and moderate or high social loneliness, while the tutorial cautioned against seeking medical diagnoses or clinical advice.
- During seven to ten days, participants customized one or more chatbots according to their moods and emotional needs, then used them for sharing, expression, or questions.
- Post-study sessions combined experience discussions with think-aloud design activities in which participants envisioned ideal emotional-support conversation partners.
- Researchers analyzed customization settings, conversation statistics, interviews, designed artifacts, logs, and diary entries using bottom-up thematic analysis.
5 RQ1. Customization Practice
Participants actively engaged with ChatLab to construct and interact with customized emotional-support chatbots, using varied settings and personas.
- 1,541 conversation rounds were recorded across 178 sessions, averaging eight sessions per participant and lasting approximately 4–37 minutes each.Participants created 118 distinctive personas, ranging from 1 to 13 personas per participant.
5.1 Persona Construction and Customization Experience
Participants constructed multiple chatbot personas for distinct emotional and reflective purposes, adapting roles to different contexts and needs.
- Participants created personas for emotional reliance, confronting stressors, intellectual discourse, self-discovery, and therapeutic support.
- Most participants used multiple personas, while some maintained one persona through incremental updates such as new names or personality traits.
- Participants tailored chatbot roles to contexts such as academic pressure or loneliness, producing different language and emotional support for each role.
5.2 Enriching The Constructed Persona
Participants enriched chatbot personas with voices and avatars that matched intended roles, conveyed emotional qualities, and shaped how supportive interactions felt.
- Participants incorporated voices, avatars, and other social cues to enrich chatbot personas.
- Participants commonly selected voices and avatars that matched the chatbot persona, such as a calm mature voice for an intellectual figure.
- Supportive personas commonly received comforting voices, familiar accents, or cheerful avatars that participants associated with warmth, relaxation, and easier self-expression.
- Participants viewed avatars as subtle cues of kindness or positivity, while acknowledging that changing an avatar might not directly change their mood.
- For stressor personas, participants did not specifically align voices or avatars with the stressors, possibly because they avoided visual and voice confrontation.
5.3 Shaping Conversation Dynamics
Participants shaped conversation dynamics by specifying identities, relationships, and emotional signals for themselves and the chatbot, supporting role-play and more personally responsive exchanges.
- Participants specified their own identities and chatbot personas to shape the relationship dynamics of conversations.
- They combined authentic and fictional anecdotes to prepare both sides for anticipated conversations and define their interaction context.
- Role-play enabled participants to explore societal experiences and, even when the chatbot failed to meet an emotional need, gain new perspectives on personal growth.
- Some participants described their own circumstances extensively while providing minimal chatbot characterization, foregrounding their personal emotional context.
- Participants used paired avatars and self-selected emotional avatars to express feelings and elicit conversational responses, even when the chatbot could not interpret the sentiment.
- Choosing an avatar for themselves could increase conversational engagement through nuanced, non-verbal self-expression.
5.4 Promoting Open and Honest Discussions
Participants customized chatbots to support open, candid, and emotionally intense discussions rather than merely receiving polite encouragement. They sought personas with autonomy, opinions, and vivid emotional expression, though chatbots often remained too neutral.
- Participants sought candid, honest, and emotionally intense responses when confronting frustrations or inner conflicts.
- Several participants configured chatbots with independent thoughts and autonomy so they could lead conversations expressively rather than follow preprogrammed responses.
- Participants used unconventional personas, such as a mysterious tarot reader, to make interactions more human-like and less logically constrained.
- Many participants wanted chatbots to express emotions through distinctive personalities and opinions instead of maintaining an official, polite, neutral role.
- Chatbots often responded too neutrally and calmly, making attempts to create vivid, emotionally expressive personas unsuccessful.
6 RQ2: Design Opportunities to Enhance Customizability
Participants proposed expanding chatbot customization beyond prompts by letting AI learn from digital traces, environments, emotional and bodily states, and selectively managed memories. They also emphasized user control over access, retention, and customization complexity.
- 6.1 Customizing Alternative Sources for AI to Learn About Users: Participants valued sharing personal backgrounds and communication preferences but found short descriptions insufficient and lengthy manual input difficult.
- 6.1.1 Leveraging digital trace to infer emotional status: Participants wanted to customize which digital traces AI could access and how it analyzed online activity to infer emotionally relevant context.
- 6.1.1 Leveraging digital trace to infer emotional status: Suggested digital-trace applications included using social-media views as conversation topics, replying to posts, and analyzing searches, readings, and music.
- 6.1.2 Recognizing physical environment to create shared dynamics: Participants envisioned uploading current surroundings or dynamically updating location and weather as conversation backgrounds for more immersive, personally related interactions.
- 6.1.2 Recognizing physical environment to create shared dynamics: Participants emphasized that environmental awareness should provide control over whether and when AI accesses location information because of privacy and security concerns.
- 6.1.3 Sensing bodily and emotional states to tailor support: Participants wanted AI to sense emotional and bodily states, while recognizing that current technology struggles to achieve this level of emotional intelligence.
- 6.1.3 Sensing bodily and emotional states to tailor support: Proposed mechanisms included emotion-recognition controls, explicit mood inputs, and self-tracking integrations for more attentive, tailored responses.
- 6.2 Adjusting what to remember and what to forget: Participants proposed adjustable memory settings to support tailored performance, self-reflection, privacy, and changing preferences about past conversations.
7 Discussion
Participants used customization to explore their identities and emotional needs, shape relational dynamics with chatbots, and envision richer multimodal and memory features. These practices supported open conversations while raising privacy and agency considerations.
- Customization as an Exploratory and Reflective Journey: Customization helped participants reflect on their feelings, thoughts, desires, and identities while preparing chatbot conversations.
- Diverse Personas and Relationship Dynamics: Participants created multiple personas for emotional reliance, confronting stressors, intellectual discourse, self-discovery, and therapeutic support.
- Building Emotional Connections Through Multiple Channels: Participants used avatars and voices to express moods, convey responsiveness, and make conversations feel more emotionally engaging.
- Opportunities for Expanded Customization: Participants proposed proactive emotional support through social-media access, sentiment analysis, bodily sensing, and integration of personal health data.
- Privacy and Data Boundaries: Greater access to digital and physical data may improve emotional understanding but increases privacy risks, requiring boundaries, anonymization, and consent.
- Opportunities for Expanded Customization: Participants wanted assistants to guide persona creation and to generate personas from uploaded images or voices, reducing customization effort.
- Adjustable Memory and Privacy: Memory features should let users edit, delete, block, and manage conversation memories while preserving agency over privacy and support needs.
8 Limitations and Future Work
The study’s conclusions are bounded by ChatLab’s incomplete feature set and its culturally narrow sample. Future work should test well-being effects over longer periods and across more diverse populations.
- Limitations: ChatLab lacked features such as voice input and fine-grained avatar construction, so future studies should evaluate whether customization improves mental well-being over time.
- Limitations: All participants came from Asian cultures, limiting generalization to other cultural backgrounds and motivating cross-cultural research.
- Future Work: Future work will explore image-generation customization and the use of personal health data from smart sensing devices in well-being support.
9 Conclusion
Using Research through Design, the study deployed ChatLab in a week-long field study with 22 participants, followed by interviews and design activities. Participants constructed diverse personas and shaped chatbot relationships to address emotional challenges.
- The study used an RtD approach with ChatLab, involving 22 participants in a week-long field study followed by interviews and design activities.
- Participants created chatbot personas beyond conventional supportive roles, including personas for confronting stressors, intellectual discourse, and self-discovery.