Think Twice: Improving Privacy Awareness with Tailored LLM-Powered Interventions

Paper Title

Think Twice: Improving Privacy Awareness with Tailored LLM-Powered Interventions

Publication Info

• Topic area: Enhancing privacy awareness in social media photo sharing using AI-driven interventions.
• Keywords: Privacy awareness, social media, large language models, privacy interventions, cognitive load, user autonomy, photo sharing, interdependent privacy, context-aware AI, privacy decision-making.

Background and Problem

• Problem / challenge: Social media users often share photos without considering potential privacy violations, especially in context-specific and subjective scenarios like memes. Existing moderation systems and generic privacy nudges lack the contextual specificity needed to address these nuanced privacy concerns effectively.
• Significance: Addressing privacy violations in photo sharing is critical to protecting individuals from emotional, relational, and reputational harm, while also respecting user autonomy in decision-making.
• Motivation and related work: Prior research has explored privacy nudges and educational interventions, but these often lack contextual depth and fail to stimulate reflective thinking. Advances in LLMs offer a new opportunity to deliver adaptive, context-aware privacy guidance that balances effectiveness with user autonomy.

Solution

• Proposed approach: Development of two LLM-powered privacy interventions—categorical and granular—that provide tailored privacy insights for social media photo sharing.
• Novelty:
  1. Introduction of context-aware, LLM-generated privacy interventions that adapt to specific photo content.
  2. Comparison of intervention types (categorical, granular, and universal) in terms of their impact on sharing behavior and cognitive load.
  3. Assessment of cognitive load metrics (intrinsic and germane) to evaluate the usability and engagement of interventions.
  4. Exploration of user perceptions and qualitative feedback to refine intervention design.
• Procedure and key techniques:
  • Generated privacy interventions using ChatGPT-4 and validated them through human review.
  • Conducted an online survey-based experiment (N = 214) with four conditions: no intervention, universal, categorical, and granular.
  • Measured sharing likelihood, intrinsic cognitive load, and germane cognitive load across intervention types.
  • Analyzed quantitative data using Kruskal-Wallis and Dunn’s post-hoc tests, and qualitative data using thematic analysis.

Results

• Concrete findings:
  • Categorical and granular interventions significantly reduced sharing likelihood for privacy-sensitive memes compared to no intervention (effect sizes: small to medium).
  • Categorical interventions imposed lower intrinsic cognitive load than granular interventions (p = 0.02).
  • Germane cognitive load was consistent across all intervention types, indicating active engagement.
• Advantage over baselines:
  • Categorical and granular interventions outperformed the universal intervention in reducing sharing likelihood for specific privacy categories (e.g., children, personal identifiable information, shaming).
  • Universal interventions were only effective for highly salient privacy issues (e.g., child-related content).
• Experiments / evaluation:
  • Dataset: 68 memes across 13 privacy violation categories.
  • Metrics: Sharing likelihood (5-point Likert scale), intrinsic and germane cognitive load (9-point Likert scale).
  • Validation: Human review and participant ratings of intervention accuracy.
• Limitations and future work:
  • Limited generalizability due to U.S.-centric participant pool and hypothetical sharing scenarios.
  • Variability in LLM performance across different architectures and prompts was not fully explored.
  • Future work should investigate multimodal interventions, cultural differences, and within-subject designs to analyze individual-level effects.

Summary

This study demonstrates the potential of LLM-powered privacy interventions to enhance privacy awareness in social media photo sharing. Categorical and granular interventions effectively reduced sharing likelihood for privacy-sensitive memes, with categorical interventions achieving this while maintaining lower cognitive load. Participants valued the autonomy and contextual relevance of these interventions, which prompted reflective decision-making. These findings suggest that LLM-based interventions can be integrated into social media platforms to support informed, autonomous user decisions. Future research should address ethical considerations, cultural diversity, and scalability to optimize the design and deployment of such tools.