Out of One, Many: Using Language Models to Simulate Human Samples
Open accessAbstract
We propose and explore the possibility that language models can be studied as effective proxies for specific human subpopulations in social science research. Practical and research applications of artificial intelligence tools have sometimes been limited by problematic biases (such as racism or sexism), which are often treated as uniform properties of the models. We show that the “algorithmic bias” within one such tool—the GPT-3 language model—is instead both fine-grained and demographically correlated, meaning that proper conditioning will cause it to accurately emulate response distributions from a wide variety of human subgroups. We term this property algorithmic fidelity and explore its extent in GPT-3. We create “silicon samples” by conditioning the model on thousands of sociodemographic backstories from real human participants in multiple large surveys conducted in the United States. We then compare the silicon and human samples to demonstrate that the information contained in GPT-3 goes far beyond surface similarity. It is nuanced, multifaceted, and reflects the complex interplay between ideas, attitudes, and sociocultural context that characterize human attitudes. We suggest that language models with sufficient algorithmic fidelity thus constitute a novel and powerful tool to advance understanding of humans and society across a variety of disciplines.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Existing literature commonly employs three strategies for creating agent profiles: handcrafting, LLM-generation, and dataset alignment.
“Existing literature commonly employs the following three strategies.”
From A survey on large language model based autonomous agents (2024) · cites this paper as [29]
The Introduction describes creating simulated subjects by conditioning GPT-3 on demographic backstories from real survey participants, supporting an example of dataset-aligned profiles. It does not identify handcrafting, LLM-generation, and dataset alignment as three commonly used strategies in the literature. (LLM, conf=0.84)
Experimental — accuracy not yet measuredLLM-based agents can be utilized to study political science and economy.
“Political science and economy: LLM-based agents can also be utilized to study political science and economy [29,105,106].”
From A survey on large language model based autonomous agents (2024) · cites this paper as [29]
The Introduction and Section 3 describe using GPT-3 to simulate respondents and study U.S. politics and public opinion. Section 7 mentions econometric models as an analysis method, but the paper does not demonstrate studying the economy; that part of the claim is unsupported. (LLM, conf=0.84)
Experimental — accuracy not yet measuredMulti-task evaluation uses a set of diverse tasks from different domains to evaluate agent generalization capability in open-domain environments.
“In this method, people use a set of diverse tasks from different domains to evaluate the agent, which can effectively measure the agent generalization capability in open-domain environments [29,80,153,163,165,166,167].”
From A survey on large language model based autonomous agents (2024) · cites this paper as [29]
The Discussion describes evaluating GPT-3 across three studies, multiple measures, and data sources, which supports the use of varied evaluations. But it explicitly says the study is restricted to a specific domain; it does not establish that tasks from different domains measure agent generalization in open-domain environments. Page numbers are unavailable in our copy. (LLM, conf=0.83)
Experimental — accuracy not yet measuredLLMs have demonstrated the capability to generate human-like experimental samples in studies from political science, psychology, and behavioral science.
“Their experiments have demonstrated the capability of LLMs to generate human-like experimental samples in studies from political science [9, 11], psychology [29, 97], and behavioral science [11].”
From AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society (2025) · cites this paper as [9]
The Introduction reports that GPT-3 generated simulated respondents whose answers mirrored human samples in studies of U.S. politics and public opinion. Study 3 describes these as survey-based silicon samples that could help researchers design later experiments, not as samples demonstrated in experiments. The separate psychology citations in the citing sentence are not attributed to this paper. (LLM, conf=0.83)
Experimental — accuracy not yet measuredCommon human similarity metrics for LLM-based agents include trajectory/location accuracy, dialogue similarities, and mimicry of human responses.
“Typical examples include trajectory/location accuracy [38,162], dialogue similarities [79,102], and mimicry of human responses [29,102].”
From A survey on large language model based autonomous agents (2024) · cites this paper as [29]
Silicon samples may be useful for pretesting surveys before they are dispatched to large groups of human respondents, or for imputing missing data.
“Though silicon samples will not soon displace survey research with human respondents, they may still be very useful for pretesting surveys before they are dispatched (at considerable cost) to large groups of human respondents, or imputing missing data (11, 17).”
From Can Generative AI improve social science? (2024) · cites this paper as [11]
GPT-3 can accurately impersonate respondents to a nationally representative public opinion survey from a range of different demographic backgrounds.
“For example, Argyle et al. indicate GPT-3 can accurately impersonate respondents to a nationally representative public opinion survey from a range of different demographic backgrounds (11).”
From Can Generative AI improve social science? (2024) · cites this paper as [11]
The demographic representation in the outputs of LLMs can be improved by using prompts featuring personas from a broad demographic spectrum.
“For instance, the demographic representation in the outputs of LLMs can be improved by using prompts featuring personas from a broad demographic spectrum 93 .”
From Machine culture (2023) · cites this paper as [93]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 37 · Cited by 783
Research works printed but not linked: 6
- [10]10. Brown, T. B., et al. 2020. "Language Models Are Few-Shot Learners." Advances in Neural Information Processing Systems 33: 1-25.
- [13]13. Campbell, A., P. E. Converse, W. E. Miller, and D. E. Stokes. 1960. The American Voter. Chicago, IL: University of Chicago Press.
- [28]28. Magleby, D. B., C. J. Nelson, and M. C. Westlye. 2011. "The Myth of the Independent Voter Revisited." In Facing the Challenge of Democracy: Explorations in the Analysis of Public Opinion and Political Participation, edited by P. M. Sniderman, and B. Highton, 238-266. Princeton, NJ: Princeton University Press.
- [30]30. Mason, L. 2018. Uncivil Agreement. Chicago, IL: University of Chicago Press.
- [31]31. Mayson, S. G. 2018. "Bias In, Bias Out." Yale Law Journal 128: 2218.
- [33]33. Radford, A., J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever. 2019. "Language Models Are Unsupervised Multitask Learners." OpenAI Blog 1 (8): 9.
- Authors
- Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D.
- Affiliations
- Brigham Young University (US)
- Corresponding author
- Argyle, Lisa P.
- Venue
- Political Analysis
- Publisher
- Cambridge University Press (CUP)
- Published
- 2023-02-21
- Volume
- 31
- Issue
- 3
- Pages
- 337-351
- ISSN
- 1047-1987 (ISSN-L)
- 1476-4989
- OpenAlex
- W4321455981
- Semantic Scholar
- f4e612658bde9db88abfd455b99f181fdc536996
- Cited by
- 783
- References
- 37
- Citations by year
- 2026: 323 · 2025: 252 · 2024: 147 · 2023: 60 · 2022: 1
- FWCI
- 195.92
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- Political Analysis (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- RePEc: Research Papers in Economics (submittedVersion)
- arXiv (Cornell University)
- Funders
- Brigham Young University
- Topics
- Social Sciences › Safety Research › Ethics and Social Impacts of AI
- Social Sciences › General Social Sciences › Computational and Text Analysis Methods
- Physical Sciences › Physics and Astronomy › Statistical and Nonlinear Physics › Opinion Dynamics and Social Influence
- Keywords
- GPT-3, algorithmic bias, attitudes, large language models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
