Potential Use of Large Language Models (LLMs) for Travel Behavior Survey Research

Term Start:

July 1, 2025

Term End:

May 31, 2026

Budget:

$75,000

Keywords:

AI-Driven Benchmarking, Large Language Models in Surveys, Synthetic Travel Behavior Data

Thrust Area(s):

Data Collection Mechanisms, Data Modeling and Analytic Tools, Understanding User Needs

University Lead:

The University of Texas at Austin

Researcher(s):

Chandra Bhat

Traditional travel behavior surveys are resource-intensive and constrained by challenges such as small sample sizes, response bias, and high costs. With the rapid advancements in artificial intelligence, particularly LLMs, we now have the opportunity to explore novel, cost-efficient methods for creating synthetic data comparable to real-world survey results. However, the potential of LLMs to contribute to transportation research remains underexplored. This project seeks to evaluate the role of LLMs in replicating human decision-making in the context of travel behavior, investigate their strengths and limitations, and identify biases to develop methodologies for integrating them into survey-based design, data collection, and analysis workflows. This project is structured around three primary objectives:

  1. Comparative Benchmarking:
    • Assess the quality of synthetic datasets generated by LLMs across three models (e.g., GPT-4, Claude, and Mistral) by comparing their statistical properties and behavioral narratives to the 2022 NHTS dataset.
    • Identify key areas of similitude and deviation to establish benchmarking protocols.
  2. Predictive Validation:
    • Generate AI-based predictions for an upcoming stated-preference survey on Unmanned Aerial Vehicles before data collection occurs.
    • As human data collection proceeds, monitor similitude between AI-generated predictions for key behavioral questions and actual responses.
    • Compare all AI-generated predictions to human responses post-data collection to evaluate their validity.
  3. Sensitivity via Ablation Studies:
    • Explore sensitivity to detailed prompts by varying input attributes (e.g., socio-demographics, built environment, personal attitudes).
    • Develop prompting best practices for improved reliability and consistency in survey integration.

By benchmarking LLM responses against real-world data, this research aims to enhance data collection mechanisms, investigate inherent biases in AI-generated surveys, and develop best practices for integrating AI into travel behavior research. This work has the potential to illuminate the potential benefits (low-cost, customizable, and constantly evolving tool) and pitfalls (biases, unreliability) of using today’s LLMs in transportation data collection and modeling by investigating and evaluating how they complement traditional survey paradigms while ensuring transparency and accountability. The research outputs from this project will be added to the Travel Behavior Data (TBD) Hub funded through the TBD UTC and under development till 2029. A separate tab for AI-generated silicone samples, benchmarks, and data comparisons for revealed and stated preference surveys will be added to the TBD hub providing transparent access to ongoing research.

Project scope

Scroll to Top