Privacy Control in Conversational LLM Platforms: A Walkthrough Study

Paper Title

Privacy Control in Conversational LLM Platforms: A Walkthrough Study

Publication Info

• Topic area: Privacy control mechanisms in conversational large language model (LLM) platforms.
• Keywords: Privacy control, conversational LLMs, data governance, user data management, natural language control, memory features, shared data, customization, user-centered design.

Background and Problem

• Problem / challenge: Existing research on conversational LLMs has focused on privacy risks but lacks an understanding of how platforms implement user-facing privacy controls for managing data such as accessing, editing, deleting, and sharing.
• Significance: Privacy control is critical as conversational LLMs process sensitive user data through natural language inputs, raising concerns about data ownership, retention, and sharing.
• Motivation and related work: Prior studies have explored privacy risks, perceptions, and mental models in conversational LLMs but have not comprehensively analyzed interface-level privacy controls. This study aims to address this gap by examining six widely used platforms.

Solution

• Proposed approach: An expert-driven application walkthrough of six conversational LLM platforms to analyze privacy control mechanisms, focusing on interface-level features.
• Novelty:
  1. Identification of emerging paradigms in privacy control specific to conversational LLM platforms.
  2. Analysis of interaction-derived data units such as memory snippets and customized objects.
  3. Exploration of natural language-based privacy controls and shared data governance.
  4. Empirical insights for designing scalable, user-friendly privacy controls.
• Procedure and key techniques:
  • Conducted walkthroughs of six platforms: Character.ai, ChatGPT, Claude, Gemini, Meta AI, and Pi.
  • Analyzed governance-related information from institutional materials (e.g., Privacy Policies).
  • Examined technical features for controlling data such as chat history, memory, and customized objects.
  • Documented control execution methods, including graphical user interfaces (GUIs) and natural language (NL) commands.

Results

• Concrete findings:
  • Platforms offer controls for chat history, memory, and customized objects, with varying granularity and execution methods.
  • Natural language control enables intuitive interaction but introduces ambiguity in user intent and system interpretation.
  • Shared data mechanisms create overlapping ownership between sharers and sharees, complicating governance.
• Advantage over baselines:
  • Identified unique privacy control paradigms tailored to conversational LLMs, which differ from traditional platforms with static data fields.
  • Highlighted innovative features like memory snippets and NL-based control that enhance user interaction.
• Experiments / evaluation:
  • Walkthroughs covered six platforms, analyzing features such as memory portals, customization interfaces, and shared data controls.
  • Data governance practices were compared using a revised privacy policy annotation scheme.
• Limitations and future work:
  • Temporal scope limits findings to features available during the study period (Nov 2024–Jan 2025).
  • Focused on U.S.-based platforms; cross-cultural analysis of LLMs in different regulatory contexts is needed.
  • Future research should include user studies to examine real-world interactions with privacy controls.

Summary

This study analyzed privacy control mechanisms across six conversational LLM platforms, uncovering unique features such as interaction-derived data units (e.g., memory snippets, customized objects), natural language-based control, and shared data governance. The findings highlight challenges in designing scalable and user-friendly privacy controls for conversational interactions. Empirical insights from this study provide a foundation for improving transparency, usability, and multi-user governance in LLM-powered platforms. Future work should explore cross-cultural comparisons and user-centered evaluations to further refine privacy control designs.