Key Summary
- A research team at Nanyang Technological University in Singapore is using ChatGPT to simulate how households may respond to government consumer support payments.
- The project is part of an OpenAI-backed policy research program and received API credits, access to frontier models, and research funding.
- The researchers are examining South Korea’s 2025 first-round livelihood recovery consumer coupon program.
- They are comparing the model’s results with anonymized financial transaction data and a Bank of Korea report on consumer support effects.
- The team says prediction error has fallen from about 25% early in development to around 10%.
Main Story
Researchers supported by OpenAI are testing whether large language models can be used to estimate the impact of public policy on consumer behavior, with a current focus on South Korea’s consumer coupon program.
OpenAI said it recently introduced the project during a media session on the use of artificial intelligence in policy research. The work is being carried out by a team led by Prof. Kwon Hyuk-koo at Nanyang Technological University in Singapore. According to the report, the study is one of 14 independent projects selected under OpenAI’s industrial policy research support initiative.
The team is using ChatGPT to run simulations based on household profiles that summarize income and spending habits, along with policy conditions. The goal is to estimate how spending may change over a four-week period after support is paid and whether the payment creates additional consumption.
The article says the researchers are focusing on the first payment round of South Korea’s 2025 livelihood recovery consumer coupon scheme. Their approach contrasts with agent-based models, which generally require researchers to define formulas and rules one by one. The team says large language models can reduce that burden and interpret more complex patterns of consumer behavior more flexibly.
Prof. Kwon said the key question is not simply whether the entire subsidy is spent, but whether the money only replaces spending that would have happened anyway or actually generates extra consumption. He said the broader aim is to shift policy debate away from politically driven arguments and toward evidence-based design.
To check the model’s performance, the researchers are using anonymized transaction records from household accounting apps and cross-checking them with a Bank of Korea report released in June that analyzed the effects of consumer support payments using credit card company data. The report says the model’s error rate has improved from roughly 25% in the early stage to about 10% now.
On privacy concerns, Kwon said the team buys and uses anonymized data with user consent and inputs only categorized or summarized information, rather than raw location coordinates or original payment logs. He also said OpenAI API data are not reused for model training, which he said lowers privacy risk. In addition, the team is controlling for possible future-information bias by selecting model versions during the first verification stage.
The researchers plan to complete validation of predictive accuracy by the end of this year using actual implemented policies. They also aim to finish a second phase by the end of March next year, comparing policy alternatives such as extending the payment period or restricting how the funds can be used.
Background
The supplied text says this project is part of OpenAI’s support program for industrial policy research and involves 14 independent research projects worldwide. It also notes that the team is examining a South Korean consumer support program and using Bank of Korea analysis as a reference point. No additional background was provided.
Items Requiring Further Verification
- The reported prediction error rate of “about 10%” is presented without methodological detail in the supplied text.
- The claim that OpenAI API inputs are not reused for model training reflects the speaker’s explanation in the article and was not independently verified in the text.
- The statement that the model will be validated against “actual implemented policies” by year-end describes a future plan, not a completed result.
- The second-phase timeline and the comparison of policy options remain planned outcomes rather than verified findings.