Large language models (LLMs) are increasingly used as agents to simulate human behavior, yet their fidelity in complex decision-making under uncertainty remains insufficiently understood. To address this gap, we developed a comparative framework that benchmarked LLM-simulated risk preferences against empirical human behavior. Using demographic profiles from surveys conducted in Sydney, Hong Kong, and Nanjing, we constructed role-playing prompts and evaluated three LLMs on abstract lottery-choice tasks. We adopted the classical constant relative risk aversion (CRRA) framework as a domain-neutral “standard ruler” to compare risk attitudes. The analysis yielded three main findings. First, off-the-shelf LLMs do not exhibit a universal risk profile: The two GPT models are more risk-averse than human benchmarks, whereas Gemini is more risk-seeking. Second, prompt language systematically affects simulated risk attitudes, with English-to-Chinese switching inducing a more conservative shift in most cases. Third, LLMs do not reliably reproduce the empirical heterogeneity of human risk preferences, tending either to generate overly concentrated distributions or unrealistically large dispersion. Taken together, these findings show that off-the-shelf LLMs remain vulnerable to model-family-specific miscalibration, language-sensitive distortions, and failures in distributional fidelity. Rigorous empirical calibration is therefore necessary before off-the-shelf LLMs can be reliably deployed in computational social science and choice modeling.

This research addresses the growing concern of balancing personalized services with data privacy in the ride-hailing industry. While personalized pricing and matching strategies, fueled by travelers’ personal data, can optimize platform revenue, they also expose users and platforms to significant privacy risks. The correlation between personalized pricing, waiting times, and personal information might be exploited by third-party agents to infer sensitive user attributes, resulting in potential economic losses for the platform and severe consequences for users, including compromised privacy and potential discrimination. Existing privacy protection methods often fall short in providing robust and quantifiable guarantees. To overcome these limitations, this study introduces a privacy-preserving approach for personalized pricing and matching within ride-hailing platforms. The proposed approach leverages the bounded Laplace (BL) mechanism and parallel composition to inject noise into the order price and waiting time feedback provided to travelers. This study rigorously demonstrates that the proposed approach satisfies differential privacy. Furthermore, the proposed approach outperforms other classic privacy-preserving methods in terms of platform revenue. This superior performance is validated through extensive numerical experiments using realistic ride-hailing data.