Twins of people
Models that stand in for people: a patient, one person's views, or a whole population, and the evidence on how well they do it.
Places are full of people, and many questions put to a model of a place are about them. The entries here model the people themselves. Digital twins of people sets out the three kinds, and Synthetic people goes through the evidence.
Terms
Patient twins in medicine, language-model agents built from interviews with single people, and simulated populations built from the census or from language models.
-
A made-up set of households and people for a place, built so that its totals and mixes match the census, letting a model hold everyone without holding anyone's real records.
Questions
-
Who lives here, what would a change do to their days, and how would they react to it, before anyone tries it on the real place?
Articles
-
A 70-minute interview on The Generalist with Joon Sung Park of Simile, on simulating real people with language models, from Schelling's model to pricing tests for CVS Health.
-
The Santa Fe Institute's report on its September 2026 working group: why the organisers think the moment has come, what a twin of a cancer patient would need, and why trust and validation will decide whether twins are used.
-
Stanford-led study, 2024, revised 2026: agents built from two-hour interviews with 1,052 Americans predicted their survey answers at 83% of the people's own two-week retest consistency.
People & society open sourcerestricted data
- Large language models that replace human participants can harmfully misportray and flatten identity groups article draft
Nature Machine Intelligence paper, 2025: across four language models and 3,200 people in 16 identities, a model speaking as a group sounds like outsiders imitating it and gives too narrow a range.
-
Political Analysis paper, 2023, from Brigham Young University: GPT-3, given the backgrounds of real US election survey respondents, reproduced the overall pattern of their vote choice and attitudes.
People & society open data
- Simulated responses from large language models are weak proxies for human economic preferences across 74 countries article draft
Nature Human Behaviour, October 2026: twelve model setups answered a survey of risk, patience, trust and generosity for 67,201 people in 74 countries, explaining about 1% of individual variation.
People & society open sourceopen data
-
The 2022 Stanford paper that filled an imagined online community with thousands of language-model members, so a designer could see what it might turn into before anyone joined.
-
Models of a population at the level of its people: residents built from the census, run through their days, and given a voice by language models, with what the evidence says about each layer.
-
Political Analysis paper, 2024: ChatGPT personas matched US survey respondents' average feelings toward 11 groups but gave too little spread, wrong regression results and shifting answers.
People & society open data
-
2025 journal paper: GPT-3.5, playing 1,905 respondents to the 2017 German Longitudinal Election Study, matched their reported vote 46% of the time and leaned toward the Greens and the Left.
Tools
-
Company that says its simulated populations, grounded in demographic, behavioural and outcome data, let clients test a product, price, message or strategy before committing; it publishes its own accuracy benchmark against published surveys.
People & society closed sourceclosed data
-
Open-source platform from Tsinghua University's FIB Lab for simulating societies of language-model agents; version 1 simulated city life for over 10,000 agents, and version 2 adds AI assistants that design and run the experiments.
People & societyCities & places partly open sourceopen participation
-
Python library from Google DeepMind for building simulations in which language-model characters act and a "game master", borrowed from tabletop role-playing games, decides what their actions do.
People & society open sourceopen participation
-
Open-source Python package from Expected Parrot for running surveys and experiments on language models and simulated respondents with set traits, with tools to send the same survey to real people for comparison.
People & society open sourceAPIopen participation
-
Research code from Stanford for language-model characters that remember, reflect and plan, shown in Smallville, a small game town of 25 agents; a 2024 follow-up repository builds agents from interviews and survey answers and lets researchers put questions to them.
People & society open sourcepartly open dataopen participation
-
Company founded by the Stanford researchers behind generative agents that sells simulations of how customers, employees or populations would respond to a change, from models it says are trained on data about real people.
People & society closed sourceclosed data
-
Experimental Python library from Microsoft that simulates people with set personalities, beliefs and jobs using a language model, for simulated focus groups, interviews and market surveys.
People & society open sourceopen participation