💻 computer science

Text-Based Personas for Simulating User Privacy Decisions

The paper introduces Narriva, a method that generates text-based synthetic privacy personas grounded in historical user decisions and structured by privacy theories, achieving high predictive accuracy and significant token efficiency while effectively reproducing both individual and population-level privacy behaviors across diverse datasets.

Original authors: Kassem Fawaz, Ren Yi, Octavian Suciu, Rishabh Khandelwal, Hamza Harkous, Nina Taft, Marco Gruteser

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kassem Fawaz, Ren Yi, Octavian Suciu, Rishabh Khandelwal, Hamza Harkous, Nina Taft, Marco Gruteser

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a digital bodyguard for millions of people. This bodyguard's job is to make privacy decisions on behalf of its user: "Should I let this app see my location?" "Is it okay to share my health data with this research study?"

The problem? Asking every single human to answer these questions every time a new app pops up is exhausting, expensive, and slow. It's like asking a million people to fill out a 100-page form every time they walk through a door.

This paper introduces Narriva, a clever new way to solve this. Instead of asking humans to fill out forms constantly, Narriva creates a text-based "digital twin" (a persona) for each person. This twin is a short, readable summary of who they are and how they think about privacy.

Here is the story of how Narriva works, explained through simple analogies:

1. The Problem: The "Raw Data" vs. The "Resume"

Imagine you want to hire a chef.

The Old Way (Raw Data): You hand the hiring manager a 500-page notebook containing every single meal the chef has ever cooked, every ingredient they've touched, and every time they burned a toast. It's accurate, but it's a nightmare to read, takes forever to process, and is impossible to carry around.
The Narriva Way (The Persona): Instead, you ask an expert to read that 500-page notebook and write a one-page resume. This resume summarizes the chef's style: "Loves spicy food, hates wasting ingredients, always checks the oven twice."

Narriva does exactly this for privacy. It takes a user's massive history of privacy choices (the 500-page notebook) and compresses it into a tiny, human-readable "Privacy Resume" (the persona).

Result: It shrinks the data by 80–95%. It's like turning a library into a single, perfect bookmark.

2. The Secret Sauce: "Practice Makes Perfect"

You might think, "Just write a summary, and we're done." But the authors realized that a simple summary isn't enough. Sometimes the summary gets the chef's style wrong.

So, Narriva uses a feedback loop, like a coach training an athlete:

Draft: It creates a draft persona based on the user's past answers.
Test: It asks the AI, "If this persona were the user, how would they answer this new question?"
Compare: It checks the answer against what the real user actually said in the past.
Coach: If the AI got it wrong, a "coach" (another AI) tells the persona: "Hey, you said you hate sharing location, but you agreed to share it for a pizza coupon. Fix your summary to explain that exception!"
Repeat: It does this a few times until the persona is a perfect match.

3. The Big Discovery: Actions Speak Louder Than Words

The researchers tested this on five different groups of people and found something fascinating, which they call the "Privacy Paradox."

The "Talkers": If you ask people, "Do you care about privacy?" they say, "Yes, I'm very careful!"
The "Doers": But when asked, "Will you share your location for a free app?" many say, "Sure, why not?"

The paper found that past actions are the best predictor of future actions.

If you build a persona based on what people said they believe (attitudes), the AI guesses wrong.
If you build a persona based on what people actually did (behaviors), the AI is 88% accurate.

Analogy: It's like predicting if someone will run a marathon.

Asking them, "Do you love running?" might get a "Yes!"
But looking at their past running logs (did they actually run 5 miles last week?) tells you the truth. Narriva trusts the logs, not the talk.

4. Why This Matters: The "Traveling Passport"

The most exciting part of Narriva is that these personas are portable.

Imagine you have a "Privacy Passport" that summarizes your privacy preferences.

Today, you use it to decide if a social media app can see your photos.
Tomorrow, you can take that same passport to a new smart home device or a health app. The device reads your passport and instantly knows, "Ah, this person is okay with sharing health data for research but hates sharing their location with advertisers."

This solves the "Cold Start Problem." Usually, when you join a new app, you have to fill out boring settings or the app has to guess. With Narriva, your "Privacy Passport" travels with you, so the app knows you immediately without you having to explain yourself again.

5. The Limits: People Change

The paper also notes that these personas aren't magic crystals. They are snapshots in time.

If you create a persona in 2014, it might not work well in 2024 because the world has changed (new apps, new fears, new laws).
However, the study showed that personas can travel across different studies and time periods surprisingly well, as long as the gap isn't too huge (like 9 years).

Summary

Narriva is a system that turns a person's messy, long history of privacy choices into a short, smart, and accurate "Privacy Resume."

It's efficient: It cuts down data size by 90%.
It's accurate: It learns from what people do, not just what they say.
It's portable: It allows users to carry their privacy preferences from one app to another, making the internet feel less like a minefield and more like a place where your personal rules are respected automatically.

In short, it's like giving every user a personal privacy bodyguard that knows exactly who they are, so they don't have to keep explaining themselves to every new app they meet.

1. Problem Statement

Simulating human privacy decisions is critical for aligning autonomous agents with user intent and conducting scalable, cost-effective privacy studies. However, existing approaches face a trade-off between four competing dimensions:

Predictiveness: Accuracy in simulating individual decisions on novel questions.
Usability/Transparency: Human-readability and editability of the simulation model.
Scalability: Maintaining bounded context lengths for efficient inference.
Representation: Authentically capturing population diversity without relying on generic LLM stereotypes.

Current methods fail to balance these:

Raw History/Logs: High predictiveness but poor scalability (growing context) and usability (unstructured, hard to audit).
Demographic Stereotypes: High scalability but low predictiveness and representation (biased, generic).
Natural Language Statements: High usability but low predictiveness in novel scenarios.

The core challenge is to generate text-based synthetic privacy personas that are concise, grounded in real data, and capable of accurately simulating both individual and population-level privacy behaviors.

2. Methodology: The Narriva Framework

The authors propose Narriva (Narrative Personas for Privacy), a framework that generates structured, text-based personas by compressing raw user data into concise narratives grounded in privacy theories.

Core Components

Data Grounding: Unlike methods relying on demographics, Narriva grounds persona generation in large-scale, real-world survey datasets (e.g., Pew Research Center, CHI, PETS).
Theoretical Templates: Personas are not free-form; they are structured using established privacy decision-making frameworks:
- Basic: Chain-of-thought reasoning.
- Privacy Calculus: Rational actors weighing risks vs. benefits.
- Bounded Rationality: Decisions influenced by cognitive biases and heuristics.
- Protection Motivation Theory (PMT): Decisions driven by threat appraisal and coping mechanisms.
Iterative Optimization Loop:
- Generation: An LLM generates an initial persona from a subset of a user's historical responses ( $r_{gen}$ ) using a specific template.
- Evaluation: A prediction model tests the persona against the same historical data to calculate accuracy.
- Feedback & Correction: A feedback model analyzes errors and generates corrections (focusing on conciseness, generalization, and predictiveness).
- Refinement: The process iterates (typically 2-3 times) to refine the persona until it consistently predicts the user's historical decisions.
Prediction: The optimized persona is used to simulate responses to unseen questions ( $q_{eval}$ ) by prompting an LLM with the persona narrative and the new question.

3. Key Contributions

Narriva Framework: A novel system that compresses high-dimensional user data into human-readable, theory-driven text personas.
Grounding in Behavior: Demonstrates that grounding personas in past behavioral decisions (rather than attitudes or demographics) is essential for high-fidelity simulation.
Optimization Strategy: Introduces an iterative, lightweight optimization process that significantly improves predictive accuracy over single-pass summarization.
Cross-Study Generalization: Shows that personas synthesized from one study can accurately reproduce aggregate behaviors in independent studies, provided the temporal gap is small.
Token Efficiency: Achieves massive reductions in prompt size (80–95%) compared to raw in-context learning while maintaining or improving accuracy.

4. Experimental Results

The framework was evaluated across five diverse datasets (SPA, Pew PP1, Pew W49, Pew W127, CAuthN) involving over 10,000 respondents.

Key Findings

High Predictive Accuracy:
- Narriva achieves up to 88% individual-level accuracy on dense datasets (e.g., SPA), significantly outperforming non-personalized LLM baselines.
- It matches the accuracy of using raw historical examples but with far fewer tokens.
Token Reduction:
- Narriva reduces prompt token counts by 82% to 95% compared to raw examples (e.g., compressing ~8,900 tokens to ~430 tokens in the SPA dataset).
Population-Level Fidelity:
- Personas accurately reconstruct population distributions, achieving a TVComplement (1-TVD) of up to 0.85 (where 1.0 is perfect match).
- They effectively smooth out individual noise while preserving aggregate trends.
Behavior vs. Attitude:
- Behavioral questions are strong predictors of future decisions.
- Attitude questions (general concerns) are poor predictors of concrete behavioral choices, highlighting the "privacy paradox."
Template Selection:
- No single privacy theory (e.g., Calculus vs. Bounded Rationality) dominates globally.
- Dynamic selection of the best-fitting template per user (via a calibration step) boosts individual accuracy by ~5–10% compared to a static "Basic" template.
Temporal Limits:
- Personas generalize well across studies separated by a few years (e.g., 2019 to 2023).
- Performance degrades significantly over long temporal gaps (e.g., 2014 to 2023) due to shifts in technology and privacy norms.

5. Significance and Implications

Scalable Privacy Research: Enables researchers to run large-scale "synthetic panels" to test privacy policies, UI designs, and experimental conditions without the cost and fatigue of human participants.
Personalized Privacy Assistants: Solves the "cold start" problem for AI privacy assistants. A user's validated text persona can be ported across platforms to immediately enable personalized decision-making without re-collecting raw history.
Auditability: By converting raw logs into structured, theory-based narratives, the system makes user preferences transparent and editable, fostering trust in autonomous agents.
Methodological Shift: Moves the field away from relying on demographic stereotypes or unstructured logs toward behavior-grounded, theory-informed synthetic data.

Conclusion

Narriva demonstrates that it is possible to balance the competing demands of predictiveness, usability, scalability, and representation in privacy simulation. By compressing real-world behavioral data into concise, optimized text personas, the system achieves high-fidelity simulations that are both computationally efficient and interpretable, laying the groundwork for next-generation personalized privacy tools.