BoJ researchers test AI personas on household inflation views with mixed results
Serguey Asael Shinder1 min readThe Bank of Japan's Institute for Monetary and Economic Studies has published discussion paper 2026-E-14, dated November 2026, by Yusuke Takahashi and Kazuki Otaka. It asks whether personas generated by large language models can stand in for the varied households of an economic simulation, using the Bank's own household survey as the yardstick.
Each persona receives the attribute profile of a real survey respondent, and the authors compare the model's inflation expectations with that respondent's answers one to one. They split the match into three parts: the spread of answers, effects at the level of attributes such as age or income, and agreement at the level of the individual.
The results are mixed. Personas reproduce the direction of most attribute effects on short-horizon expectations, and conditional averages over a one-year horizon are partly recovered. But the spread of answers stays a fraction of the human level in every configuration tested, and the correspondence weakens as the horizon lengthens, to nothing for five-year expectations.
One addition helped. Following theories of information frictions, the authors added cognitive details drawn from the same respondents, such as awareness of monetary policy and attention to prices. Describing a respondent's awareness of monetary policy can more than double how closely the persona tracks one-year expectations, which the paper reads as a way to bring information frictions into a simulation as persona text rather than as calibrated parameters. The views are the authors' and not necessarily the Bank's.
What it means
For a central bank that relies on household surveys, the paper sets a limit on synthetic respondents: they echo the short-term average but not the disagreement or the long-run anchor, which are what policymakers most need to see.