How NTU’s AI Policy Simulator Would Test Cash Handouts Before They Launch

A cash handout can be redeemed without generating the same amount of new spending: a household might use it for a purchase it would have made anyway. Researchers at Nanyang Technological University (NTU) are developing an AI policy simulator to help policymakers examine that distinction before choosing how to distribute support.25

What would the simulator compare?

The four-person Nanyang Business School team—Hyeokkoo Eric Kwon, Jaecheol Park, Mo Jiayun and Kayoung Shin—aims to model how consumers might respond to alternatives such as a $100 voucher. Its forecasts would examine whether recipients save money, spend more overall or change what they buy, rather than treating voucher use alone as proof of an economic boost.25

The project is listed by OpenAI as “Privacy-Preserving LLM Agents for Auditable Government Policy Simulation.” That is the research project’s name; reporting describes the tool itself as unnamed.85

How would AI agents use spending data?

The proposed system draws on anonymised transaction histories from roughly 1.5 million users of an unnamed South Korean budgeting app, covering 2023–2025. The aim is to give simulated consumer agents a grounding in varied financial circumstances and spending patterns.23

The described privacy process starts with transaction information collected through bank notifications. Identifiers are removed, while figures such as spending and income—and details such as purchase times—are grouped into ranges. Depending on the sensitivity of a profile, its characteristics must be shared by at least three, five or 10 users. These measures are intended to reduce the chance that a simulated profile exposes one person’s financial history; anonymisation should not be mistaken for a guarantee against re-identification.2

Unlike a simulation that assigns everyone in a category the same fixed spending rule, an agent could respond to several characteristics together. Income and age, for example, might jointly affect how a household uses a voucher. That flexibility is the proposed benefit, not yet proof that the agents predict people better than survey-informed or rule-based models.23

How will the forecasts be checked?

The team plans to compare simulations with observed spending before and after South Korea’s July and September 2025 handouts. Those handouts were limited to eligible retailers and had to be used by November 2025, giving researchers a real policy episode against which to examine predicted purchases and spending changes.2

That comparison can help expose implausible agent behaviour or invented explanations. To establish a forecasting advantage, however, the tool would need to predict outcomes it had not already been shown, outperform relevant conventional baselines and account for other changes affecting spending. The available reporting does not provide a detailed evaluation protocol or published error rates. Claims of greater accuracy should therefore be treated as a research objective, not a settled result.23

What do the funding and release target mean?

NTU’s project is one of 14 selected by OpenAI from more than 400 submissions. Kwon lists its award at US$100,000; reporting says half is cash and half is OpenAI model credit, supporting about six months of development and testing.8104

The team is targeting an open-source release by February 2027 and hopes to adapt the simulator to different populations. Plans also include richer agent profiles reflecting interests and political views, alongside explanations for predictions.4 Those are development goals. Whether the tool transfers reliably beyond the South Korean data—and whether its explanations reflect sound predictions—will depend on validation, not on the agents’ ability to produce convincing answers.