Abstract
Modern smartphone users are frequently overwhelmed by permission requests, leading to decision fatigue and suboptimal privacy choices. This research explores whether Large Language Models (LLMs) can act as personalized assistants to automate these decisions based on natural-language privacy statements. By evaluating models like GPT-4o and Mistral against a novel dataset of over 14,000 human decisions, the study finds that LLMs can achieve up to 89% agreement with user majorities and effectively "nudge" users toward safer privacy practices.
Introduction
The constant barrage of access control prompts on modern operating systems creates a significant cognitive burden. Users often grant permissions indiscriminately just to dismiss the interruption, a phenomenon known as decision fatigue. This paper investigates a "human-in-the-loop" solution: using LLMs to understand a user's high-level privacy preferences and apply them to specific, contextual requests.
Methodology: Designing an AI Privacy Assistant
The researchers developed a framework where an LLM makes access control decisions by synthesizing three inputs: a system prompt defining its role as an expert, a user-provided natural-language privacy statement (e.g., "I only share data when essential"), and the specific context of the request (e.g., "The app FoodGuide is requesting Location to search for restaurants").
The system relies on in-context learning, avoiding the need for computationally expensive fine-tuning. This approach allows the model to remain flexible and easily adaptable to individual user preferences expressed in plain English.
Data Collection
To evaluate the system, the authors compiled a unique dataset consisting of:
- 307 Privacy Statements: Collected from participants via static forms and interactive chatbots.
- 14,682 Decisions: Human responses to 111 distinct permission tasks across 14 popular Android apps.
- User Feedback: Over 1,400 ratings of the LLM's decisions and justifications.
Key Insights: Reasoning and the Privacy Paradox
A core finding of the study is the power of LLM justifications. When the model disagreed with a user, its explanation often convinced the user to change their mind. This was particularly effective for "Deny" recommendations, where 59% of users opted to follow the LLM's safer suggestion.
The study also highlighted a "privacy paradox": users often write strict privacy statements but act more permissively in practice. In these instances, the LLM tended to follow the written statement more faithfully than the user followed their own stated rules, acting as a consistent enforcement mechanism. Furthermore, the researchers found that LLM confidence scores (token probabilities) correlate with user consensus, providing a reliable metric for when the system should defer to the human user.
Quantitative Results
- High Alignment: Generic LLMs agreed with the majority of users in 75% to 89% of cases.
- Personalization Impact: Incorporating individual privacy statements improved agreement by 2% to 5% on average, with some users seeing improvements as high as 50%.
- Safety and Automation: By applying confidence thresholds, a personalized GPT-4o model could autonomously handle 64% of requests without making any security-critical errors.
- Expert Alignment: In clear-cut security cases, advanced models like GPT-4o correctly denied access 100% of the time, surpassing the accuracy of the user majority.
Limitations and Future Work
While promising, the study was conducted in a controlled, hypothetical environment. Real-world behavior may differ when actual personal data is at risk. Additionally, the research focused on English-speaking populations in the US and UK, leaving cultural variations in privacy norms for future study. The authors also noted the potential for adversarial "prompt injection" attacks by malicious apps and suggested that future work should explore more complex agentic environments.