“Don’t Look, But I Know You Do”: Norms and Observer Effects in Shared LLM Accounts

Paper Title

“Don’t Look, But I Know You Do”: Norms and Observer Effects in Shared LLM Accounts

Publication Info

• Topic area: Social dynamics and design implications of shared generative AI accounts.
• Keywords: LLM account sharing, observer effect, implicit norms, cognitive traces, privacy, appropriation, social translucence, negotiation tools, multi-user environments.

Background and Problem

• Problem / challenge: Generative AI platforms like LLMs are designed for individual use but are increasingly shared among multiple users, creating tensions around privacy, coordination, and behavioral changes due to visibility.
• Significance: Understanding shared use is critical as LLMs become common in collaborative, educational, and creative contexts, where single-user designs fail to address multi-user realities.
• Motivation and related work: Prior research on account sharing in streaming and enterprise systems highlights motivations like convenience, trust, collaboration, and cost-saving but does not address the unique challenges posed by reasoning traces in LLM environments. This study builds on theories of appropriation, implicit norms, and observer effects to explore these gaps.

Solution

• Proposed approach: A mixed-methods study combining a survey of 245 users and interviews with 36 participants to investigate shared LLM accounts, focusing on norms, observer effects, and design implications.
• Novelty:
  1. Identification of a 2×2 typology of LLM account sharing (owner participation × cost-sharing) and its implications for coordination and motivations.
  2. Analysis of how reasoning traces amplify observer-effect tensions, extending concepts of awareness and social translucence to linguistic traces.
  3. Design strategies for negotiation tools, social translucence, and graded visibility controls to adapt single-user LLM interfaces for multi-user use.
• Procedure and key techniques:
  • Conducted a survey to map sharing patterns and implicit norms.
  • Performed semi-structured interviews to explore lived experiences and observer effects.
  • Thematic coding of qualitative data to identify sharing types, norm dynamics, and behavioral changes.

Results

• Concrete findings:
  • Four types of account sharing: owned-casual, ownerless-casual, owned-split, and ownerless-split, shaped by owner participation and cost-sharing.
  • Privacy norms (63%), boundary-setting norms (24%), and access norms (13%) emerged but were fragile and inconsistently interpreted.
  • Observer effects manifested as snooping (intentional and unintentional) and self-censorship, with users alternating between observer and observed roles.
• Advantage over baselines: Unlike prior studies on streaming or enterprise systems, this work highlights reasoning traces as a unique coordination surface in LLM sharing, emphasizing cognitive sharing and intellectual vulnerability.
• Experiments / evaluation: Mixed-methods approach with survey (n=245) and interviews (n=36), analyzed through thematic coding and intercoder reliability (κ = 0.78–0.79).
• Limitations and future work:
  • Reliance on self-reports may underrepresent sensitive behaviors.
  • Findings focused primarily on ChatGPT due to its high adoption.
  • Cross-sectional design limits insights into temporal changes; future work should include longitudinal studies, cross-cultural comparisons, and behavioral log analysis.

Summary

This study explores the social dynamics of shared LLM accounts, identifying four sharing types based on owner participation and cost-sharing. Implicit norms around privacy, boundaries, and access are fragile, often violated, and shaped by reasoning traces that amplify observer effects. Users alternate between snooping and self-censorship, creating tensions that highlight unmet design needs. The paper proposes design strategies like negotiation tools, social translucence, and graded visibility controls to support multi-user environments. These findings extend CSCW concepts and offer practical insights for adapting LLM platforms to collective use while balancing privacy and collaboration.