Does AI based synthetic travel behavior data replicate human cognitive biases?

Term Start:

September 16, 2026

Term End:

September 15, 2027

Budget:

$150,000

Keywords:

AI/LLM, Cognitive bias, Synthetic data

Thrust Area(s):

Data Collection Mechanisms, Data Modeling and Analytic Tools

University Lead:

University of Washington

Researcher(s):

Don MacKenzie

This project will develop methods to test for cognitive biases in AI-generated synthetic data used as substitutes for human survey responses in travel behavior analyses. Moving beyond recent benchmarking and prompt engineering work, this project will ask: even when synthetic data look statistically reasonable, do they reproduce canonical cognitive biases that influence travel choices? In other words, we test whether synthetic data are behaviorally realistic, not merely statistically plausible. We will implement a compact, transportation-specific battery of experimentally controlled stated-choice and judgment tasks to test for biases such as status quo bias, reference dependence and loss aversion around time/cost changes, framing, anchoring, and zero-price effects in toll/parking/fare presentation, and ambiguity aversion in travel-time reliability. We will field this battery in parallel to (i) a commercial synthetic panel product (e.g., Qualtrics) and (ii) LLM-based agents under multiple prompting/persona strategies, comparing each to a modest human reference data set. The human data may comprise existing data sets or data newly collected from known human respondents, as long as question wording is consistent with the AI data collection. The direction and magnitude of any cognitive biases demonstrated in the synthetic data will be compared with those in the human reference data.

Project scope

Scroll to Top