What is human simulation?
Human simulation is the use of a model to produce what people would say or do in a situation, for one person or for a whole population, without asking or observing them at that moment. With language models it now covers survey answers, the results of experiments, conversations, and small simulated societies.
Also called human behavior simulation, LLM social simulation, social simulation, simulating human behavior, generative agent simulation.
Three levels of the same question
A human simulation can be right at one level and wrong at another, so it helps to say which level is meant.
- The individual. Does the model give the answer this particular person would give? This is what a digital twin attempts.
- The population. Does the spread of simulated answers match the spread of real ones? Errors in opposite directions cancel across a population, so the population level is the easier test.
- The effect. If something changes, such as a price or a message, does the simulated result move the way the real one would?
How language models entered the field
Several research groups began using language models for human simulation around 2022 and 2023. Aher, Arriaga and Kalai proposed the Turing Experiment: have a language model stand in for a representative sample of participants and see whether a classic study replicates. They reproduced the Ultimatum Game, garden path sentences and the Milgram shock experiment, and found a distortion in a fourth study, where simulated people were more accurate than real ones.
The economist John Horton argued that language models are implicit computational models of humans, which he called Homo silicus, and can be given preferences and information and then observed. Argyle and colleagues conditioned a model on real survey respondents' backstories and called the result a silicon sample.
Computer scientists at Stanford went from answers to behavior. Park and colleagues built 25 characters with memory and planning living in a small simulated town, and the following year built agents of 1,052 real Americans from two-hour interviews.
What it can do today
Predicting the effects of experiments is the strongest result in human simulation so far. A 2026 paper in Nature found that language models can predict the results of social science experiments.
Simulating individuals is harder than simulating effects. The Stanford agents built from two-hour interviews reached 85% of people's own test-retest consistency on a standard survey. Centaur, a foundation model of human cognition, captured the choices of people it had never seen in psychology experiments better than existing cognitive models.
Where it breaks
Simulated people vary less than real ones. A comparison with the American National Election Study found averages close to the real data and less variation in the answers, far less on some questions.
Training a model to be a helpful assistant works against simulation. The SimBench benchmark found that instruction tuning improves answers on questions where people agree and hurts them where people disagree, which is where research is most useful.
The accuracy of human simulation also falls with distance from the training data. One study measured a correlation of minus 0.70 between a country's cultural distance from the United States and how closely model answers resembled its people.
How it is measured
A simulation of people is graded against people: answers that were held out, real survey results, and how consistent people are with themselves when asked twice. That last number is the ceiling, because no simulation can match a person better than the person matches themselves. Jean Technologies collects the published benchmarks on its open evaluations page.
What it is used for
In market research, human simulation shows up as synthetic respondents and digital twins. In product and AI work, human simulation shows up as a user simulator that tests an interface or trains an agent. In social science, human simulation is a way to pilot an experiment before running it on people.
Human simulation and related terms
- Synthetic respondents
human simulation applied to surveys and interviews in market research.
- Digital twins of customers
simulation at the level of one specific, real person.
- User simulator
simulation of people using a product or talking to an AI agent.
- Foundation model of human behavior
a model built for this job, trained on what people do, where most simulation today uses a general assistant.
Questions people ask
- What is human simulation?
- Human simulation is the use of a model to produce what people would say or do in a situation, for one person or for a whole population, without asking or observing them at that moment. With language models, human simulation covers survey answers, experiment results, conversations and simulated societies.
- Can language models simulate human behavior?
- Partly. Language model predictions of experimental effects correlate strongly with the real effects, though they overestimate their size, and agents built from long interviews come close to people's own consistency on standard surveys. Simulated people vary less than real people, and language models are less accurate for groups far from their training data.
- What is LLM social simulation?
- LLM social simulation is the use of language models to stand in for people in surveys, experiments or simulated societies. Early examples of LLM social simulation include Turing Experiments, silicon samples, Homo silicus and generative agents.
- How is a human simulation evaluated?
- A human simulation is evaluated against real people: held-out answers, real survey results, and the human ceiling. The human ceiling is how consistently people give the same answer when asked twice, and no simulation can match a person better than that.
- Which companies work on human simulation?
- Companies working on human simulation for research and decisions include Simile, Aaru, Artificial Societies, Synthetic Users and Jean Technologies. Jean Technologies offers digital twins built from real interviews, simulated users for product testing, and human simulators for training AI agents.
What Jean offers
Jean Technologies is the #1 AI market research platform to learn from real people and digital twins.
- Jean Technologies offers an AI interviewer that runs AI-moderated interviews with real people, by voice or text, in any language, for research teams that need the reasons behind a number from more people than they can interview themselves.
- Jean Technologies offers digital twins, synthetic respondents each built from a real person's interview, for teams doing AI market research who want to test concepts, prices, bundles and messages before fieldwork, a practice also called decision simulation.
- Jean Technologies offers simulated users that use a website, app, chatbot or voice agent before launch, for product teams that want to see where people drop off and why.
- Jean Technologies builds Alyosha, a foundation model of human behavior trained on what people say, what they do, and why, for companies that want digital twins and simulations run on a model of people. Alyosha has the highest overall score on Jean's open evaluations of how well models simulate people: 84.6 across the SOUL benchmark tasks.
Sources
- Aher, Arriaga, Kalai (2023), Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies, ICML
- Horton (2023), Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?
- Argyle et al. (2023), Out of One, Many: Using Language Models to Simulate Human Samples, Political Analysis
- Park et al. (2023), Generative Agents: Interactive Simulacra of Human Behavior
- Park et al. (2024), Generative Agent Simulations of 1,000 People (arXiv v1)
- Ashokkumar, Hewitt et al. (2026), Large language models can predict the results of social science experiments, Nature
- Binz et al. (2025), A foundation model to predict and capture human cognition, Nature 644
- Bisbee et al. (2024), Synthetic Replacements for Human Survey Data? The Perils of Large Language Models, Political Analysis
- Hu et al. (2025), SimBench: Benchmarking the Ability of Large Language Models to Simulate Human Behaviors
- Atari et al. (2023), Which Humans?