What the paper found

Piskorski, Seru, Zhang and Zhao (NBER Working Paper 35756, September 2026, https://www.nber.org/papers/w35756) linked every private-housing purchase in Singapore from 1995 to 2019 to the government directory and asked whether civil servants bought near new Mass Rapid Transit stations before those stations were announced. They did: compared with similar buyers outside the civil service, officials’ purchase rate within 1 km of a future station was 0.5 to 0.7 percentage points a year higher (more than 60% above the average rate), concentrated one to two years before the announcement and not statistically significant outside that window.

The buying was concentrated among the people you would expect to know: mid-level officials, and staff in agencies involved in rail planning. Real-estate agents and company directors, who are sophisticated about property, showed no such excess, which counts against officials simply reading public information better. People registered at the same address as officials (the paper calls them relatives) also bought more before announcements (borderline significant, starting a year sooner) and kept buying afterwards. On homes later sold, returns beat the market by about 2.4 percentage points a year (6.8 against homes away from stations), worth somewhere between a fifth and three-fifths of an average official’s income.

After 2011 and 2012, when a run of corruption scandals led Singapore to tighten its laws and enforcement, the excess buying by officials and their households was no longer detectable. That is the point of the paper: Singapore is meant to be the case where decades of hard enforcement produced an integrity culture that runs on its own, and the authors read the data as a culture that was eroding until enforcement was turned back up.

As of this writing it is a working paper that has not been peer-reviewed, and the 2012 result rests on a single episode in which other things were also changing.

The AI part

The viral version, as summarised on Digg, says researchers used AI to scan every property deal and cross-reference 141,000 public servants, and reports 12% annual returns.

The detection is standard matched-sample econometrics, and ChatGPT’s only job, described in an appendix, was sorting job titles into senior, middle and junior ranks so the authors could ask which level the buying came from. The main result does not use it. The 12% is accurate but raw, before adjusting for the market, whereas the excess over the market is 2.4 points. The 141,000 is the size of Singapore’s entire public service, quoted in the paper as background.

The appendix reproduces the prompt, which opens: “You are a GPT with a formal and academically rigorous approach, specializing in analyzing and tagging data for Singapore government employees.”