Exploring the Effects of Different Chatbot Voice Identities on Self-Disclosure

Paper Title

Exploring the Effects of Different Chatbot Voice Identities on Self-Disclosure

Publication Info

• Topic area: Influence of chatbot voice identity on user self-disclosure and interaction dynamics.
• Keywords: Chatbots, self-disclosure, voice identity, own voice, family voice, stranger voice, human-computer interaction, voice synthesis, social distance, mental health.

Background and Problem

• Problem / challenge: While voice-based chatbots are increasingly used to promote self-disclosure, the impact of the chatbot's voice identity (e.g., own voice, family member's voice, or a stranger's voice) on user behavior and disclosure remains underexplored.
• Significance: Understanding how voice identity influences self-disclosure can guide the design of chatbots for mental health, self-reflection, and relational support.
• Motivation and related work: Prior research shows that chatbots can elicit more disclosure than humans due to perceived anonymity and neutrality. Voice synthesis technologies now allow chatbots to use cloned voices, which may evoke trust and familiarity. However, the relationship between voice identity and disclosure depth has not been sufficiently studied.

Solution

• Proposed approach: A mixed-method study examining how chatbot voice identities (own voice, family member’s voice, stranger’s voice) affect user impressions and self-disclosure over 14 days of daily interactions.
• Novelty:
  1. Investigates the role of voice identity in shaping self-disclosure in a longitudinal, in-the-wild setting.
  2. Explores how impressions and disclosure patterns evolve over time across different voice conditions.
  3. Provides design guidance for tailoring chatbot voices to promote specific types of self-disclosure.
• Procedure and key techniques:
  • Participants (N=61) were randomly assigned to one of three chatbot voice conditions: own voice (Condition S), family member’s voice (Condition F), or stranger’s voice (Condition O).
  • Daily 15-minute conversations were conducted for 14 days, with participants responding to progressively sensitive questions.
  • Surveys measured impressions (e.g., attractiveness, trustworthiness) on Days 1, 7, and 14.
  • Conversational data were analyzed for response length, disclosure depth, and casualness.
  • Semi-structured interviews provided qualitative insights.

Results

• Concrete findings:
  • Own-voice chatbots (Condition S) elicited the deepest and most sustained self-disclosure, with participants describing them as resembling a close friend or twin.
  • Family-voice chatbots (Condition F) prompted relational reflections and detailed interpersonal disclosures but showed topic-sensitive variation in disclosure willingness.
  • Stranger-voice chatbots (Condition O) elicited less disclosure, with some participants expressing discomfort and resistance.
• Advantage over baselines:
  • Own-voice chatbots increased perceived attractiveness and trust over time, leading to deeper introspective disclosures.
  • Family-voice chatbots facilitated relational self-reflection and comfort but were less effective for highly sensitive topics.
  • Stranger-voice chatbots were less effective overall, with declining disclosure depth over time.
• Experiments / evaluation:
  • Mixed-method approach combining quantitative analysis (e.g., response length, disclosure depth) and qualitative insights from interviews.
  • Linear mixed-effects models assessed temporal changes in impressions and disclosure.
  • Closeness to family members was analyzed as a moderator in Condition F.
• Limitations and future work:
  • Demographic imbalance and limited gender diversity in voice samples.
  • Lack of long-term memory in the chatbot, which may have influenced user perceptions.
  • Limited exploration of highly sensitive or nuanced topics.
  • Future studies should examine additional voice identities (e.g., friends, colleagues) and implement memory features.

Summary

This study demonstrates that chatbot voice identity significantly influences user impressions and self-disclosure. Own-voice chatbots fostered deep introspective disclosure and self-reflection, while family-voice chatbots encouraged relational reflections but showed topic-sensitive variation. Stranger-voice chatbots elicited less disclosure, with some users expressing discomfort. These findings highlight the importance of tailoring chatbot voices to specific user needs and contexts. The results provide actionable design guidelines for leveraging voice identity to enhance user engagement and support mental health and self-reflection. Future research should address demographic diversity, explore additional voice identities, and incorporate memory features to deepen understanding.