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

Framework

• LLMs can serve as Homo silicus: computational models given endowments, information, preferences, and scenarios whose behavior is explored through simulation rather than deduction.Their natural-language flexibility allows prompts to be adjusted, including with conventional economic theory, and newer methods can improve fidelity to human responses.Source: §1 · p. 2

Conceptual issues

• LLMs capture social information from economic activities, decision-making heuristics, common social preferences, and scientific research encoded in training data, whose sources are not perfect.Source: §1 · pp. 2–3

Simulated experiments

• The paper presents five simulated experiments selected because they are simple to describe and implement, with clear qualitative results that can be compared to AI outcomes.They include experiments motivated by Kahneman et al. (1986) and simulations exploring Oprea (2024b).Source: §1 · pp. 3–4

Kahneman et al. (1986)

• Large price increases are judged more negatively, right-leaning agents are more sanguine about gouging, and these effects remain robust across dozens of experimental permutations.The study varies both the price increase and respondents’ political leanings, extending Kahneman et al. (1986).Source: §1 · p. 3

Charness and Rabin (2002); Samuelson and Zeckhauser (1988)

• Endowed social preferences steer dictator-game choices toward equity, efficiency, or narrow self-interest; without endowment, some models choose efficiently while others are selfish, enabling more human-like two-stage agents.In framing experiments, some LLMs strongly favor whichever budget option is presented as the status quo, whereas others show little status quo bias.Source: §1 · pp. 3–4

Oprea (2024b); Horton (2025)

• Simulations provide suggestive support for Oprea’s (2024b) complexity-driven interpretation of risk attitudes, while minimum wages shift hiring toward more experienced applicants and higher hourly wages.The hiring result is directionally consistent with Horton (2025) across LLMs with varying capabilities.Source: §1 · p. 4

Methods and reproducibility

• The paper updates its experiments across contemporary open-source and proprietary LLMs and provides an open-source package intended to make simulations easy to replicate, share, update, and scale.The package supports button-push reproduction with almost any LLM and enables researchers to explore new models or ideas with minimal technical overhead.Source: §1 · pp. 4–5

Applications and limitations

• LLM simulations can pilot experiments cheaply, explore parameter and wording sensitivity, generate realistic-looking data, and inform power calculations, but require empirical confirmation (Ludwig et al., 2025).The paper reports that its experiments ran in minutes for a trivial amount of money.Source: §1 · p. 5
• Prompting AI agents with social-science theory can dramatically improve predictive capabilities (Manning and Horton, 2025), and simulations can guide empirical work or theory-building.The paper frames simulation as an engine of discovery, analogous to economists’ use of toy models to identify promising patterns.Source: §1 · p. 5

Related literature

• Since 2023, an extensive and rapidly growing literature testing AI agents, including thousands of experiments, spans every social-scientific discipline, with some simulations enabling previously infeasible studies of humans.Source: §1 · p. 5

Paper contributions

• The paper’s contributions are to help economists incorporate evolving AI-simulation practices and to connect LLM experimentation to economics’ research paradigm and foundational assumptions such as rationality.It argues that LLM experimentation is more akin to economic theory than it may superficially appear.Source: §1 · p. 6

2.1 Fairness as a constraint on profit-seeking (Kahneman et al., 1986)

• AI agents reproduce political differences and a dose-response pattern in fairness judgments: left-leaning agents find price hikes less fair, while larger hikes are generally judged less fair.These results recapitulate and extend Kahneman et al. (1986), whose original subjects judged 82% of price-gouging cases unfair or very unfair.Source: §2.1 · p. 7, Figure 1 · p. 8

2.1.1 Testing generalizability and possible memorization

• The experiments use prompt-endowed political personas, multiple languages, scenario variations, and adversarial prompts to test whether fairness patterns generalize beyond the snow-shovel vignette.The design addresses concerns that memorization of the original study or political associations could make the initial simulation brittle.Source: §2.1.1 · p. 9
• Across translated, alternative, adversarial, and original prompts, right-leaning agents generally rate price gouging as fairer, and the relationship cannot be inverted.The translated version shows some category-specific differences, while a $1 price hike reduces fairness ratings by 0.004–0.009 points across permutations.Source: §2.1.1 · pp. 10–11

2.2 A social preferences experiment (Charness and Rabin, 2002)

• The section next introduces Charness and Rabin (2002) as a social-preferences experiment involving choices between allocations with trade-offs.The provided passage establishes the experiment’s allocation trade-offs but does not report its findings here.Source: §2.2 · p. 11
• Persona-less AI agents differed substantially from human participants, with baseline behavior varying across models: Llama-3-70B appeared more selfish, while Gpt-4o, Claude-Sonnet-3.5, and Deepseek were more efficiency-minded.These differences persisted despite the original experiment likely being present in the models’ training data.Source: §2.2 · pp. 12–13, Figure 3 · p. 12
• Theory-grounded personas produced disciplined, theory-consistent choices: efficient agents favored maximum combined payoffs, inequity-averse agents minimized payoff differences, and self-interested agents maximized their own payoffs.A few exceptions occurred, including Deepseek and Llama-3-70B failing to select the inequity-averse option in the spiteful Berk23 scenario.Source: §2.2 · p. 13

2.2.1 Constructing a calibrated sample of AI agents

• Off-the-shelf persona-less agents may be useful for low-stakes testing but have poor predictive reliability for empirical applications, whereas calibrated theory-based agents can support prediction in novel settings without prior human data.The supplied status quo-bias subsection only states that Samuelson and Zeckhauser (1988) studied decision-making scenarios in which people were more likely to retain the status quo.Source: §2.2.1 · p. 13
• The calibration method constructs an optimized nonnegative mixture of self-interested, inequity-averse, and efficient agent types whose combined choices best reproduce aggregate human choices, minimizing MSE subject to weights summing to 1.Source: §2.2.1 · pp. 13–14
• Calibrated persona mixtures reduced out-of-sample MSE to 0.094 versus 0.182 for persona-less controls, indicating closer alignment with human choices across sequential games.The weighted samples’ MSE was about half that of the persona-less controls, supporting transfer across structurally distinct game formats.Source: §2.2.1 · p. 15, Figure 4 · p. 15

2.3 Status quo bias in decision-making (Samuelson and Zeckhauser, 1988)

• The status-quo replication creates 24 AI agents from twelve safety beliefs and car-ownership identities, exposing each to four framings across four funding allocations and five LLMs.Agents make repeated choices without memory unless prior decisions are explicitly supplied, enabling a within-subject design that would be difficult with human subjects because they could detect repeated framing manipulations.Source: §2.3 · p. 16
• AI agents exhibit strong status quo bias: they shift preferences toward the status quo, and allocations farther from it are less likely to be chosen across all models.The status-quo specification fits the data slightly better than the distance specification, while Gpt-4o exhibits the bias to a lesser extent.Source: §2.3 · pp. 17–18

2.4 A complexity view of prospect theory (Oprea, 2024b)

• The experiment extends Oprea’s central experiment by endowing each AI agent with $50 and eliciting certainty equivalents for 100 box sets with probabilities p∈{0.05, 0.10, . . . , 0.95}.It uses three math-ability personas, plus a baseline agent, and runs each agent 10 times on five LLMs at temperature 1, generating 15,200.Source: §2.4 · p. 19
• Across models, lottery and mirror conditions generally move in lockstep: both show the characteristic X-shape or both are flat at the expected value, with no lottery-only fourfold pattern.Less advanced models—Gpt-3.5-Turbo and Claude-Haiku-3—show pronounced deviations between small and large probabilities.Source: §2.4 · p. 19
• The AI recapitulation supports Oprea’s claim that the fourfold pattern is not a special phenomenon of risk, appearing consistently across five models and thousands of simulations.The figure reports mean certainty equivalents with 95% confidence intervals across LLMs and mathematical-ability assignments.Source: §2.4 · p. 20
• More capable models select expected value across conditions, whereas assigning them to be “very bad at math” produces an increasingly pronounced fourfold pattern aligned with prospect theory.The authors find no robustly reported relationship that would explain this pattern as a default learned application, and variation across personas makes confusion with lotteries unlikely.Source: §2.4 · p. 21
• The authors’ preferred explanation is that LLMs learned systematic relationships from examples of humans making decisions under complexity, though human data remain necessary for definitive conclusions.This interpretation is consistent with the similarity between lottery and mirror conditions and with stronger prospect-theoretic responses among agents worse at math.Source: §2.4 · p. 21

2.5 Labor-labor substitution under a minimum wage (Horton, 2025)

• A minimum wage increased the hired worker’s wage by about $1.3/hour without a reference wage, and the effect grew by nearly 50% when a reference wage was provided.The regressions included model fixed effects and standard errors clustered at the model level; hiring was mandatory for the AI employer.Source: §2.5 · p. 22
• The minimum wage also increased hired-worker experience by about a month without a reference wage and by roughly three months with one, consistent across dozens of independently developed models.These results reproduce the labor-labor substitution margin reported by Horton (2025), with the reference wage strengthening the experience effect.Source: §2.5 · pp. 22–23

3.1 Why LLMs can simulate human behavior

• LLMs draw on social-world information through two paths: explicitly memorized social-science findings and implicitly ingested behavior captured in human-generated text.The implicit path includes communication, decisions, interactions, and perceptions represented across non-scholarly sources.Source: §3.1 · pp. 24–25
• Because both information streams are filtered and incomplete, AI simulations should not be treated as unequivocal evidence about human behavior.Peer review, self-presentation, and proprietary training-data curation shape what enters the model, creating trade-offs between capability and openness.Source: §3.1 · pp. 25–26
• LLMs learn generalizable human behavioral patterns from rich training data, transformer architectures, and post-training processes that translate context into predicted actions.Their training data encode social-science findings and human behavior in action, while post-training steers responses toward cooperative, instruction-following behavior.Source: §3.1 · pp. 23, 26

3.2 Homo silicus as theory

• Homo silicus simulations are more akin to economic theory when they formalize mechanisms, generate testable predictions, and reveal directions for further modeling.Unlike a single baseline model, simulations can explore many empirically informed specifications without requiring each equilibrium to be derived analytically.Source: §3.2 · pp. 26–27
• Even when simulations mislead, they can make expensive data collection more useful by targeting efforts where the marginal value of information is highest.A simulated job seeker accepting lower startup wages because of layoff risk, promotion timelines, wage growth, and insurance value illustrates how simulations can suggest mechanisms and empirical designs.Source: §3.2 · p. 27
• Homo silicus can generate predictions in virtually any natural-language setting, unlike conventional models constrained by fixed parameters, structural assumptions, or training domains.LLMs can adapt a prompt’s setting and extend mappings to structurally distinct games without retraining, although they may fail to generalize unexpectedly.Source: §3.2 · pp. 27–29
• Economic theory supplies explicit instructions that LLMs execute flexibly, making assumptions, objectives, and decision rules transferable and auditable across new settings.The approach supports checking behavior on simple test cases before scaling simulations, and calibrated theoretical personas predicted human responses better than persona-less LLMs.Source: §3.2 · p. 29
• Improved predictive accuracy suggests that LLMs encode meaningful relationships between interpretable theory and human behavior, while atheoretical traits generally do not improve fidelity.Manning and Horton (2025) find that hobbies or favorite television shows do little to improve fit, whereas theory-based social-preference personas do; these patterns remain structured signals rather than proof of human mechanisms.Source: §3.2 · p. 30

3.3 Using Homo silicus to navigate uncharted territory

• Figure 7 frames AI simulations as part of the relationship among the social world, social-science knowledge, and propositions studied in the literature.The figure distinguishes true propositions about the social world from propositions that have been studied and recorded in social science.Source: §3.3 · p. 30
• Figure 7 frames AI simulations as a confusion matrix: they help when supporting true or rejecting false propositions, but amplify errors through false positives or false negatives.Simulations provide a selective, noisy slice of the social world rather than definitive truth.Source: Figure 7 · p. 31, §3.3 · p. 31
• Prediction-powered inference can combine a ground-truth human sample with model predictions to produce unbiased, more precise estimates at lower data-collection cost.The method trains outcome predictors on observed covariates and corrects their predictions using the ground-truth sample.Source: §3.3 · p. 32
• For novel settings without ground truth, researchers should use theory-grounded agents validated on related human benchmarks, because benchmark accuracy supports—but cannot guarantee—nearby generalization.Charness and Rabin allocation-game experiments illustrate this approach, and Manning and Horton (2025) show improved prediction in entirely novel settings.Source: §3.3 · pp. 32–33
• Simulation reliability is expected to be higher in domains richly represented in training data, including text-captured laboratory psychology and digital marketing experiments.Trusting simulations in settings well captured by internet text is presented as a sound heuristic, although the training corpus’s boundaries remain uncertain.Source: §3.3 · p. 33
• Fine-tuning can improve simulation fidelity, but it is unwieldy for ideation and may cause catastrophic forgetting of information from the original training data.The passage contrasts fine-tuning with effective prompting and theory-grounded agent design.Source: §3.3 · p. 33

4 Conceptual issues

• Following Friedman (1953), the authors argue that Homo silicus should be evaluated by whether it generates useful results, especially by accelerating human knowledge.They nevertheless examine non-Friedman responses to common critiques of AI simulations.Source: §4 · p. 34

4.1 The “garbage in, garbage out” critique

• LLMs are not simple weighted averages of noisy training data, but their corpus may overrepresent “humans creating public writing” and stated rather than revealed behavior.The authors argue that internet text includes implicit and explicit social information, including reasoning, intentions, decisions, purchases, and ratings.Source: §4.1 · pp. 34–35

4.2 “Memorization” and “performativity” problems

• LLM behavior contradicts simple memorization because models differ from humans on scenarios in their training data and can generalize flexibly like a good student.Performativity may also benefit simulations when internalized theories and empirical facts help predict human behavior, assuming those theories are correct.Source: §4.2 · pp. 35–36

4.3 Black boxes with incoherent world models

• LLMs need not have fully coherent world models to produce credible predictions, because instruction-following can support simpler, perturbation-robust tasks than reconstructing reality.The New York City routing example shows brittle inferred maps that omit streets, invent others, and perform worse under detours, while broader scaling effects may improve capability.Source: §4.3 · pp. 36–37

4.4 Are these “just simulations?”

• Unlike conventional agent-based models, Homo silicus is not under direct researcher control, yet natural-language prompts flexibly map beliefs and preferences to purposeful simulated actions.Researchers can influence the model with endowments, but behavior remains determined by the underlying model.Source: §4.4 · p. 37
• Flexible LLM reasoning makes Homo silicus less vulnerable to the Lucas critique (Lucas, 1976), but weak world models can reduce persona prompts to caricatures.LLM agents can reason about environmental changes rather than follow hard-coded behavioral rules, whereas prompts such as “respond as a 30-year-old male” may fail when complex identity interactions are poorly represented.Source: §4.4 · p. 37

4.5 It is easy to “prompt hack” AI experiments

• Because AI experiments are easy to run, researchers may iterate prompts until obtaining favorable results, creating a problem conceptually similar to p-hacking.Robustness checks using prompt permutations, temperatures, languages, and contexts, together with public simulation code, improve transparency; replication can require only a few dollars and one click.Source: §4.5 · p. 38

4.6 Causal inference with AI simulations is challenging

• AI simulations face causal-identification challenges because editing a prompt to change one factor may inadvertently cause other factors to change; in the hiring experiment, the hired worker’s wage and experience increased.Source: §4.6 · p. 39

• Experiments with dozens of LLMs recover core qualitative patterns from analogous human-subject experiments while surfacing deviations informative for future research.Source: §5 · p. 40
• Homo silicus is best understood as theory in flexibly executable form, enabling fast, low-cost exploration, design stress tests, and power-calculation calibration.Source: §5 · p. 40
• Open-source software standardizes AI experiments, supports model switching and updates, and makes the paper’s experiments button-push reproducible with extensive documentation.Source: §5 · p. 40
• The key limitation is the missing map of when AI simulations track Homo sapiens, requiring researchers to identify high- and low-fidelity settings candidly.Source: §5 · pp. 40–41
• AI simulations are most informative when mechanisms are clearly instructable, domains are well represented in digital text, and related human-data benchmarks are available.Source: §5 · p. 41
• Automated scientific loops could propose hypotheses, instantiate agents, run and score simulations, and iterate, potentially accelerating research productivity (Manning et al., 2024).Source: §5 · p. 41

• Figure A1 shows Charness and Rabin (2002) Simple Tests choices by model type, endowed “personality,” and framing.The figure reports the fraction of AI agents choosing each option.Source: Figure A1 · p. 48, A Figures from the original draft · p. 48
• Figure A2 shows AI-agent opinions on Kahneman et al. (1986)’s price-gouging snow-shovel question under endowed political views and different scenarios.The figure reports the fraction of agents choosing each opinion.Source: Figure A2 · p. 49, A Figures from the original draft · p. 49
• Figure A3 shows the distribution of preferred car-safety budgets by status-quo framing.The figure reports the fraction of AI agents choosing each option by framing.Source: Figure A3 · p. 49, A Figures from the original draft · p. 49
• Table A1 reports the effects of imposing a minimum wage on the hired worker’s observed wage and experience.Source: Table A1 · p. 50, A Figures from the original draft · p. 50

• Pre-training uses massive text datasets to learn a conditional probability distribution over subsequent text, shaped by architectural and hyperparameter choices.Source: B A brief primer on large language… · p. 51
• LLMs generate responses by repeatedly predicting the most likely next word or token from the prompt and previously generated text.Tokens may be sentences, words, word parts, or characters.Source: B A brief primer on large language… · p. 51
• RLHF post-training uses human evaluations to train a reward model and adjust LLM parameters toward responses receiving higher predicted feedback.Supervised fine-tuning can additionally teach prompt-following or a desired writing style.Source: B A brief primer on large language… · p. 51
• Fully trained LLMs optimize both next-word prediction and reward-model feedback, while further fine-tuning can make responses more consistent with a persona or writing style.Source: B A brief primer on large language… · p. 51
• Temperature controls response sampling: at zero the LLM deterministically selects the highest-probability response, whereas higher values make outputs more stochastic and uniform.Source: B A brief primer on large language… · p. 51

• Linked Jupyter notebooks provide instructions, comments, and sharing or download links for replicating simulations and exploring new ideas.The notebooks share a common structure and include the Kahneman et al. (1986) experiment.Source: C Additional Figures and Tables · p. 52, Figure A4 · p. 52
• Table A2 reports regression estimates of how political beliefs and snow-shovel price increases affect AI agents’ fairness assessments across four Kahneman et al. (1986) experiment permutations.Fairness is measured continuously from 1 (“Very Unfair”) to 4 (“Completely Fair”), with Moderate as the reference political-belief class and $15 as the price baseline.Source: Table A2 · p. 53, C Additional Figures and Tables · p. 53
• The supplemental regressions cover baseline, translated-prompts, prompt-variations, and adversarial-prompt specifications, with significance indicated at ∗p<0.05.The Δ Price coefficient measures the fairness-rating change for each $1 increase above the $15 baseline.Source: C Additional Figures and Tables · p. 53