💬 NLP

Reproducing human biases in route choice using large language models: Toward scalable behavioral modeling

This paper demonstrates that large language models can effectively reproduce human route choice biases and prospect-theoretic decision behaviors without explicit parameter specification, offering a scalable alternative to traditional methods for large-scale behavioral modeling.

Original authors: Jiangtao Han, Shoufeng Ma, Shuxian Xu, Geng Li, Shuai Ling, Ning Jia, Zhengbing He

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiangtao Han, Shoufeng Ma, Shuxian Xu, Geng Li, Shuai Ling, Ning Jia, Zhengbing He

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 figure out why people make weird choices when they are in a rush. Sometimes, a driver will pick a route that is usually faster but has a tiny chance of a massive traffic jam, just because they are in a hurry. Other times, they pick a boring, slow route just to be safe. Scientists call this "human bias," and for a long time, they've used a complex math recipe called Cumulative Prospect Theory (CPT) to describe it. This recipe has special ingredients (numbers) that tell us how much people fear losing time versus how much they love saving it.

But here's the problem: getting those special numbers usually requires asking thousands of real humans to fill out boring surveys or play fake games in a lab. It's slow, expensive, and hard to get a truly diverse crowd.

Enter the Large Language Model (LLM). Think of an LLM not as a calculator, but as a super-smart, super-creative actor who has read every book, story, and forum post ever written. The researchers in this paper asked a big question: Can we just ask this digital actor to "pretend" to be a human traveler, and will it naturally make the same weird, biased choices that real people do?

The Big Experiment: A Virtual Traffic Jam

To find out, the researchers built a giant virtual world. They didn't just ask one AI; they created 3,000 unique digital travelers. Each one had a specific "personality card" (like a profile). Some were cautious retirees, some were stressed parents, some were spontaneous college students, and some were busy couriers.

They then dropped these 3,000 agents into a simulated road network with two choices:

Route A: A steady, reliable path.
Route B: A path that might be super fast, but also has a chance of a huge delay (like a traffic accident or a road closure).

The researchers set up two different "moods" for the experiment:

The "Gain" Mood: Everyone is trying to arrive early. If they get there early, they get a "reward."
The "Loss" Mood: Everyone is trying to avoid being late. If they are late, they suffer a "loss."

What Happened? (The Magic and the Math)

The results were surprisingly human. The digital actors didn't just pick the mathematically "perfect" route. Instead, they showed the exact same biases that real humans do:

When things were going well (Gains): The cautious agents (like the retirees) stuck to the safe route. They didn't want to risk a delay.
When things were going badly (Losses): The same cautious agents started taking risks! When they knew they were going to be late anyway, they switched to the risky, fast route, hoping to squeeze out a few minutes to save face.

This is exactly what the CPT math predicts: We hate losing more than we love winning, and when we are already losing, we become gamblers.

The researchers then took the choices made by these 3,000 digital agents and ran them through the CPT math recipe. They calculated the "special numbers" (parameters) that best fit the AI's behavior.

They found a loss aversion number of 1.43. This means that in these simulations, the agents felt the pain of being late 1.43 times more intensely than the joy of arriving early.
They found that the agents were less sensitive to small changes in time (a property called diminishing sensitivity), just like real people.

Did the AI Get it Right?

To check if their new "AI method" was any good, they compared the numbers the AI generated against numbers found in real-world studies with actual humans.

The AI's numbers predicted human choices with an error rate of only 0.0036 in a complex three-route test.
This was actually better (lower error) than some of the classic numbers used by scientists for years, and much better than using numbers from totally different fields (like gambling).

What This Means (and What It Doesn't)

The paper suggests that we might not need to interview thousands of tired commuters anymore. Instead, we can use these digital actors to generate massive amounts of behavioral data quickly and cheaply. It suggests that AI can "learn" human irrationality just by reading about it and role-playing.

However, the authors are careful. They didn't prove that AI is human. They showed that in these simulations, the AI behaves in a way that matches human patterns. They also noted that their digital agents are currently "independent thinkers" who don't talk to each other or learn from traffic jams in real-time.

So, while we haven't solved the mystery of the human mind, this study suggests that if you want to model how a crowd of people might act in a traffic jam, you might just need to ask a very well-read, very creative robot to pretend to be them. And surprisingly, the robot might just do a better job than a spreadsheet ever could.

Technical Summary: Reproducing Human Biases in Route Choice Using Large Language Models

Problem Statement

Human decision-making, particularly in route choice, frequently deviates from the assumption of full rationality, exhibiting systematic behavioral biases. Cumulative Prospect Theory (CPT) is the dominant framework for modeling these biases, capturing mechanisms such as reference dependence, loss aversion, and nonlinear probability weighting. However, the large-scale application of CPT in simulation and agent-based modeling is hindered by a critical bottleneck: the accurate estimation of CPT parameters.

Traditional methods rely on surveys and controlled experiments to calibrate these parameters. While effective for specific groups, these approaches are difficult to generalize, often fail to capture the full diversity of human decision-making across different occupations and personality types, and are constrained by limited sample sizes and high data collection costs. Consequently, there is a need for a scalable method to generate diverse behavioral data and estimate prospect-theoretic parameters without relying solely on expensive, real-world data collection.

Methodology

This paper proposes a framework for conducting large-scale virtual behavioral experiments using Large Language Models (LLMs) to simulate human route-choice behavior under uncertainty. The methodology consists of four core components:

1. Construction of LLM-based Human Agents

The study employs a nonparametric prompting method to construct heterogeneous human agents. Each agent is defined by a standardized prompting architecture comprising three modules:

Profile: Defines individual attributes (occupation, personality, risk preference, travel purpose, time pressure), behavior features (sensitivity to travel time variability, risk tolerance), and constraint conditions. The study defines 10 distinct agent profiles (e.g., retired workers, civil servants, couriers, students) to capture demographic and behavioral heterogeneity.
Planning: Guides the agent to decompose complex tasks, evaluate routes based on expected time and variability, and prioritize options consistent with their assigned identity and psychological preferences.
Action: Enforces decision rules where risk-averse agents prioritize stability and risk-seeking agents prioritize time savings. Outputs are strictly structured as JSON to ensure compatibility with statistical analysis.

2. Construction of the Travel Environment

The environment is defined by Configuration (static task objectives, route options, and risk types) and State (dynamic probabilities of disruptions like accidents or congestion and their impact on travel time). The study designs scenarios involving both Gain (arriving earlier than a reference time) and Loss (arriving later than a reference time) domains, with routes offering identical mean travel times but different risk distributions.

3. Experimental Simulation

Using OpenAI's GPT-5, the study simulates 3,000 agents across 11 distinct cases involving two-route and three-route choice scenarios. The temperature parameter is set to 1 to balance behavioral diversity with reproducibility. Each scenario is executed three times to ensure robustness.

4. Parameter Estimation and Evaluation

CPT Parameter Estimation: The study fits the LLM-generated choice data to CPT formulations to estimate the sensitivity parameters ( $\alpha$ for gains, $\beta$ for losses) and the loss-aversion coefficient ( $\lambda$ ). The reference point ( $x_0$ ) is inferred dynamically from the distribution of expected travel times generated by the agents.
Validation: The estimated parameters are used to predict route choices in new scenarios. The results are compared against:
1. LLM-simulated choices.
2. Parameters estimated from real-world behavioral experiments (e.g., Xu et al., 2011; Zhang et al., 2018).
3. Classic parameters from non-travel domains.
Metrics: Prediction accuracy is evaluated using the sum of squared residuals between predicted and observed choice probabilities.

Key Results

Reproduction of Biases: LLMs successfully reproduced systematic human biases. In gain domains, agents generally exhibited risk aversion (preferring stable routes), while in loss domains, they shifted toward risk-seeking behavior (preferring routes with higher variability but potential for earlier arrival).
Parameter Estimation: The LLM-simulated data yielded CPT parameter estimates of $\alpha = 0.4$ , $\beta = 0.64$ , and $\lambda = 1.43$ . These values indicate pronounced diminishing sensitivity and significant loss aversion (agents are 1.43 times more sensitive to losses than equivalent gains).
Predictive Accuracy: The parameters estimated from LLM simulations demonstrated the lowest prediction error (0.0036) when applied to a three-route scenario, outperforming parameters derived from real-world surveys (0.0042) and classic non-travel parameters (0.0110).
Context-Dependent Behavior: The agents exhibited nuanced, context-dependent decision-making. For instance, while generally risk-averse, certain agent types (e.g., couriers) shifted to riskier routes when the alternative guaranteed a loss (lateness), demonstrating that LLMs can simulate complex trade-offs rather than rigid adherence to a single "risk-averse" label.
Robustness: The method maintained high predictive accuracy across different reference points and scenario configurations, including cases where gains and losses were defined under the same reference point.

Significance and Claims

The paper claims that its primary significance lies in demonstrating that generative AI models can serve as a scalable alternative for modeling human decision processes. Specifically:

Methodological Feasibility: It validates that LLMs can reproduce non-rational human choice biases and prospect-theoretic effects without explicit specification of CPT parameters, relying instead on standardized prompting to encode behavioral profiles.
Scalability: The approach offers a low marginal cost, scalable computational paradigm for generating large-scale, heterogeneous behavioral data, addressing the bottleneck of parameter estimation in traditional agent-based modeling.
Foundation for Future Research: The study posits that LLM-based agents provide a promising foundation for next-generation large-scale agent-based simulations and AI-driven behavioral research, capable of capturing individual differences and group-level emergent behaviors.

The authors maintain a modest stance, noting that while LLMs effectively simulate biases, future work is required to integrate learning, social interaction, and multimodal tools to further align agent perception and reasoning with real-world decision processes.