Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

LLMs Model Non-WEIRD Populations: Experiments with Synthetic Cultural Agents

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that LLM agents primed with web-scraped cultural profiles can reproduce the economic behavior of non-WEIRD small-scale societies in classic experiments.

desk verdict The paper's transparent RAG pipeline for LLM cultural agents is worth discussing, but its central validation fails because the synthetic dictator game measures acceptance of proposed splits, not chosen offers. read the letter →

arxiv 2501.06834 v1 pith:NIDRZECE submitted 2025-01-12 cs.AI cs.CLecon.GNq-fin.EC

Augusto Gonzalez-Bonorino (1) , Monica Capra (2 , 3) , Emilio Pantoja (4) ((1) Pomona College Economics Department , (2) Claremont Graduate University Economics Department , (3) University of Arizona Center for the Philosophy of Freedom , (4) Pitzer College Economics , Computer Science Department) This is my paper · ORCID
classification cs.AIcs.CLecon.GNq-fin.EC
keywords BehavioralGamesExperimentsLarge-LanguageModelsSmall-scalesocietiesSyntheticCulturalAgentsRetrieval-AugmentedGeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that large language models, primed with a web-scraped cultural profile, can serve as preliminary stand-ins for members of non-WEIRD (Western, Educated, Industrialized, Rich, Democratic) small-scale societies in classic economic experiments. The authors build six synthetic cultural agents—Hadza, Machiguenga, Tsimané, Aché, Orma, and Yanomami—and run them through dictator, ultimatum, and endowment-effect tasks. They report substantial cross-cultural variability in the agents' choices, an absence of purely self-interested behavior, and qualitative resemblance to real human subjects wherever comparison data exist. If true, the method gives behavioral researchers a cheap, ethical way to generate hypotheses and pilot protocols for populations that are hard to reach in the field. For the Yanomami, a population with no prior experimental data, the same pipeline produces novel, testable predictions about their likely behavior.

What carries the argument

The carrying mechanism is a Search + Retrieval-Augmented Generation pipeline for building cultural profiles. For each tribe a web search on 'What characterizes the X tribe?' returns the top ten results; those pages are scraped, split into fixed-length text chunks, embedded in a vector store, and the most similar chunks are injected into a profile-writing prompt. The resulting profile—covering lifestyle, cultural practices, economic system, social organization, kinship, and core values—then becomes the system prompt for a language model that answers the experimental tasks using the strategy method: accept or reject each split from 0% to 100% of a day's wage, repeated one hundred times per agent. For the endowment effect, the pipeline adds a multimodal step in which images of two food items are shown to the agent, mirroring the field protocol where objects are placed in front of participants.

What would settle it

Search the retrieved chunks and scraped sources for each tribe for the benchmark game statistics—for example, 'Hadza modal offer 0-10%' or 'Machiguenga mean ultimatum offer 25% and 4.8% rejection.' If any such number appears in the context the language model sees, the qualitative resemblance is at least partly an echo of the input. The decisive test, which the paper itself names in its limitation section, is to profile a population with no existing experimental data, run the synthetic agents first, and then compare against a field implementation of the same protocol.

Watch

Extended reading notes

Core claim

The central claim is that synthetic cultural agents—language models given a retrieval-augmented cultural profile as their system prompt—produce experimental behavior that qualitatively resembles that of real human subjects from the same societies. In the dictator game, the synthetic Aché are the most generous and the Yanomami the closest to a self-interested homo economicus, an ordering the authors read as consistent with anthropological accounts of egalitarian hunter-gatherers versus more family-level cooperators. In the ultimatum game, offers and rejection rates vary significantly across tribes, with the Hadza agent rejecting low offers more often than the horticulturalist Tsimané and Machiguenga agents, again tracking the human benchmark data. The paper also reports that no synthetic agent, including the unstudied Yanomami, shows purely self-interested behavior. The authors frame these results as validating the method for hypothesis generation and protocol piloting, explicitly complementary to human-subject field research.

Load-bearing premise

The validation assumes that the web text used to build each profile does not already contain the very experimental results being used as the benchmark; the paper acknowledges its profiles rely on outsider-written accounts from the same literature as the comparison data, and it provides no audit to rule out this contamination.

Editorial extensions

If this is right

  • Researchers can pilot dictator, ultimatum, and endowment-effect protocols on synthetic agents before field work, lowering cost and reducing the ethical burden of studying hard-to-reach communities.
  • For unstudied populations, the method yields concrete predictions; the Yanomami agent behaves close to self-interest as a dictator and proposer but rejects low offers as a responder, a pattern a field study could test.
  • The cross-cultural ordering in the synthetic data tracks known drivers of experimental behavior: family-level cooperation, market integration, community size, and sharing norms.
  • Because profiles are built from retrievable, updateable sources, new information such as market integration can be added to the knowledge base and the resulting behavioral shifts observed, making the agents dynamic.
  • The paper positions SCAs as a complement, not a substitute, for human-subject research, aimed at hypothesis generation and protocol refinement before expensive fieldwork.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The resemblance is demonstrated only where human data already exist; the Yanomami result is a prediction, not a validation, and the paper's own Section 4 flags that a prospective field study is the natural next validation step.
  • A contamination audit would be cheap and decisive: print the retrieved chunks for each tribe and check for any sentence containing a number from the benchmark papers; if found, the resemblance is partly an echo.
  • The pipeline is portable to other experimental paradigms, such as public-goods games or risk and time preference elicitation, so the generality of the claim could be tested without new methodology.
  • The endowment-effect demonstration rests on a single Aché example; a systematic multimodal comparison across all six agents and multiple item pairs would be needed to turn that illustration into a result.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper introduces Synthetic Cultural Agents (SCAs), LLM-based agents instantiated with cultural profiles built via a Search + RAG pipeline from web sources, and reports their behavior in dictator, ultimatum, and a multimodal endowment-effect task. Agents represent six small-scale societies (Hadza, Machiguenga, Tsimané, Aché, Orma, Yanomami). The authors claim substantial cross-cultural variability, absence of purely self-interested behavior, qualitative resemblance to human experimental data where available, and novel testable hypotheses for unstudied populations such as the Yanomami. The paper includes code and data on GitHub.

Significance. If the claims were supported, the method would be a valuable low-cost tool for hypothesis generation and protocol piloting in cross-cultural experimental economics, particularly for hard-to-reach populations. Strengths include transparent Search + RAG pipeline with traceable sources, public code/data, repeated trials, and an honest discussion of limitations. However, the validation currently rests on an instrument that does not measure the human decision variables and on informal qualitative comparisons; these issues must be resolved before the methodological claim can be accepted.

major comments (4)
  1. [Appendix A.2.1, A.2.2; Tables C.3 and C.6 vs C.1 and C.2] The dependent variable in the SCA dictator game is not the human dependent variable. The prompt in A.2.1 asks "An experimenter proposes that you give at most X... Do you accept this distribution? Yes/No." Tables C.3 and Figure 2 therefore report acceptance counts for eleven exogenously proposed splits, whereas the human benchmarks in Table C.1 are mean/median/modal offers chosen by the participant. A distribution of yes/no answers over hypothetical splits is not an offer distribution; for example, the Hadza SCA's 7/100 acceptances of the $0 split in Table C.3 cannot be compared with a human modal offer of 0–10% in Table C.1. The same problem applies to UG proposers in Table C.6. Only the UG responder task, where an actual offer is accepted or rejected, is construct-comparable. This invalidates the abstract's "qualitative resemblance" claim and the specific comparisons in §B.1 and §B.2.1.
  2. [§3.3, §4, Appendix B] No formal comparison to human data is reported. The paper asserts qualitative resemblance without defining a resemblance criterion, reporting effect sizes, confidence intervals, or statistical tests comparing SCA distributions to the human benchmark distributions in Tables C.1 and C.2. The CMH tests in §3.1–3.2 and Appendix B only establish that acceptance rates differ across synthetic tribes; they say nothing about whether the synthetic agents match human behavior. The authors should specify an operational measure of resemblance (e.g., overlap of modal/median offers, distributional distance, or pre-registered "qualitative" rules) and apply it separately to each validation population.
  3. [§2.1.1, §4, Table A.2] The validation is exposed to source leakage. The profiles are built from the top k=10 Google results for queries such as "What characterizes the [tribe] tribe?", with sources including Wikipedia and National Geographic (Table A.2); the authors acknowledge in §4 that profiles are based on accounts written mostly by outsiders. The same literature contains the experimental results used as benchmarks (Henrich et al. 2001/2005; Marlowe 2004; Apicella et al. 2014), and these results may also be in the LLM's pretraining. No audit is provided to show that retrieved chunks and pretrained knowledge excluded the benchmark game outcomes. Without such an audit, the observed "resemblance" may be partly an echo of the input, and the paper's claim of independent validation is not established. I note the paper itself flags the validation limitation in §4, but the point is load-bearing for the central claim and needs a concrete remedy.
  4. [§3.3, Tables C.3–C.7, B.1] The "absence of purely self-interested behavior" and "Yanomami closer to homo economicus" findings are not supported by the current instrument. In the dictator task, acceptance counts like the Yanomami row in Table C.3 (1,0,0,1,2,0,0,0,0,0,0) describe responses to proposals, not chosen offers; a subject can reject every proposed split yet be selfish or not. Moreover, the counts are noisy and non-monotonic (e.g., Aché accepts 3/100 at 0% but 0/100 at 50% and 2/100 at 90%, Table C.3), and with temperature = 1 and 100 repetitions no confidence intervals are reported. The claim that no SCA is purely self-interested should be re-derived using choice-based measures, with appropriate uncertainty quantification, after correcting the instrument.
minor comments (6)
  1. [A.1.1, Figure A.1] The spelling "Hazda" appears in the code and search query in A.1.1 and Figure A.1; this should be "Hadza" to ensure exact replication of the pipeline.
  2. [Tables C.3, C.6, C.7] The table captions should state explicitly that each cell is the number of "yes" responses out of 100 independent trials with temperature = 1, and the independence assumptions underlying the CMH test should be discussed.
  3. [§3.1, §B.1] The abbreviation "CMHM" is used inconsistently; spell out Cochran–Mantel–Haenszel and define the reported test statistic in one place.
  4. [Data Availability] The contact email "monica.capra@cguedu" appears to be missing a period in the domain; it should be "monica.capra@cgu.edu".
  5. [§B.2] The Hadza rejection statistics are attributed to [2] in one part of the text and to [17] in another; please verify the correct citations.
  6. [Figures 2 and 3] Figures 2 and 3 are difficult to read in grayscale and lack a clear legend; please add distinguishable markers and axis labels such as "count of yes responses out of 100".

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity; validation is against external published human data, with an acknowledged same-literature leakage risk that is not shown to be realized.

full rationale

The paper's chain is empirical: Search + RAG builds cultural profiles from retrieved web text; the profiles are used as system prompts; SCAs then give binary accept/reject responses; and those responses are compared with published human results from Henrich et al. and others. No parameter is fitted to the target game outcomes, and none of the paper's equations or definitions make the predicted behavior equal to an input by construction. The discussion section does flag a limitation: 'our tribal profiles are based on articles, stories, and accounts written mostly by WEIRD people' and 'our current validation method relies on replicating patterns of behavior observed in the existing literature.' This is a genuine leakage risk, since retrieved chunks or the LLM's pretraining could contain the benchmark experiment results. However, the hard rule requires exhibiting a specific reduction, and the paper does not show that any retrieved chunk (Table A.2 lists general ethnographic sources) or profile contains the benchmark game data. The construct mismatch between the strategy-method accept/reject responses and human chosen offers is a validity problem, not a circularity problem. Self-citations (refs. 29 and 30) are peripheral Discussion background and not load-bearing. Hence no circular step is demonstrated; score 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central validation depends on assumptions about profile fidelity, prompt-task equivalence, independence of repeated draws, and absence of benchmark leakage. The paper states some of these as limitations (profile fidelity, validation method) but does not test them, and the dictator prompt equivalence is never discussed.

free parameters (4)
  • k (number of search results retrieved per tribe) = 10
    Chosen for the RAG pipeline; no sensitivity analysis is reported and profile quality may depend on it.
  • Chunk size and overlap for RAG indexing = 2000 words with 200 overlap
    Chosen ad hoc; no robustness checks are reported, and retrieval content depends on these values.
  • Temperature for experimental decisions = 1.0
    Set to ChatGPT API default; it controls response stochasticity and the 100-repeat distributions.
  • Number of repetitions per offer level = 100
    Chosen to capture within-agent variation; no statistical power justification is given, and the resulting counts are small and noisy (often 0-20 yes responses out of 100).
assumptions (5)
  • domain assumption Cultural profiles retrieved from top-k web sources (Wikipedia, National Geographic, etc.) provide a faithful representation of each tribe's norms and decision-making.
    Section 2.1.1 and Discussion state that profiles are built from web scraping of sources written mostly by WEIRD authors; the authors acknowledge this limitation.
  • domain assumption The yes/no contingent-split responses in the dictator game measure the same willingness-to-give construct as standard dictator game offers.
    Appendix A.2.1 frames the decision as accepting an experimenter's proposal to give 'at most' a stated amount, which is not the standard dictator task and could be interpreted differently by the model.
  • domain assumption The 100 repeated draws from each SCA are independent observations suitable for the Cochran-Mantel-Haenszel and chi-square tests.
    Section 2.2.3 repeats tasks 100 times and Results treat these as independent; repeated sampling from the same prompt with the same model may be autocorrelated and does not represent 100 independent human participants.
  • domain assumption The LLM's pretraining and the RAG-retrieved context do not already contain the specific experimental results used as validation benchmarks.
    Results compare SCA behavior to Henrich et al. and Apicella et al.; no source audit rules out leakage from the profile-building context or from the model's training data.
  • standard math Standard statistical tests (CMH, Fisher) are applicable to the stratified binary data.
    The paper uses them without diagnosing sparse cells or dependence; acceptable as standard tools but assumptions are not checked.
invented entities (1)
  • Synthetic Cultural Agent (SCA)
    purpose: An LLM instance conditioned on a RAG-built cultural profile, used to produce experimental choices in economic games.
    The paper postulates that a text profile plus in-context learning yields a valid proxy for human cultural behavior; the only evidence is qualitative resemblance to the same literature used to build the profiles, and no independent falsifiable prediction has been tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLMs Model Non-WEIRD Populations: Experiments with Synthetic Cultural Agents." pith.science (2026). https://pith.science/paper/NIDRZECE

@misc{pith2026250106834,
  author       = {Pith},
  title        = {Pith review of: LLMs Model Non-WEIRD Populations: Experiments with Synthetic Cultural Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NIDRZECE}},
  note         = {Machine review of arXiv:2501.06834}
}
read the original abstract

Despite its importance, studying economic behavior across diverse, non-WEIRD (Western, Educated, Industrialized, Rich, and Democratic) populations presents significant challenges. We address this issue by introducing a novel methodology that uses Large Language Models (LLMs) to create synthetic cultural agents (SCAs) representing these populations. We subject these SCAs to classic behavioral experiments, including the dictator and ultimatum games. Our results demonstrate substantial cross-cultural variability in experimental behavior. Notably, for populations with available data, SCAs' behaviors qualitatively resemble those of real human subjects. For unstudied populations, our method can generate novel, testable hypotheses about economic behavior. By integrating AI into experimental economics, this approach offers an effective and ethical method to pilot experiments and refine protocols for hard-to-reach populations. Our study provides a new tool for cross-cultural economic studies and demonstrates how LLMs can help experimental behavioral research.

Figures

Figures reproduced from arXiv: 2501.06834 by the authors.

Figure 1
Figure 1. Framework for Constructing Experiments with Synthetic Cultural Agents [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Dictator’s Count to Range of Proposed Splits by Tribe and ChatGPT [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Ultimatum Game Counts to Range of Offers by Tribe and ChatGPT [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Will Scaling Improve Social Simulation with LLMs?

    cs.CL 2026-07 conditional novelty 7.0 of 10

    Using 85 controlled and 35 public LLMs, the authors show social-simulation accuracy generally improves with compute, but some behavioral and low-resource tasks do not scale.

Reference graph

Works this paper leans on

50 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    New York: Farrar, Straus and Giroux, 2020

    Joseph Henrich.The WEIRDest People in the World: How the West Became Psychologically Pe- culiar and Particularly Prosperous. New York: Farrar, Straus and Giroux, 2020. isbn: 978- 0374173227

    work page 2020

  2. [2]

    The weirdest people in the world?

    J. Henrich, S. Heine, and A. Norenzayan. “The weirdest people in the world?” In:Behavioral and Brain Sciences33 (2-3 2010), pp. 61–83.doi: 10.1017/s0140525x0999152x

    doi:10.1017/s0140525x0999152x 2010

  3. [3]

    "Economic Man

    Joseph Henrich et al. “"Economic Man" in Cross-Cultural Perspective: Behavioral Experiments in 15 Small-Scale Societies”. In:Behavioral and Brain Sciences(2005), pp. 795–815.doi: 10.1017/ s0140525x05000142

    work page 2005

  4. [4]

    Fairness in simple bargaining experiments

    Robert Forsythe et al. “Fairness in simple bargaining experiments”. In:Games and economic be- havior 6.3 (1994), pp. 347–369

    work page 1994

  5. [5]

    An experimental analysis of ultimatum bargaining

    Werner Güth, Rolf Schmittberger, and Bernd Schwarze. “An experimental analysis of ultimatum bargaining”. In:Journal of economic behavior & organization3.4 (1982), pp. 367–388

    work page 1982

  6. [6]

    Anomalies: The Endowment Effect, Loss Aversion, and Status Quo Bias

    Daniel Kahneman, Jack L. Knetsch, and Richard H. Thaler. “Anomalies: The Endowment Effect, Loss Aversion, and Status Quo Bias”. In:Journal of Economic Perspectives5.1 (1991), pp. 193–

    work page 1991

  7. [7]

    Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?

    John J. Horton. “Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?” In: (2023). arXiv:2301.07543 [econ.GN]

    arXiv 2023

  8. [8]

    Luke Hewitt et al.Predicting Results of Social Science Experiments Using Large Language Models

    work page

Show all 50 references
  1. [9]

    A Turing Test: Are AI Chatbots Behaviorally Similar to Humans?

    Qiaozhu Mei et al. “A Turing Test: Are AI Chatbots Behaviorally Similar to Humans?” In:arXiv preprint arXiv:2312.00798 (2024)

    2024 arXiv

  2. [10]

    Qingxiu Dong et al.A Survey on In-context Learning. 2024. arXiv: 2301.00234 [cs.CL] . url: https://arxiv.org/abs/2301.00234. 13

    2024 arXiv

  3. [11]

    Argonauts of the Western Pacific: An Account of Native Enterprise and Adventure in the Archipelagoes of Melanesian New Guinea

    Bronislaw Malinowski. Argonauts of the Western Pacific: An Account of Native Enterprise and Adventure in the Archipelagoes of Melanesian New Guinea. London: George Routledge & Sons, Ltd., 1922

    1922

  4. [12]

    New York: Basic Books, 1973

    Clifford Geertz.The Interpretation of Cultures: Selected Essays. New York: Basic Books, 1973

    1973

  5. [13]

    Ofir Press et al.Measuring and Narrowing the Compositionality Gap in Language Models. 2023. arXiv: 2210.03350 [cs.CL]

    2023 arXiv

  6. [14]

    Patrick Lewis et al.Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. 2021. arXiv: 2005.11401

    2021 arXiv

  7. [15]

    Jianlv Chen et al.BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. 2024. arXiv:2402.03216 [cs.CL]

    2024 arXiv

  8. [16]

    Dictators and Ultimatums in an Egalitarian Society of Hunter-Gatherers: The Hadza of Tanzania

    Frank Marlowe. “Dictators and Ultimatums in an Egalitarian Society of Hunter-Gatherers: The Hadza of Tanzania”. In:Foundations of Human Sociality: Economic Experiments and Ethno- graphic Evidence from Fifteen Small-Scale Societies.Ed.byJosephHenrichetal.Oxford:Oxford Universit...

    2004

  9. [17]

    In Search of Homo Economicus: Behavioral Experiments in 15 Small-Scale Societies

    Joseph Henrich et al. “In Search of Homo Economicus: Behavioral Experiments in 15 Small-Scale Societies”. In:American Economic Review91.2 (2001), pp. 73–78.doi: 10.1257/aer.91.2.73. url: https://www.aeaweb.org/articles?id=10.1257/aer.91.2.73

    2001 doi

  10. [18]

    Evolutionary Origins of the Endowment Effect: Evidence from Hunter- Gatherers

    Coren L. Apicella et al. “Evolutionary Origins of the Endowment Effect: Evidence from Hunter- Gatherers”. In: American Economic Review104.6 (2014), pp. 1793–1805.doi: 10.1257/aer. 104.6.1793

    2014 doi

  11. [19]

    DoesCultureMatterinEconomicBehavior? UltimatumGameBargainingAmong the Machiguenga of the Peruvian Amazon

    JosephHenrich.“DoesCultureMatterinEconomicBehavior? UltimatumGameBargainingAmong the Machiguenga of the Peruvian Amazon”. In:American Economic Review(2000). doi: 10. 1257/aer.90.4.973

    2000

  12. [20]

    Collective action in action: Prosocial behavior in and out of the laboratory

    Michael Gurven and Jeffrey Winking. “Collective action in action: Prosocial behavior in and out of the laboratory”. In:American Anthropologist110.2 (2008), pp. 179–190

    2008

  13. [21]

    Altruistic Cooperation During Foraging by the Ache, and the Evolved Human Predis- position to Cooperate

    Kim Hill. “Altruistic Cooperation During Foraging by the Ache, and the Evolved Human Predis- position to Cooperate”. In:Human Nature 13.1 (2002), pp. 105–128.doi: 10.1007/s12110- 002-1016-3. 14

    2002 doi

  14. [22]

    Market Integration and Fairness: Evidence from Ultimatum, Dictator, and Public Goods Experiments in East Africa

    Jean Ensminger. “Market Integration and Fairness: Evidence from Ultimatum, Dictator, and Public Goods Experiments in East Africa”. In:Foundations of Human Sociality: Economic Experiments and Ethnographic Evidence from Fifteen Small-Scale Societies. Ed. by Joseph Henrich et al....

    2004

  15. [23]

    https : / / www

    The Editors of Encyclopaedia Britannica.Yanomami. https : / / www . britannica . com / topic / Yanomami. Accessed: 2024-07-10. 2024.url: https://www.britannica.com/topic/Yanomami

    2024

  16. [24]

    Lo.LLM economicus? Mapping the Behavioral Biases of LLMs via Utility Theory

    Jillian Ross, Yoon Kim, and Andrew W. Lo.LLM economicus? Mapping the Behavioral Biases of LLMs via Utility Theory. 2024. arXiv: 2408.02784 [cs.CL]. url: https://arxiv.org/abs/ 2408.02784

    2024 arXiv

  17. [25]

    2023.doi: 10.31234/osf.io/5b26t

    Mohammad Atari et al.Which Humans?Sept. 2023.doi: 10.31234/osf.io/5b26t. url: osf.io/ preprints/psyarxiv/5b26t

    2023 doi

  18. [26]

    Behavioural Development Economics: Lessons from Field Labs in the Developing World

    Jeffery Carpenter and Juan Camilo Cardenas. “Behavioural Development Economics: Lessons from Field Labs in the Developing World”. In:The Journal of Development Studies 44.3 (2008), pp. 311–338. doi: 10.1080/00220380701848327

    2008 doi

  19. [27]

    Dictatorgames:ametastudy

    ChristophEngel.“Dictatorgames:ametastudy”.In: Experimental Economics14.4(2011),pp.583–

    2011

  20. [28]

    Generation Next: Experimentation with AI

    Gary Charness, Brian Jabarian, and John A List. Generation Next: Experimentation with AI. Working Paper 31679. National Bureau of Economic Research, Sept. 2023. doi: 10 . 3386 / w31679. url: http://www.nber.org/papers/w31679

    2023

  21. [29]

    Rationality and Utility: Economics and Evolutionary Psy- chology

    C. Monica Capra and Paul H. Rubin. “Rationality and Utility: Economics and Evolutionary Psy- chology”. In: Evolutionary Psychology in the Business Sciences. Ed. by Gad Saad. Springer, 2011, pp. 319–338

    2011

  22. [30]

    The evolutionary psychology of economics

    Paul H Rubin and C Monica Capra. “The evolutionary psychology of economics”. In:Applied Evolutionary Psychology. Ed. by S Craig Roberts. Oxford: Oxford University Press, 2011

    2011

  23. [31]

    Evolutionary models of preference formation

    Ingela Alger and Jörgen W. Weibull. “Evolutionary models of preference formation”. In:Annual Review of Economics12 (2020), pp. 329–354

    2020

  24. [32]

    Is altruism evolutionarily stable?

    Helmut Bester and Werner Güth. “Is altruism evolutionarily stable?” In:Journal of Economic Behavior & Organization34.2 (1998), pp. 193–209

    1998

  25. [33]

    Self-Prompting Large Language Models for Zero- Shot Open-Domain QA

    Junlong Li, Zhuosheng Zhang, and Hai Zhao. “Self-Prompting Large Language Models for Zero- Shot Open-Domain QA”. In: (2023). arXiv:2212.08635 [cs.CL]. 15

    2023 arXiv

  26. [34]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei et al. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”. In: (2023). arXiv: 2201.11903 [cs.CL]

    2023 arXiv

  27. [35]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao et al. “ReAct: Synergizing Reasoning and Acting in Language Models”. In: (2023). arXiv: 2210.03629 [cs.CL]

    2023 arXiv

  28. [36]

    Liu et al.Lost in the Middle: How Language Models Use Long Contexts

    Nelson F. Liu et al.Lost in the Middle: How Language Models Use Long Contexts. 2023. arXiv: 2307.03172 [cs.CL]

    2023 arXiv

  29. [37]

    Exchange Asymmetries Incorrectly Interpreted as Evidence ofEndowmentEffectTheoryandProspectTheory

    Charles R. Plott and Kathryn Zeiler. “Exchange Asymmetries Incorrectly Interpreted as Evidence ofEndowmentEffectTheoryandProspectTheory”.In: American Economic Review97.4(2007), pp. 1449–1466. doi: 10.1257/aer.97.4.1449

    2007 doi

  30. [38]

    Camerer.Behavioral Game Theory: Experiments in Strategic Interaction

    Colin F. Camerer.Behavioral Game Theory: Experiments in Strategic Interaction. Princeton Uni- versity Press, 2011

    2011

  31. [39]

    A selfish origin for human food sharing: Tolerated theft

    N. G. Blurton Jones. “A selfish origin for human food sharing: Tolerated theft”. In:Ethology and Sociobiology 5.1 (1984), pp. 1–3.doi: 10.1016/0162-3095(84)90030-X

    1984 doi

  32. [40]

    Costly punishment across human societies

    Joseph Henrich et al. “Costly punishment across human societies”. In:Science 312.5781 (2006), pp. 1767–1770. doi: 10.1126/science.1127333. 16 Data Availability

    2006 doi

  33. [44]

    This repository is publicly accessible and includes detailed README files for navigation and usage

    Repository: The Python code for creating synthetic cultural agents (SCAs), the interac- tive platform, and the data generated from SCA simulations are available in our GitHub repository: https://tinyurl.com/mr4yx3tj. This repository is publicly accessible and includes detailed...

  34. [45]

    Data Types: The repository contains raw SCA output data, processed experimental data, configuration files, and source code for the interactive platform

  35. [46]

    Large Language Models:We used GPT-3.5 turbo for creating SCAs, and GPT4, GPT4o, and Claude 3 Opus for creating the profiles

  36. [47]

    License: All data and code are freely available for academic use under the MIT License

  37. [48]

    Supplemental Data: Tables containing data on the number of observations for each pro- posed endowment split in the Dictator and Ultimatum Games can be found in the Supple- mental Materials (Appendices)

  38. [49]

    Data Sufficiency: All data supporting the findings of this study are available within the paper and its supplementary materials, ensuring full replicability of the results

  39. [50]

    parameterized

    Additional Information: Readers are encouraged to contact the corresponding author (monica.capra@cguedu) for any additional data-related queries or specific requests. For Online Publication LLM Experiments with Synthetic Tribesmen Appendix A Methods A.1 Build Cultural Profile ...

    2024

  40. [206]

    doi: 10.1257/jep.5.1.193

    doi

  41. [610]

    doi: 10.1007/s10683-011-9283-7

    doi

  42. [2024]

    io / dl / Predicting % 20results % 20of % 20social % 20science % 20experiments%20using%20large%20language%20models.pdf

    url: https : / / samim . io / dl / Predicting % 20results % 20of % 20social % 20science % 20experiments%20using%20large%20language%20models.pdf

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.