🤖 AI

Guided Persona-based AI Surveys: Can we replicate personal mobility preferences at scale using LLMs?

Original authors: Ioannis Tzachristas, Santhanakrishnan Narayanan, Constantinos Antoniou

Published 2026-06-26

📖 5 min read🧠 Deep dive

Original authors: Ioannis Tzachristas, Santhanakrishnan Narayanan, Constantinos Antoniou

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

The Big Idea: Can AI Pretend to Be People?

Imagine you are a city planner trying to figure out how people in Germany like to get around. Do they prefer walking, driving, or taking the bus? Usually, to find this out, you have to send out thousands of paper surveys, pay people to fill them out, and wait months for the results. It's expensive, slow, and sometimes people lie or forget to answer.

This paper asks a new question: Can we use a super-smart computer brain (called a Large Language Model, or LLM) to write these surveys for us? Instead of asking real people, can we ask the AI to "pretend" to be 10,000 different people and tell us what they would do?

The researchers wanted to see if this "fake" data looks and acts just like "real" data.

The Problem with Old Methods

Think of traditional surveys like trying to paint a portrait of a crowd by asking every single person to stand still and pose. It takes forever.

The paper says that previous attempts to use AI to guess what people do were a bit clumsy. They were like a child trying to draw a crowd: they got the general idea (there are old people and young people), but they missed the details (like how a rich teenager might drive a car, while a poor teenager might take the bus). The AI didn't understand the connections between who a person is and what they do.

The New Solution: "Personas" as LEGO Sets

The researchers came up with a clever new way to teach the AI. They called it the "Guided Persona-based AI Survey."

Imagine you are building a massive LEGO city.

The Old Way: You just dump a bucket of LEGO bricks on the table and hope they look like a city. (This is the "Naive" method).
The Better Way: You sort the bricks by color and shape first, then build. (This is the "Structured" method).
The Winning Way (Guided Persona): You build specific, detailed characters first. You create a "Persona" by snapping together specific LEGO pieces:
- Age: 30–39
- Job: Full-time employee
- Money: High income
- Family: Two kids under 6
- House: A family home with a garage

Once the AI has built this specific "character" (the Persona), you ask it: "If you were this specific person, would you walk to work or drive?"

Because the AI understands the whole story of that specific person, it gives a much more realistic answer. The researchers created over 15,000 of these unique "Personas" to cover every type of person in Germany.

The Test: The "Gold Standard"

To see if their new method worked, the researchers compared their AI-generated answers against a real, massive survey called MiD 2017. Think of the MiD 2017 as the "Gold Standard" or the "Answer Key" that was created by actually asking real people.

They tested six different ways of asking the AI to generate data:

Naive AI: "Just guess what people do." (Result: Messy and inaccurate).
Structured AI: "Make sure the number of old and young people matches the real world." (Result: Better, but still missed the details).
Guided AI: "Make sure the numbers match AND the average answers match." (Result: Good, but not perfect).
Naive Persona: "Create specific characters, but guess how many of each exist." (Result: Okay, but the character mix was wrong).
Structured Persona: "Create specific characters and match the real-world mix." (Result: Better).
Guided Persona (The Winner): "Create specific characters, match the real-world mix, AND make sure the answers match the real-world statistics."

The Results: The Winner Takes All

The results were clear. The Guided Persona-based AI Survey was the only method that truly nailed it.

The Analogy: If the real survey is a photograph, the "Naive" AI method looked like a blurry sketch. The "Guided Persona" method looked like a high-definition photo that was so perfect, it was almost impossible to tell it was made by a computer.
The Numbers: The researchers used math to measure the "distance" between the AI's answers and the real answers. The Guided Persona method had a near-zero error score. It successfully captured the complex rules of life, like how having a car often depends on your age, your job, and your family size all at once.

Why This Matters (According to the Paper)

The paper concludes that this method is a game-changer for three main reasons:

Privacy: You can study how people behave without ever needing to see their real, private names or addresses. You are just studying "fake" people who act like real ones.
Cost & Speed: It's much cheaper and faster to generate 10,000 AI responses than to hire people to fill out paper surveys.
Flexibility: You can ask the AI, "What if everyone suddenly had more money?" or "What if it rained every day?" and get an instant answer based on the rules of human behavior, without waiting for a real experiment to happen.

In short: The researchers proved that if you give an AI a detailed "character sheet" (a Persona) and tell it exactly how real people behave, it can create a perfect fake survey that helps scientists understand the real world without the hassle of asking real people.

Technical Summary: Guided Persona-based AI Surveys

Problem Statement

Traditional survey methods for transportation systems and social science research face significant hurdles, including high operational costs, inefficiencies in data collection, and limited scalability. While existing synthetic data generation techniques have focused on demographic simulation or imputation, they often fail to capture the intricate dependencies between sociodemographic attributes (e.g., age, education, economic status) and specific behavioral preferences (e.g., mobility choices). Consequently, these methods struggle to replicate complex, real-world correlations without explicit pre-definition, limiting their utility for high-fidelity scenario planning.

Methodology

This study proposes a novel Guided Persona-based AI Survey methodology to generate synthetic mobility data. The approach leverages Large Language Models (LLMs), specifically the GPT-4o API, to create artificial survey responses that align with real-world benchmarks.

1. Dataset and Benchmarking
The study utilizes the MiD 2017 (Mobility in Germany 2017) dataset as the ground truth. This comprehensive national survey includes detailed demographic distributions, transport modes, and trip frequencies. The dataset was normalized, merging "not specified" responses into the most popular categories to ensure consistency.

2. Persona Definition
A "Persona" is formally defined as a tuple $P(A, E, M, S, H)$ representing a unique combination of five attributes:

Age Group ( $A$ )
Education Level ( $E$ )
Main Activity ( $M$ )
Economic Status ( $S$ )
Household Type ( $H$ )

The population percentage of a specific Persona ( $\pi_P$ ) is calculated using conditional probabilities to respect interdependencies among attributes (e.g., the probability of a specific economic status given age, education, and activity).

3. Comparative Framework
The study evaluates six distinct survey generation methods against the MiD 2017 benchmark:

Naive AI Survey: Generates a population based on general knowledge without alignment to real-world benchmarks.
Structured AI Survey: Aligns the synthetic population structure with MiD 2017 demographics but does not explicitly align response distributions.
Guided AI Survey: Adds constraints to both population structure and response averages based on MiD 2017 statistics.
Naive Persona-based AI Survey: Defines all possible attribute combinations (Personas) and simulates responses using general knowledge.
Structured Persona-based AI Survey: Aligns Persona densities with MiD 2017 demographic data.
Guided Persona-based AI Survey: Builds upon the Structured Persona method by incorporating expected response statistics from MiD 2017 into the LLM prompts.

Prompts were tailored for each method, ranging from general instructions to advanced directives explicitly instructing the LLM to maintain realistic correlations between attributes based on the benchmark data.

Key Contributions

Novel Methodology: Introduction of the "Guided Persona-based" approach, which systematically integrates sociodemographic and behavioral attributes to generate synthetic responses, a technique not previously applied to travel surveys.
Scalable Synthetic Data: Demonstration of a cost-effective, privacy-preserving alternative to traditional surveys that can generate large-scale datasets (e.g., 15,840 unique Personas) encompassing diverse scenarios.
Capturing Complex Dependencies: The study shows that LLMs, when guided by specific personas and real-world constraints, can effectively model the non-linear dependencies between demographic factors and mobility preferences, overcoming limitations of traditional imputation techniques.

Results

The performance of the methods was evaluated using statistical alignment metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Jensen-Shannon (JS) Distance, Entropy, Conditional Entropy, and Cramér's V.

Superior Performance of Guided Persona-based AI: This method achieved near-perfect alignment with real survey data, recording an exceptionally low MAE of 0.03, RMSE of 0.17, and JS Distance of 0.0016. It achieved a Cramér's V of 1.00, indicating perfect association strength between variables compared to the real data.
Performance of Guided AI Survey: This method also performed robustly (MAE: 3.37, RMSE: 4.65) but was less precise than the Persona-based variant, suggesting that explicit Persona structuring enhances accuracy.
Limitations of Other Methods: Naive and Structured AI surveys (both non-Persona and Persona-based without guidance) exhibited significantly higher error metrics (e.g., Naive AI Survey MAE: 10.53) and struggled to capture complex demographic-response relationships. The Structured Persona-based method showed moderate performance but lacked the precision of the guided approach.

Significance and Claims

The paper claims that the Guided Persona-based AI Survey represents a transformative advancement in survey methodology. By combining demographic alignment with response constraints, it enables the creation of high-fidelity synthetic datasets that are scalable, cost-efficient, and privacy-preserving.

The authors assert that this approach offers valuable opportunities for transportation planning and social science research, allowing researchers to:

Analyze population-level trends without accessing sensitive individual data.
Simulate "what-if" scenarios and explore hypothetical questions not present in original surveys.
Generate data that accurately reflects complex interdependencies between demographic attributes and travel behaviors.

The study concludes that while LLMs offer powerful capabilities for synthetic data generation, the specific integration of "Personas" with "Guided" constraints is critical for achieving high-fidelity results that match observed real-world patterns. Future work is suggested to explore hybrid models for computational efficiency and to expand applications to other transportation modes and datasets.