Blog · Field notes · Oct 21, 2026 · 5 min
Census-built populations, agents grounded in interviews and language models speaking for whole groups have all advanced in 2025-26. The evidence now shows clearly which of them match real people, and the dividing line is what each simulated person is built from.
Simulated people have become a serious research instrument. In 2025 and 2026 the field produced national synthetic populations of hundreds of millions, agents grounded in hours of interview, a foundation model of human choice in experiments, and the first large tests of language models answering surveys in place of the people who would have answered them. The results now point to a simple rule: a simulated person is as good as what it is built from. People built from records of real people hold up. People built from what a model has read about a group tend to drift towards the group's stereotype.
The oldest form of synthetic population is also the most mature. In January 2025 Rineer and colleagues at RTI published a national synthetic population for the United States in Scientific Data: 120,754,708 households and 303,128,287 people, drawn from the 2015-2019 American Community Survey microdata and placed across the country. Populations like this are fitted so that their totals match official tables, and they reproduce census categories closely. They are the base layer of epidemic models, transport models and energy demand studies.
What they hold is structure: who lives where, with whom, at what age. What they do not hold, by design, is behaviour over time. A census-fitted population is a photograph. To ask what it will do, something has to make it live.
The strongest result for individual simulation comes from Stanford. Park and colleagues interviewed 1,052 Americans for two hours each and built an agent for each from the transcript. In the 2026 revision of the paper, interview-based agents reproduce participants' General Social Survey answers 83% as accurately as the participants reproduce their own answers two weeks later, rising to 86% when interviews and survey answers are combined. Agents given only demographics reach 74%.
Centaur, published in Nature in July 2025, makes a similar case for behaviour in the lab. Fine-tuned on Psych-101, trial-by-trial data from more than 60,000 participants making over 10 million choices across 160 experiments, it predicted held-out participants better than established cognitive models and carried over to new cover stories and new task structures.
Both results share a feature: the simulated person is anchored in data about that person, or about people doing that exact kind of task.
The weakest results come from the most popular shortcut: describe a demographic to a language model and ask it to answer as a member. Wang, Morgenstern and Dickerson, in Nature Machine Intelligence in 2025, showed that models used this way misportray identity groups and flatten the variety inside them, and traced the problem to limits in how current models are trained.
The largest direct test arrived in September 2026. Pew Research Center built an AI counterpart for every respondent in three waves of its American Trends Panel, each from that person's real profile, and asked the counterparts the survey. The best configuration, using Opus 4.6, had an average absolute error of 11.4 points on one politics wave and 14.7 on a global attitudes wave. Richer profiles helped, from 16 points down to 13.1 in Pew's configuration tests, and more reasoning did not. Pew's title states its conclusion: AI cannot really stand in for human survey-takers.
At the scale of whole societies, AgentSociety from Tsinghua simulated the social lives of more than 10,000 language-model agents with about 5 million interactions, and used them to study polarisation, inflammatory messages, universal basic income and hurricanes. Simulators like this make it possible to pose questions about a society that no survey can. The open question for all of them is the one Pew asked: matched against the measured answer, how close are they?
Read together, the evidence separates cleanly. Structure fitted to records holds. Individual agents grounded in what a specific person said hold well. Model voices asked to speak for a group miss, and they miss in a characteristic way: they play the group's stereotype and get the overall level wrong. A model voice can still earn a place, but as one input weighed against real data rather than as the population itself.
That matters because synthetic populations are moving from research into decisions: policy trials, market forecasts, public health planning. A population that plays stereotypes will confirm the assumptions it was given. A population fitted to records, and checked on answers it never saw, can surprise its makers, which is the point of asking.
Living Worlds fits lives to records first, and its register shows both sides of the line.
Model voices, scored as registered: answering as each of 1,936 people, Gemini 3.8 Flash split gun-permit opinion by party and age 16.4 points from the 2024 General Social Survey, playing each group's stereotype (voice not met, 0 of 6). A small language model answering as each of 2,872 people missed welfare opinion by 58.9 points (voice not met). A frontier model asked directly for each group's share, calibrated on eleven other questions, reached a mean error of 6.2 points, ahead of simple baselines, but landed inside the survey's own noise on 1 of 5 questions where 3 were needed (Not met). Asked about births, prompted voices overstated the total about threefold, 5.24 births per woman against 1.62 (not consistent).
Where a voice did earn a place, it was as a measured input. Counted as 50 respondents alongside 50 real ones per group, a model's estimates came closer than 75 real respondents alone, though 100 real respondents came closer still and on one question the model added nothing (Met). A voice trained on lived records gave its world's birth rate, 1.75 against 1.62 (Met). An agent that remembers what a person said predicted their answers four years later 2.5% better than a persona, in every panel and on every question (Met); given only what had happened to the person, it did no better (Not met).
Lives fitted to records: five fresh 2024 questions split by party and age all landed inside the survey's own sampling noise (Met, 5 of 5), and across 31 African countries in Afrobarometer round 10, 1,556 unseen questions split by place and age landed inside the noise in 30 of 31 countries (Met). Each was registered before any answer was read and scored once.
Sources
Blog
Language models answering as real people, scored once on survey answers nobody had read: a frontier model met 0 of 6, prompted voices overstated births threefold, and carrying the last shares forward beat a small model. Lives fitted to records came closer, a model's estimate counted as a prior meant fewer people to ask, and subgroup splits landed inside the survey's own noise in the United States and in 30 of 31 African countries.
A population lived month by month in the United States from 1950 matched how many children Americans have had and how many brothers and sisters they report, closer than a frontier model answering from what it has read. The father's side, first counted through partnerships, missed, and became the next test.
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.