用于支付应用调查的生成式人工智能：人工智能对隐私和技术的看法

Generative AI for surveys onpayment apps: AI views on by Koji Takahashi and Joon Suk Park Monetary and Economic Department March 2026 JEL classification: M31, C83, C45, D12, L86Keywords: ChatGPT, generative artificial agents, privacy BISWorking Papers are written by members of the Monetary and EconomicDepartment of the Bank for International Settlements, and from time to time by othereconomists, and are published by the Bank. The papers are on subjects of topical This publication is available on the BIS website (www.bis.org). ©Bank for International Settlements 2026. All rights reserved. Brief excerpts may be ISSN 1020-0959 (print)ISSN 1682-7678 (online) Generative AI for Surveys on Payment Apps:AI Views on Privacy and Technology Koji Takahashi†Joon Suk ParkMarch 2, 2026 Abstract This study uses ChatGPT to simulate survey responses about payment apps, focusing onprivacy and perceived benefits. By designing prompts that mirror real user characteristics,the generated responses align with findings from a Dutch survey, especially when groupedby privacy concern. Privacy-concerned agents view apps less favorably, while users show Keywords: ChatGPT, generative artificial agents, privacy paradox, Westin index, survey,payments 1Introduction In this decade, we have witnessed the rapid and transformative evolution of artificial intelli-gence (AI). In some specific tasks, researchers report that the AI’s intelligence level has reacheda level comparable to that of human beings (Hendrycks et al. (2020); Choi et al. (2021); Kung A growing number of models and applications of generative AI (GenAI) have been devel-oped by researchers, companies, and various institutions. One potential application of GenAIis as a tool for market surveys. Market surveys require significant costs and time for implemen-tation. In addition, it is not easy to obtain responses from samples that accurately represent the In this paper, we apply GenAI as generative agents to survey the usage of payment apps,with a particular focus on perceptions of privacy and benefits. To ensure the validity of surveyresponses generated by GenAI, we compare them with the empirical findings of Brits andJonker (2023), who document that Dutch consumers’ use of financial apps largely reflects arational privacy calculus—carefully weighing perceived benefits against privacy risks ratherthan exhibiting paradoxical behavior. This comparison allows us to evaluate whether GenAI (2023); Li (2023)). However, user perceptions of privacy issues are inherently complex. Previous Our study suggests that GenAI has the potential to serve as a complementary tool to privacy issues.Specifically, GenAI could be used to help researchers brainstorm in orderto create survey questions and conduct simulations of surveys before embarking on surveyswith actual human beings.Using ChatGPT-4o for simulating surveys, first, we find thatgenerative agents’ views on payment app benefits and privacy are similar to actual surveyresults. Privacy-concerned agents view financial apps less favorably and see more risks, evenwithout specifying this tendency in prompts. Second, the views of generative agents on thebenefits of financial apps are more widely dispersed than those on risk, which is consistent However, there are caveats when generative agents are applied to surveys. First, the gener-ative agents do not generate as much variation as the responses of actual humans. This limitedvariation fails to capture the wide range of responses found in actual surveys. Another caveatis that most generative agents are classified as “privacy fundamentalists,” which is inconsistentwith the actual survey. In other words, ChatGPT generates a bias in views on privacy and risksrelated to payment apps. To examine the source of this bias, we simulated another experiment personas does not necessarily help to generate better synthetic surveys. The rest of the paper is organized as follows. Section 2 discusses the literature. In Section3, we illustrate a general methodology for applying GenAI to market surveys.In Section 4, we introduce the study by Brits and Jonker (2023) that we replicate and report our replication 2Literature review Our paper is mainly related to three strands of literature. The first strand of studies investigateseconomic experiments using GenAI. Using a large language model (LLM), Horton (2023) andMa et al. (2023) implemented economic experiments, which are motivated by classic exper-iments in the behavioral economic literature.They find that ChatGPT can generate similarresults to the original studies. Their findings suggest that ChatGPT is able to replicate humancognitive processes in a number of ways.For example, ChatGPT can make inferences from Second, our paper is closely related to the study by Brand et al. (2023), which explores thepossibility of using generative AI for market and social science research.Brand et al. (2023)use ChatGPT to conduct market research on consumer