The Impact of a Chatbot's Ephemerality-Framing on Self-Disclosure Perceptions

Problem Statement
Self-disclosure, the act of sharing thoughts and feelings, is heavily influenced by the perceived nature of the relationship between individuals. While human-computer interaction (HCI) research has traditionally focused on designing conversational agents (CAs) that build long-term, persistent relationships to foster trust, the impact of framing a chatbot's interaction as ephemeral (temporary) versus persistent remains under-explored. With large language models (LLMs) increasingly offering both "memory" features for continuity and "temporary chat" modes for privacy, there is a need to understand how these distinct temporal framings affect user comfort and willingness to self-disclose. Specifically, it is unclear whether a chatbot framing itself as a "Stranger" (forgetting past interactions) or a "Familiar" companion (remembering past interactions) better facilitates the sharing of sensitive emotional information versus low-risk factual information.

Methodology
The authors conducted a user study employing a 2 × 2 mixed factorial design to investigate the effects of Ephemerality-Framing (between-subjects: Stranger vs. Familiar) and Disclosure Type (within-subjects: Factual vs. Emotional).

• Participants: 127 US-based participants completed two chatting sessions across two days.
• Conditions:
  • Stranger Condition: The chatbot framed itself as a "friendly stranger" that would not remember the conversation after the session ended. The interface displayed no chat history.
  • Familiar Condition: The chatbot framed itself as a "friendly companion" that would remember past interactions and grow familiar with the user. The interface displayed previous chat history.
• Procedure: Participants engaged in two sessions. In one session, the chatbot elicited Factual-disclosure (hobbies, interests, preferences), and in the other, Emotional-disclosure (feelings of loneliness, social connections, relationship challenges). The order of these disclosure types was randomized.
• Measures:
  • Quantitative: Post-session surveys measured honesty, comfort, enjoyment, desire to continue, and perceived closeness (using the Inclusion of Other in the Self scale). Response length (word count) was also analyzed.
  • Qualitative: Open-ended questions explored user perceptions of comfort, intrusiveness, and the influence of the chatbot's framing on their willingness to disclose.
• Implementation: Chatbots were hosted on Qualtrics using a mix of pre-scripted utterances (to establish the framing) and GPT-4o generated responses (to acknowledge user input).

Key Results
The study yielded significant findings regarding the interaction between ephemerality framing, disclosure type, and the order of interaction:

1. Comfort in Emotional Disclosure: When Emotional-disclosure was sought in the first session, participants in the Stranger condition reported significantly higher comfort levels compared to those in the Familiar condition. The Stranger framing provided a sense of anonymity and reduced fear of judgment.
2. The Role of Order (Factual First): When Factual-disclosure was sought in the first session (followed by Emotional-disclosure in the second), the comfort gap between conditions disappeared. In this scenario, participants in the Familiar condition reported significantly higher enjoyment and a stronger desire to continue using the chatbot.
3. Qualitative Insights:
   • Stranger: Users felt this framing afforded anonymity and reduced social risk, making it suitable for sensitive topics, especially in initial interactions. Some described the experience as similar to "journaling."
   • Familiar: While seen as friendly, the Familiar framing was sometimes perceived as intrusive if sensitive topics were introduced immediately. However, when rapport was first established via low-risk Factual-disclosure, users felt more comfortable with the Familiar chatbot for subsequent Emotional-disclosure.
4. Closeness: Perceived closeness increased in the second session for all participants, but the Familiar condition showed a stronger association with closeness when rapport was built through prior interaction.
5. Beliefs in AI: Users who believed chatbots possessed higher intelligence reported greater enjoyment, desire to continue, and perceived closeness, though they tended to provide shorter responses (omitting peripheral details).

Key Contributions

• Empirical Evidence on Ephemerality: The paper provides quantitative and qualitative evidence that the "ephemerality-framing" of a chatbot significantly impacts self-disclosure perceptions, challenging the assumption that persistent memory and familiarity are always superior for engagement.
• Context-Dependent Effectiveness: The study demonstrates that the optimal framing depends on the type of disclosure and the stage of the relationship. Ephemeral framing is more effective for immediate, sensitive emotional disclosure, while persistent framing becomes more beneficial for enjoyment and continued engagement once low-risk rapport has been established.
• Design Implications for LLMs: The findings offer specific guidance for designing CAs with memory capabilities, suggesting that "temporary chat" modes may be preferable for initial sensitive interactions, whereas persistent memory is better suited for long-term relationship building after initial trust is established.

Significance and Claims
The authors claim that their work highlights the nuanced role of temporal framing in human-chatbot interactions. They argue that while personalization and familiarity are common design goals, they are not universally beneficial. Specifically, the paper posits that framing a chatbot as a transient "Stranger" can create a safer, more cathartic space for open emotional sharing by mitigating perceived social risks and privacy concerns. Conversely, a "Familiar" framing is most effective when users have already established a baseline of trust through low-risk interactions.

The study concludes that designers should consider using ephemeral framing selectively, particularly for initial or one-off encounters involving sensitive topics, while allowing chatbots to gradually transition to persistent, familiar modes as relationships develop. The authors also note that the mere acknowledgement of memory retention (or lack thereof) can prime user expectations regarding the depth and sensitivity of the conversation, suggesting that managing these expectations is critical for ethical and effective CA design.