17 SEP 2026 — A team at Nanyang Technological University has won OpenAI funding to build a tool that predicts what people actually do with a government handout: spend it, save it, and on what.
The grant is US$100,000, about S$126,800, and it is one of fourteen projects chosen worldwide from more than four hundred proposals. Half of it is cash and half is credit for OpenAI's models. The work runs for six months.
What the tool is trained on
The system learns from transaction records belonging to more than a million users of a South Korean mobile budgeting app, covering 2023 to 2025. The app builds its picture by reading bank notifications, messages and emails, which means the data records what people did rather than what they later told a survey they did.
The design's core is its data source. Traditional models of consumer response rely on surveys or aggregate statistics, which are slow and depend on self-reporting. This tool uses records of what people actually spent.
The Straits Times reports the tool forecasts outcomes with far greater accuracy than traditional simulation models. No figure has been published for that comparison, and it is a claim about a system still being built.
The team has stated its privacy design
Hyeokkoo Eric Kwon, the Nanyang Business School professor leading the team, described a specific threshold rather than a general assurance: "There should never be a situation where the data can only be matched to one profile."
In practice that means a minimum of three, five or ten users sharing identical characteristics before a record is usable, with the threshold rising as the data gets more sensitive. Names and account numbers are removed and amounts are grouped into ranges rather than held as exact figures.
Anyone who has followed re-identification research will recognise the reasoning: an anonymised transaction history is one of the easiest things in the world to re-identify, because a handful of purchases at particular shops on particular days is close to a fingerprint. A minimum group size is the standard defence against re-identification. Stating the number is the difference between a design and a promise.
Why this question matters in this region
Cash transfers are a standing instrument of policy across Southeast Asia, not an emergency measure. Singapore distributes vouchers and payouts as a routine part of its budget, Malaysia and Indonesia run large targeted transfer programmes, and the question every finance ministry has is the same one: how much of a payout comes back as consumption, how quickly, and from which households.
In practice, that response is estimated rather than known. A tool trained on what people actually spent, if it generalises, answers a question that ministries currently answer with models built on assumptions about how people behave.
The open question is how well the model generalises. The training data is South Korean, and consumption is shaped by local factors: savings culture, household debt, credit availability and the size of the informal economy. A model that predicts Seoul well may predict Jakarta badly, and nothing published so far tests that.
What OpenAI is buying
Fourteen grants of a hundred thousand dollars is not a research programme in any meaningful financial sense. It is a small amount of money spread across a set of questions, with half returned to the funder as usage of its own models.
The grant buys position. Policy evaluation has old tooling and conservative institutions, and the first credible entrant sets the terms. Funding fourteen academic teams to work on government policy design costs less than a single engineering hire and puts the company's models inside the workflow of the people who will later decide how such systems are procured.
This is not a criticism of the research, which is open and worthwhile, but an observation about the funder's strategy.
What to watch
Start with the open-source release, targeted for February 2027. A model of consumption response that ministries can run themselves is a public good. One that requires a commercial API to use is a product with a paper attached.
Then validation outside Korea. For readers here, the test is whether the tool predicts a Singaporean or Malaysian payout, and that needs local transaction data with the same privacy floor.
Adoption is the hardest test. Tools like this usually founder not on accuracy but on procurement, data-sharing agreements and the fact that the agency holding the transaction records is rarely the one designing the policy.
