Relational Gains, Privacy Strains: Exploring Users’ Perceptions and Experiences with ChatGPT’s Memory Feature

Paper Title

Relational Gains, Privacy Strains: Exploring Users' Perceptions and Experiences with ChatGPT's Memory Feature

Publication Info

• Topic area: User perceptions and privacy concerns regarding ChatGPT's memory feature.
• Keywords: ChatGPT, memory feature, privacy, user perceptions, personalization, AI memory, mental models, expectancy violations, relational benefits, design implications.

Background and Problem

• Problem / challenge: Users lack understanding of ChatGPT's memory feature, leading to privacy concerns and unmet expectations. The feature's opacity and potential risks, such as data misuse and profiling, remain underexplored.
• Significance: Understanding user perceptions is critical to designing AI systems that balance personalization with privacy, ensuring user trust and satisfaction.
• Motivation and related work: Previous research highlights privacy risks in AI memorization and users' limited awareness of memory features. However, gaps remain in understanding how users perceive and experience these features, particularly in balancing relational benefits with privacy concerns.

Solution

• Proposed approach: A qualitative study using semi-structured interviews with 20 ChatGPT users to explore their perceptions, mental models, and suggestions for improving the memory feature.
• Novelty:
  1. Empirical documentation of user experiences with ChatGPT’s memory feature, focusing on relational benefits and privacy risks.
  2. Analysis of expectancy violations and their impact on users' mental models.
  3. Design recommendations to enhance transparency, user control, and privacy protection.
• Procedure and key techniques:
  • Conducted interviews with participants who were aware of ChatGPT's memory feature.
  • Asked participants to interact with the feature and reflect on its operation.
  • Analyzed data using open coding to identify themes related to perceptions, privacy concerns, and design suggestions.

Results

• Concrete findings:
  • ChatGPT's memory was perceived as unforgetful, detailed, accurate, and lacking emotions, distinguishing it from human memory.
  • Users valued relational benefits, such as personalization and a sense of belonging, but were concerned about privacy risks and outdated or irrelevant memories.
  • Most participants experienced negative expectancy violations, such as unexpected data retention and limited transparency.
• Advantage over baselines: Provides user-centered insights into the interplay between personalization and privacy in AI memory, addressing gaps in prior research.
• Experiments / evaluation: Semi-structured interviews with 20 participants, focusing on their interactions with ChatGPT's memory feature and subsequent reflections.
• Limitations and future work:
  • Findings may not generalize to other LLMs or cultural contexts.
  • Study captures perceptions at a specific point in time; longitudinal studies are needed.
  • Future work should explore specific types of AI memory and reduce anthropomorphic language to avoid inflated perceptions of AI capabilities.

Summary

This study investigates users' perceptions of ChatGPT's memory feature, highlighting its machine-like characteristics (unforgetfulness, detailedness, accuracy, lack of emotions) and relational benefits. While users appreciated personalization, they expressed concerns about privacy risks, outdated memories, and limited transparency. Negative expectancy violations were common, emphasizing the need for better communication of memory mechanisms. The study provides actionable design suggestions, such as increasing transparency, enhancing user control, and balancing relational cues with privacy safeguards, to create a more user-aligned and responsible memory experience.