Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Social Simulation and Generative Agents

Abstract. Social simulation studies how specified actors and interactions generate collective behavior; generative agents extend this tradition through language models, memory, planning, and dynamic social environments. This fundamental research monograph traces the field from early agent-based explanations of segregation and cultural diffusion to contemporary interactive communities and synthetic-respondent research. It critically integrates original studies, methodological protocols, peer-reviewed experimental forecasts, survey-validation research, cross-national economic-preference comparisons, and independent public-opinion audits through October 2026. The synthesis establishes distinct standards for behavioral plausibility, construct validity, statistical calibration, causal identification, population representativeness, and external transport. Evidence supports useful scenario generation and some experimental-effect forecasting, while documenting substantial limitations in distributional heterogeneity, cultural correspondence, subgroup representation, and reproducibility. A research protocol for transparent, human-benchmarked generative social simulation is developed as a methodological synthesis of the cited literature, rather than a claim of new experimental data or a preregistered systematic review.

Keywords: social simulation; generative agents; agent-based modeling; computational social science; large language models; synthetic respondents; social interaction; multi-agent systems; experimental forecasting; construct validity; statistical realism; population representativeness; human–artificial relations; computational psychology

The Research Object: Social Simulation and Generative Agents

Social simulation is a method for investigating how specified actors, rules, environments, and interactions produce patterns over time. Generative agents are computational actors whose behavior is assembled through learned generative models, frequently large language models, sometimes combined with explicit memory, planning, retrieval, and action interfaces. The psychological research object is the relationship between the model's representations and observable social behavior: how simulated choices depend on cues, roles, histories, incentives, network position, and interactions with other actors. A convincing conversation and a valid representation of a population are distinct achievements, assessed against different criteria. The central scientific task is to specify which human behavior an artificial system is intended to reproduce, under which conditions, and with what independently observable error Epstein (1999) Park et al. (2023) Argyle et al. (2023).

This monograph examines single-agent synthetic respondents, interactive multi-agent environments, artificial social networks, and models that attempt to estimate responses to interventions. Its organizing questions concern explanatory architecture, psychological construct validity, empirical transport across settings, dependence among agents, and the ethics of using synthetic voices. The investigation follows an evidence-based critical narrative approach: original peer-reviewed work and methodologically transparent institutional reports provide the empirical record, while historical modeling literature supplies concepts for mechanisms and validation. The study does not claim newly collected participants, new experimental results, preregistered searches, or meta-analytic estimates. Differences between simulated plausibility, statistical calibration, experimental forecasting, and genuinely observed human behavior govern every chapter Dillion et al. (2023) Bisbee et al. (2024) Chen et al. (2026).

The boundary of the dossier is specific. Cognitive architectures, memory systems, social cognition, and decision making enter where they explain an agent's observable contributions to a social model; their general theories belong to neighboring research objects. The focus here is the inferential chain from an artificial actor through a constructed social setting to a claim about people or institutions. A single language-model output may serve as a hypothesis generator, a scenario probe, or an input to a larger causal simulation. Each use requires its own evidence. The research program becomes scientifically cumulative when its authors state the target population, operational variables, process assumptions, model version, prompt protocol, human comparator, and the uncertainty surrounding any inferred social conclusion Grimm et al. (2006) Xie et al. (2026).

Definitions, Levels, and Forms of Artificial Social Modeling

At the microscopic level, a model specifies state variables and transition rules for actors; at the mesoscopic level it generates relationships, conversations, institutions, or group formations; at the macroscopic level it describes distributions or aggregate trajectories. These levels can be linked but cannot simply be exchanged. A simulation that generates realistic-sounding individual biographies might fail to produce realistic associations among variables, and a model calibrated to a national average might misrepresent the individuals composing that average. Scientific language should therefore distinguish an agent-based generative mechanism from a synthetic survey respondent, a conversational role from an empirical person model, and a simulation of a possible future from a calibrated prediction of a particular future Epstein (1999) Xie et al. (2026).

A rule-based agent implements explicitly specified behavioral procedures; a reinforcement-learning agent adapts according to a reward structure; a language-model agent samples or selects outputs conditional on prompts and context. Hybrid designs join these mechanisms. The choice determines what can be interpreted. If cooperation arises from fixed decision rules, the result establishes the implications of those rules under the chosen environment. If it arises from language model interaction, the result additionally depends on training distributions, framing, decoding, and instructions. Without separating these sources, an apparent social mechanism can be an artifact of linguistic convention or hidden model preferences. The appropriate baseline is the simplest model that can account for the same observable pattern Schelling (1971) Shanahan et al. (2023).

The distinction between generative and predictive success is fundamental. A generative explanation aims to show how a specified process is capable of producing a pattern. Prediction requires that forecasts from one dataset or situation generalize to observations that were unavailable when the forecast was created. Description concerns a faithful account of an existing distribution, and causal inference concerns differences under interventions. An agent simulation can perform one of these tasks well and another poorly. Researchers must declare the intended inference before comparing a model with people. Formal descriptions, parameter sweeps, out-of-sample data, and experimentally manipulated environments make this declaration testable Grimm et al. (2010) Ashokkumar et al. (2026).

Historical Foundations: From Microrules to Generative Social Science

Long before language-model agents, computational social science investigated how local decisions create collective regularities. Thomas Schelling's dynamic segregation models demonstrated that the composition of neighborhoods could emerge from individual location choices even when the assumed preference was not a demand for complete segregation. The result was a carefully constructed possibility claim about local interaction and sorting, not a measurement of the motives of actual residents. Schelling's example remains pivotal because it exposes a recurring danger in interpretation: a model reproducing an observed outcome does not uniquely identify the historical cause that generated it. Alternative mechanisms can produce the same end state Schelling (1971).

Robert Axelrod subsequently modeled dissemination of culture through interactions conditioned on similarity. The model showed how local convergence could coexist with persistent global differentiation under specified assumptions. Joshua Epstein developed the methodological argument for generative social science: articulate actors and interaction rules, run the model, and explain population patterns through constructive dynamics. This tradition established replicable formal mechanisms, controlled perturbations, and transparent assumptions as core requirements. It also established the central limitation of simulations: demonstrating that a mechanism can generate an observation does not by itself establish that the mechanism operated in a historical population. Evidence external to the simulation remains decisive Axelrod (1997) Epstein (1999).

Language models changed the expressive surface of artificial social actors. They can answer open-ended questions, adapt utterances to apparent roles, synthesize remembered interactions, and propose plans without researchers hand-coding each conversational branch. Yet the classical lessons remain. Mechanisms still need documentation, alternatives still need comparison, and the target population still needs an independent measurement. The generative-agent literature expanded the space of possible simulated interactions; it did not invalidate the empirical disciplines of sampling, construct definition, model identification, or falsifiable prediction. A historical account therefore connects computational emergence, agent-based reporting protocols, and the empirical validation of contemporary language-model systems Grimm et al. (2006) Park et al. (2023) Bisbee et al. (2024).

Emergence, Interaction, and Collective Dynamics

An emergent outcome is a pattern at the collective level that depends on interaction among components and is not represented as an explicit goal of each component. Networks can amplify local imitation, segregate information, concentrate attention, or sustain different behavioral conventions across otherwise similar groups. What matters experimentally is whether the observed aggregate arises from the interaction topology, the actors' decision procedures, a common external signal, or an imposed system objective. Researchers should rerun counterfactual configurations with altered network edges, memory horizons, incentives, and communication opportunities. These interventions isolate which conditions are sufficient for the pattern in the model and reveal whether it survives reasonable variations Axelrod (1997) Centola & Macy (2007).

Social contagion research illustrates the relevance of complex reinforcement. In some settings, adoption depends on confirmation from multiple contacts rather than a single transmitted exposure. Centola and Macy's formal analysis of complex contagions helped establish why a network's bridging structure can facilitate simple information spreading while impeding practices that require repeated social endorsement. A generative multi-agent simulation may exhibit imitation, norm formation, or polarization, but these labels must refer to measured interaction sequences. The model should record who influenced whom, when actors changed their responses, and how that change compares with a no-interaction control. Agreement arising independently from shared prompting is not evidence of interpersonal transmission Centola & Macy (2007).

Real-world political exposure provides another caution. Bail and colleagues' field experiment manipulated exposure to opposing political views on social media and documented a context-specific increase in polarization among some participants. That observation concerns assigned exposure in a human sample, with the study's recruitment, measurement, and platform constraints. A synthetic platform that reproduces a similar polarization curve is consistent with one possible explanation, but it remains necessary to determine which assumed recommender rules, agent attributes, and communication channels produce the pattern. Human experiments, observational network traces, and theoretical simulations contribute complementary kinds of evidence rather than interchangeable proofs Bail et al. (2018) Jeon et al. (2025).

Psychological Constructs and the Problem of Operationalization

Psychological terms such as identity, preference, conformity, trust, altruism, loneliness, and belief require operational definitions before entering agent simulations. A survey answer expressing trust is a linguistic action within a prompt; an incentivized transfer of resources is a different behavior; a stable tendency across situations is a further inferential claim. The model may generate each form, but the construct must be connected to independent human indicators through a defensible measurement model. Researchers must specify the temporal horizon, response scale, context dependence, and criterion behavior for each construct. Otherwise, apparently coherent agents can conflate words about dispositions with dispositions inferred from conduct Dillion et al. (2023) Chen et al. (2026).

Personas often encode age, occupation, nationality, education, and prior opinions. These attributes may be useful predictors when grounded in a real sampling frame, but they do not deterministically define an individual's views. Human differences exist within every demographic category, and intersections among attributes cannot be reconstructed faithfully through stereotypical biographies. The 2025 study by Wang, Morgenstern, and Dickerson demonstrates concrete misrepresentation and flattening when language models are used as substitutes for participants from identity groups. A responsible operationalization treats category membership as measured or disclosed background information and tests whether the model recovers the variation that exists within the category Wang et al. (2025) Henrich et al. (2010).

The appropriate model of a psychological construct includes its observable indicator, expected correlations with other variables, expected sensitivity to experimental manipulation, and boundary conditions. A synthetic respondent that gives plausible answers on isolated items might still violate known within-person consistency patterns or cross-variable structures. Measurement evaluation therefore needs more than grammatical judgments. It should compare response distributions, test-retest structure, associations, temporal trajectories, and actual behavior whenever the construct purports to explain behavior. The resulting assessment may support use as a scenario model while rejecting use as a substitute respondent sample, a conclusion that reflects methodological specificity rather than failure of the entire simulation enterprise Xie et al. (2026) Bisbee et al. (2024).

Language Models as Generative Components of Social Agents

A contemporary generative agent typically contains a language model that interprets a prompt or state description and proposes a response. The system may also include persistent records, search, a separate planner, environmental state, and an action executor. The language model contributes linguistic and conceptual generalization; the surrounding software determines which information is visible, what actions are available, and whether history survives beyond one exchange. Consequently, the same underlying model may behave differently across agent architectures. Researchers should describe both components when attributing success, and should measure whether the behavior persists after removing the retrieval, planning, or messaging layer. A result belongs to a complete configuration until the contribution of its components has been experimentally isolated Park et al. (2023) Shanahan et al. (2023).

The statistical basis of a language model matters for social inference. Its outputs are conditioned on learned patterns in corpora and subsequent training, alignment, and prompting. A prediction that sounds socially natural may reflect common narrative conventions more strongly than a rare lived experience. Language models can display strong knowledge of public descriptions of society while making systematically compressed predictions about its distributions. Training corpora also reflect uneven linguistic, cultural, and institutional representation. A simulation must therefore investigate training-data overlap, the distinction between public textual descriptions and reported behavior, and the possibility that several superficially independent simulated agents draw on the same correlated model prior Bender et al. (2021) Chen et al. (2026).

Agent behavior is further shaped by decoding settings, safety instructions, order of messages, tool outputs, and model revisions. These details belong to the experimental apparatus. A reproducible study archives the exact prompt templates, system instructions, retrieval configuration, model identifier, decoding controls, available tools, and environment rules. It distinguishes a refusal triggered by provider policy from a human-like social preference and a generic cooperative answer from an inference grounded in a personal history. Comparative work should test models with equivalent information budgets and task affordances. Without this accounting, differences attributed to agent reasoning may reflect unobserved differences in interface construction Grimm et al. (2010) Weidinger et al. (2022).

Persona Construction, Identity Conditioning, and Stereotype Risk

A persona is a constructed set of descriptors, memories, goals, or behavioral instructions supplied to an artificial agent. Some personas summarize real interview material, whereas others are invented from demographic templates. This provenance is central. A real interview can constrain an agent's possible responses while introducing privacy and consent obligations; an invented biography provides a convenient experimental condition but no evidence of representing an actual person. Persona consistency across responses may be a property of persistent instructions, not a demonstrated reconstruction of individual psychology. The scientific report should identify how each attribute was obtained, which features were inferred, and whether any person was directly represented in the training or prompt context Park et al. (2024, preprint) Wang et al. (2025).

Research on simulated human samples supplies two complementary findings. Argyle and colleagues showed that language models conditioned on demographic or political information could reproduce some patterns in survey responses, motivating the concept of algorithmic fidelity. Bisbee and colleagues later demonstrated that synthetic opinion responses could show compressed variance, unstable regression estimates, sensitivity to wording, and drift over time despite reasonable aggregate averages. Both can be true because mean matching, subgroup gradients, and faithful individual heterogeneity are separate properties. Scientific evaluation must therefore establish which population moments and relationships are supported rather than treating a broad claim of persona realism as a single validated result Argyle et al. (2023) Bisbee et al. (2024).

Identity attributes are socially situated and context-dependent; they cannot be used as substitutes for the voices of marginalized populations. Wang and colleagues' four-model study, involving 3,200 human participants across 16 demographic identities, provides direct evidence that synthetic substitution can misportray and flatten the distribution of group experience. A well-designed persona experiment measures within-group errors and intersections among attributes, checks how identities are described, and reports failures without converting stereotypes into group facts. More detailed prompting may lessen a problem without eliminating its structural source. Models should be used to generate testable hypotheses about social perception, and human respondents remain the standard for claims about their own views Wang et al. (2025).

Memory, Reflection, and Temporally Extended Interaction

Social behavior unfolds through accumulated interaction. An agent may need to recall a promise, recognize a previous collaborator, revise an expectation after betrayal, or notice that a local norm has changed. Systems using external memory stores can retrieve selected past observations into the language model's current context, while reflection procedures synthesize higher-level descriptions from recorded events. Park and colleagues' generative agents architecture integrated observation, natural-language memory, reflection, and planning in a small interactive town. Its demonstrations and controlled component comparisons established an influential design pattern for temporally coherent simulations, though they do not establish general correspondence between simulated personalities and independently sampled human populations Park et al. (2023).

Several distinct mechanisms can be hidden under the term memory. Storage describes persistent records; retrieval concerns which records are reintroduced; updating changes those records; reflection summarizes, compresses, or reinterprets past events; and planning uses selected representations to guide future choices. Each operation can affect social trajectories. For example, an agent may appear forgiving because a conflict was omitted from retrieval rather than because its model of the relationship changed. Longitudinal research should log retrieved records, source timestamps, overwrite events, and traces linking an utterance to prior information. Such instrumentation supports causal interventions on memory availability and helps distinguish narrative continuity from operational continuity Park et al. (2023) Grimm et al. (2006).

Memory also introduces contamination and path dependence. An agent may read its previous generated output as if it were an independent historical event, consolidating an initially arbitrary assertion through repeated reflection. Multiple agents exchanging summaries can amplify the same invented event into apparent consensus. The research design should label simulated observations as generated data, maintain immutable source records, and separately store externally validated facts. Tests should include seeded memory errors, selective omissions, resets, and delayed interactions to assess whether the system handles social history robustly. These procedures make temporal realism assessable, while preventing the preservation of a transcript from being mistaken for evidence of autobiographical experience Shanahan et al. (2023) Park et al. (2024, preprint).

Goals, Planning, and Action Selection in Social Worlds

Generative agents often translate broad instructions into local plans: attend a meeting, seek a contact, share information, or complete a transaction. Planning makes the simulation more coherent by linking present action to imagined future states. In scientific terms, however, a natural-language plan is a representation in an engineered system, while a goal is inferred from patterns of selection under constraints. To establish goal-directed organization, experiments must change available options, deadlines, costs, and payoff structures, then observe whether choice adapts. A plan repeated in dialogue without corresponding action should be scored as verbal performance. Distinguishing utterance and consequential choice is essential when simulations are used to study incentives or collective behavior Park et al. (2023) Horton et al. (2023/2026).

Agent architectures frequently mix model-generated judgments with programmer-defined rules. An actor might say that friendship motivates a decision while the simulator's reward function actually determines the selection. Alternatively, the system may have no explicit utility function and instead rely on prompt-conditioned social scripts. These alternatives predict different responses to changes in incentives, competing responsibilities, and resource scarcity. A transparent experimental protocol specifies whether the decision function is sampled, optimized, rule-based, or selected by an external controller. Factorial ablations can then identify contributions from model prompts, stated preferences, actual reward signals, and social context Epstein (1999) Grimm et al. (2010).

Consequential social simulations require action validity. In a resource-allocation task, the agent should make an actual allocation recorded by the environment, not merely describe what it might give. In a negotiation, agreement should update commitments and constrain later choices. In a group meeting, information should reach only participants present or linked by a specified channel. These constraints make action traces suitable for comparison with human task data. They also reveal discrepancies between fluent social explanations and the behavior implemented by the agent. Valid planning claims depend on state transitions and independent outcomes, with text serving as supplementary evidence about the generated process Horton et al. (2023/2026) Dillion et al. (2023).

Environments, Social Affordances, and Interaction Protocols

An environment defines what actors can perceive, where they can move, whom they encounter, which resources they can use, and how actions change the world. These affordances are substantive psychological conditions. Scarcity, anonymity, visibility, reciprocity, and institutional enforcement can alter the incentives and meanings of social decisions. An artificial town with universal communication and no material cost differs systematically from a human neighborhood with unequal access and physical constraints. Descriptions of the environment must therefore include spatial structure, temporal scheduling, observation rules, resource accounting, and communication privileges. Otherwise patterns may reflect convenient software defaults instead of theoretically meaningful social mechanisms Schelling (1971) Park et al. (2023).

Protocols can impose simultaneous or sequential interaction. In a sequential system, early messages affect later actors; in a synchronized system, all actors act on the same previous state. This detail changes contagion, coordination, and bargaining. Researchers should explicitly identify the event ordering, the maximum context accessible to each agent, whether model calls share common knowledge, and the handling of conflicting actions. Repeating simulations under randomized update order is an informative robustness check. Network experiments likewise need a record of edge formation, message delivery, and missing information so that apparent influence can be assigned to actual contact rather than common input Axelrod (1997) Jeon et al. (2025).

Environment realism should be judged against its intended explanatory task. A minimal grid may be sufficient to study formal sorting under different tolerance thresholds. A conversation platform may require realistic attention limits, exposure ranking, and moderation procedures. A policy scenario may demand institutional eligibility rules, uneven resources, and calibrated population composition. Additional visual or linguistic detail is valuable only when it changes a construct or a testable prediction. The standard of adequacy is therefore structural fidelity to the research question, evaluated through known baseline patterns, sensitivity analysis, and independent observations. Rich fictional detail without matching mechanisms offers limited scientific value Grimm et al. (2006) Centola & Macy (2007).

Synthetic Surveys and the Representation of Public Opinion

Synthetic survey research prompts a model to answer questions as if it were a respondent or a demographically described sample. Its most accessible criterion is the resemblance between artificial and human response frequencies. However, frequency matching depends on sampling frames, question wording, response categories, item order, and the model's accessible information. Argyle and colleagues' early research illustrated patterned agreement between model-conditioned and human survey responses, which generated interest in scalable social inquiry. A valid evaluation extends beyond selected examples: the research must compare held-out survey questions, subgroup distributions, variance, correlations among measures, and prediction errors across models and dates. Each synthetic response is produced by a computation, not by contacting a sampled member of the target public Argyle et al. (2023) Bisbee et al. (2024).

Bisbee and colleagues compared persona-conditioned ChatGPT answers with American National Election Study feeling thermometers for 11 groups. The model could approximate some aggregate scores yet exhibited substantially less response variation, differences in regression relationships, wording sensitivity, and temporal instability. This illustrates a general measurement point: a favorable estimate of one population mean does not establish a trustworthy survey distribution. Compressing differences between respondents can distort confidence intervals and obscure the relationships most important for explanation. The proper analysis must account for nonindependent synthetic draws, since repeated samples from one shared model are not independent observations from a human population. Naively treating synthetic sample size as human sample size creates false precision Bisbee et al. (2024).

A September 2026 Pew Research Center methodological study compared synthetic versions of three American Trends Panel waves with responses from the actual panelists. Its reported average absolute question-level difference was 12.4 percentage points across the three waves, with consequential errors in many subgroups. The study concerned its particular prompts, model, respondent information, and survey context; it does not establish an invariant error rate for every architecture. It does establish the inadequacy of fluency and demographic matching as validation. A defensible synthetic polling study requires transparent measurement error, a contemporaneous human benchmark, appropriate weighting, item-level scrutiny, and explicit restrictions on use for public opinion inference Chapekis et al. (2026).

Forecasting Experimental Effects with Artificial Respondents

Experimental forecasting is a more demanding and more structured use of simulated agents than simply predicting an answer to a survey item. The analyst supplies treatment and control stimuli and asks simulated participants to respond under each condition; the difference is the predicted effect. In a large 2026 investigation, Ashokkumar and colleagues assembled 70 preregistered nationally representative U.S. survey experiments, including 469 effects involving 119,330 participants. Predictions from a language-model procedure were correlated with measured treatment effects and, for that task, performed comparably to pooled human forecasts. The benchmark linked predictions to independently observed experimental outcomes rather than to aesthetic assessments of plausible text Ashokkumar et al. (2026).

That result also exposes the difference between ranking effects and estimating their magnitude. The same study reported systematic overestimation of effect sizes, meaning that a simulation may indicate which intervention is more promising while misleading researchers about the expected size of real-world change. Its secondary archive of megastudies produced weaker correlations. A rigorous application must report signed effect errors, calibration slopes, uncertainty, heterogeneity across domains, and performance against trivial baselines such as predicting zero effect. Studies should also examine whether simulated predictions change after rephrasing stimuli or changing demographic conditioning, and whether forecast accuracy survives genuinely prospective tests Ashokkumar et al. (2026) Dillion et al. (2023).

An experimental agent does not become a randomized human participant merely because it is assigned to a simulated treatment arm. Randomization among model calls estimates a contrast within the artificial system. Its applicability to people depends on transportability assumptions that the system cannot establish internally. A credible pipeline therefore uses synthetic forecasts to screen hypotheses, select candidate treatments, and anticipate implementation issues, while human experiments estimate human causal effects. Where independent evidence is available, calibration adjustments can be estimated on a training archive and evaluated on a distinct test archive. This division preserves the scientific usefulness of computational forecasting without declaring model responses equivalent to observations from living subjects Ashokkumar et al. (2026) Bisbee et al. (2024).

Interactive Communities and the Generative-Agents Town

Multi-agent community simulations place artificial actors in an environment where conversations, observations, and actions can affect future decisions. Park and colleagues demonstrated an interactive town populated by 25 generative agents, with event memories, reflection, planning, and opportunities for emergent coordination. The work has become a foundational example because it linked the agent's accumulated record to plausible local behavior and provided architectural ablations. Its evidence supports an engineering possibility: language-driven components can sustain a coherent small-scale artificial community under defined conditions. It should be distinguished from a survey-validated simulation of an existing town or population Park et al. (2023).

Scientific interpretation of an artificial town requires specifying what counts as success. Possible outcomes include remembering social relations, completing invitations, spreading information over a network, honoring schedules, or producing interaction paths with known human analogues. Human evaluators can judge plausibility or consistency, but their ratings do not necessarily indicate statistical realism. An artificial community can generate many convincing anecdotes while producing implausible encounter frequencies or almost no disagreement. Measures should therefore combine qualitative trace audits with quantitative statistics on contact, influence, turnover, and rule compliance. A controlled environment provides opportunities for ablations that are harder to arrange in field research, but external validation remains a separate stage Park et al. (2023) Grimm et al. (2010).

Scaling from 25 to hundreds or thousands of agents introduces dependence, computation cost, and concentrated priors. An apparent population of many agents can contain limited independent information when all actors use similar prompts or a single foundation model. Network topology, shared system instructions, and scheduling decisions can produce collective consensus without interpersonal persuasion. Researchers should compare runs with heterogeneous and homogeneous model backends, distinct persona sampling procedures, varied temperatures, and independent initial conditions. Evaluation should investigate both repeated collective patterns and the range of variation across simulations. Large numbers of generated actors improve scenario coverage, but they do not by themselves improve population representativeness Xie et al. (2026) Wang et al. (2025).

Network Conversations, Information Diffusion, and Norm Formation

Digital social networks are convenient simulation environments because messages, connections, exposure, and responses can be represented as explicit events. A model can test how ranking algorithms, attention limits, group boundaries, and conversational tone influence cascades. Jeon and colleagues' 2025 study of generative agent-based models for social media conversations illustrates this approach, joining language-model text generation with a networked simulation. The evaluation question is whether the resulting conversations and trajectories match appropriate real-world evidence, not whether isolated posts resemble authentic social media prose. Relevant benchmarks include reply chains, participation concentration, topic persistence, interaction asymmetry, and the timing of diffusion Jeon et al. (2025).

Norm formation requires evidence that actors modify expectations or actions in response to others. An aggregate increase in similar statements is insufficient because a shared prompt could generate the similarity directly. A research design should introduce controlled exposure pathways, independent control communities, and event records identifying the timing and content of communication. It should vary whether adoption requires a single exposure or multiple reinforcing signals. The mathematical theory of complex contagion predicts that adoption thresholds and network bridges interact, providing a mechanism against which language-driven simulations can be tested. Human-network observations can then evaluate where the artificial threshold assumptions are realistic Centola & Macy (2007).

Online polarization creates a particularly demanding validation case. Political preferences are heterogeneous, shaped by institutions and history, and often react to perceived audiences. Human field experiments such as Bail and colleagues' study provide context-specific causal observations, but simulations must reproduce intervention definitions and recruitment conditions before comparisons are meaningful. Researchers should refrain from treating a stylized polarization trajectory as a validated population forecast. Sensitivity analyses should vary exposure, recommendation policies, moderation, actor turnover, and historical shocks. Models can help identify plausible pathways through which social feedback amplifies disagreement, provided the report labels these pathways as generated hypotheses requiring human validation Bail et al. (2018) Jeon et al. (2025).

Economic Games, Cooperation, and Institutional Scenarios

Economic games make aspects of social behavior operational through observable choices, budgets, information, and payoffs. Trust games, bargaining tasks, public goods decisions, and market interactions allow simulations to record consequential actions rather than conversational claims about preferences. Horton and colleagues' Homo Silicus research program investigates using language models as simulated economic actors to probe whether their choices can reproduce established experimental patterns. This provides an adaptable experimental interface for generating hypotheses about incentives. Its scientific value depends on comparison with independently measured human responses and on checking the model's sensitivity to payoff framing and selection rules Horton et al. (2023/2026).

The Global Preference Survey comparison published by Chen and colleagues in October 2026 supplies a stringent cross-national test. It matched artificial responses to 67,201 people across 74 countries for six experimentally validated preference domains. Although some demographic gradients were recovered, correspondence for cultural dimensions was much weaker, and the models were judged unreliable for representative macro-level inference or policy evaluation. The finding underscores that a model can exhibit learned knowledge of common economic concepts while failing to recover local distributions of risk-taking, reciprocity, patience, altruism, or trust. A synthetic economic participant should therefore be evaluated against the same behavioral or survey instruments the intended human population actually completed Chen et al. (2026).

In institutional scenarios, realistic action depends on constraints beyond individual preference. Legal eligibility, enforcement, group power, resource inequalities, and information asymmetry shape who can transact and under which conditions. Generative actors that freely negotiate without such constraints model a hypothetical institution, not an existing one. A transparent study should separate institutional rules from agent decision procedures and test whether group patterns persist when enforcement and resource conditions change. Useful results may identify fragile assumptions or unexpected coordination failures. Forecasts concerning policy impacts, public welfare, and marginalized groups require independent human or administrative evidence before being adopted as estimates of actual outcomes Epstein (1999) Chen et al. (2026).

Construct Validity and the Levels of Empirical Correspondence

Validation begins by specifying the correspondence a simulation is expected to achieve. Surface validity concerns whether outputs seem understandable or socially credible; task validity asks whether actions conform to experimental instructions; construct validity concerns whether a measured indicator reflects its intended psychological variable; distributional validity compares statistical patterns; and transport validity addresses performance in new situations or populations. These criteria should be tested separately. Agreement from human raters about the naturalness of dialogue is relevant to conversational interface research but cannot establish the accuracy of synthetic survey estimates. Similarly, an accurate mean under one prompt does not validate an agent's reasoning mechanism or its predictions for new institutions Dillion et al. (2023) Xie et al. (2026).

Construct validity is strengthened by convergent and discriminant tests. A model of trust should show expected associations across multiple trust measures while remaining distinguishable from generic agreeableness or a preference for positive wording. A model of social influence should respond to actual information received from peers and less to irrelevant shared labels. Researchers can manipulate information pathways, incentives, or role descriptions and test whether outputs track the construct's hypothesized causal antecedents. Failure to separate closely related constructs can produce seemingly persuasive simulations that are analytically empty. A useful validity argument therefore describes both what the simulation should reproduce and what it should resist reproducing Wang et al. (2025) Argyle et al. (2023).

Scientific comparison should report the unit of observation and its evidential status. One synthetic answer, one generated persona, one unique human, one interaction event, and one independently sampled community are different units. The effective information in repeated model calls depends on the shared foundation model, prompts, and environments. Apparent gains from multiplying calls may largely reflect replication of the same prior. Confidence intervals formed under ordinary independent-sampling assumptions can therefore be misleading. Clustered evaluations, repeated-seed designs, hierarchical error decomposition, and human benchmarks from independently recruited populations are needed to interpret where correspondence actually occurs Bisbee et al. (2024) Chen et al. (2026).

External Validity, Ecological Validity, and Population Transport

External validity concerns whether a finding persists beyond the circumstances in which it was produced. For synthetic agents, these circumstances include the model version, prompt language, sampling frame, population, setting, time, question wording, and available actions. A demonstration in an English-language fictional village is informative about that particular setting, but it offers no automatic prediction of social behavior in another language or institution. Ecological validity further requires examining the extent to which the simulation's opportunities, constraints, and incentives correspond to those in the real target activity. High conversational plausibility can coexist with unrealistic contact rates, economic stakes, or consequential choices Henrich et al. (2010) Grimm et al. (2006).

The human reference sample must also be examined. Psychological research has repeatedly faced difficulties when results from convenient, highly educated populations are generalized to all people. The 2010 analysis by Henrich and colleagues documents how the populations heavily represented in behavioral research can depart from broader human variation. The problem is intensified when synthetic actors are conditioned on text drawn disproportionately from digitally visible groups. Cross-national experiments in 2026 show that artificial economic preferences may capture some demographic trends while poorly recovering important cultural variation. A credible transport claim must specify its target, evaluate subgroup and country errors, and revisit assumptions under local language and institution changes Henrich et al. (2010) Chen et al. (2026).

Transport can be studied directly. Investigators can fit or select a model on one time period, region, or experiment archive and evaluate it on another, with predeclared metrics. They can compare original-language prompts with translated prompts, evaluate underrepresented groups separately, and measure changes after new model releases. Where a model fails, the report should record the direction and size of errors instead of averaging them away. Calibration against historical data may improve a forecast, but only prospective evaluation can establish whether a correction remains useful after distributional change. The appropriate scientific claim is conditional on a specified application and verified target distribution Bisbee et al. (2024) Ashokkumar et al. (2026).

Statistical Realism, Heterogeneity, and Calibration

Statistical realism concerns the relationships among generated observations, not simply the elegance of any individual answer. It includes marginal response frequencies, subgroup distributions, pairwise associations, multivariate prediction, temporal sequences, and the dependence between life events and background characteristics. Xie and colleagues' 2026 SSDataBench study compared 15 language models with patterns from four longitudinal and three cross-sectional social surveys, reporting compression of human heterogeneity into simplified typologies under sparse conditioning. This shows why many apparently varied persona descriptions can still produce a narrow set of answer patterns. Validation should compare an entire statistical signature with human data rather than selected headline indicators Xie et al. (2026).

Calibration asks whether predicted probabilities or effect magnitudes correspond to measured frequencies or effects. For categorical outcomes, suitable checks include class-wise error, probabilistic calibration, and frequency differences by subgroup. For continuous outcomes, distributional distance, rank correlations, and mean squared errors each address different aspects of fit. The benchmark's target must determine which measure matters most. A synthetic sample might reproduce an overall mean but understate uncertainty, yielding overconfident estimates of a subgroup difference. Studies should present raw values and confidence intervals from independent human benchmarks, disclose any tuning to the test data, and assess whether calibrations remain stable after prompt and model changes Bisbee et al. (2024) Ashokkumar et al. (2026).

Two major 2026 evaluations sharpen this distinction. Chen and colleagues found limited individual-level alignment and weaker cultural correspondence in a matched cross-national economic-preference survey. Pew Research Center found meaningful question-level discrepancies when its synthetic American Trends Panel respondents were compared with the humans represented by their profiles. These investigations use different constructs and sampling designs, so their numerical scores should not be pooled. Their shared methodological lesson is that population inference demands distribution-level validation and transparent uncertainty. Success in predicting which treatment works better does not establish accurate estimates of the public's current preferences or the full distribution of individual responses Chen et al. (2026) Chapekis et al. (2026).

Causal Identification, Interventions, and Counterfactual Logic

A simulator can produce a counterfactual narrative by changing a prompt and generating another outcome. This operation estimates the model's response to a hypothetical change; it is not automatically an estimate of the causal effect in people. Human causal inference requires an identification strategy, such as genuine random assignment, credible natural experiments, or justified structural assumptions. Simulations add value by encoding explicit assumptions and exploring their implications. Their outputs become empirically consequential when predictions are tested against interventions conducted in the intended social system. A research report should state exactly what has been randomized: human participants, simulated actors, prompts, event order, or algorithmic seeds Ashokkumar et al. (2026) Epstein (1999).

Within a multi-agent environment, feedback makes causal analysis harder because one actor's treatment can alter the exposure and decisions of others. Standard assumptions of noninterference may fail by construction. Researchers can nevertheless study intervention mechanisms through controlled network rewiring, partial saturation, staggered agent treatments, and logging of every observed contact. These tests identify causes within the artificial world. Transport to humans then requires evidence that the modeled interference pathways resemble observed ones. For social diffusion and polarization, the experimental design should distinguish exposure, message production, adoption, and later persistence, since each outcome can have different causal determinants Centola & Macy (2007) Bail et al. (2018).

Structural uncertainty creates another challenge: distinct agent rules can reproduce the same data. A successful simulation of segregation, consensus, or rapid diffusion does not uniquely validate the preferences or psychological processes assigned to actors. Competing mechanistic models should be fit to the same benchmarks and compared using out-of-sample predictions that distinguish them. Active experimentation can select interventions on which candidate mechanisms disagree most strongly. The strongest inference follows from a sequence of explicit models, discriminating predictions, human measurements, and model revision. This cycle turns a generative simulation into an instrument for theory testing rather than a source of self-confirming social explanations Schelling (1971) Grimm et al. (2010).

Replication, Ablation, Sensitivity, and Reproducible Reporting

Reproducibility requires enough information to recreate the agent configuration and its environment. For generative social simulation this includes model provider and identifier, parameter settings, seed controls where supported, all prompts, actor profiles, retrieval rules, external tools, event schedules, network topology, stopping criteria, and analysis scripts. The Overview, Design concepts, and Details protocol developed for agent-based models offers a foundation for organizing these materials. Contemporary generative systems add further dependencies, including changing hosted model versions and hidden provider-side modifications. Reports should preserve immutable run artifacts and state when exact computation is unavailable for third-party inspection Grimm et al. (2006) Grimm et al. (2010).

Ablation studies remove or alter one component at a time: persistent memory, planning, reflection, persona attributes, network communication, or access to external information. The investigator then measures changes in defined outcomes rather than relying on an impression of better or worse role-play. Park and colleagues employed component comparisons in their generative-agent work, illustrating how an architecture can be studied rather than merely exhibited. Independent replications should also vary the language model, prompts, environment, and calibration procedure. If a claimed social effect disappears when an innocuous wording choice changes, it requires reinterpretation as a fragile model-contingent output Park et al. (2023) Bisbee et al. (2024).

Replication standards from psychology underscore the importance of independently reproducing outcomes and measuring uncertainty. The Open Science Collaboration's 2015 multicenter replication project demonstrated how methodologically rigorous repetition can challenge established beliefs about effect robustness. Generative-agent research needs comparable habits adapted to computational studies: public benchmark specifications, multiple independent implementations, held-out tasks, clear denominators, and negative results. Statistical testing must account for the shared structure of runs, actors, and language-model outputs. A reproducible social simulation is one whose assumptions and results can be inspected; a robust social inference additionally requires successful tests against evidence from human communities Open Science Collaboration (2015) Xie et al. (2026).

Culture, Social Position, and Representational Inequality

Human societies contain variation shaped by language, institutions, economic circumstances, historical experience, and unequal access to opportunities. An artificial sample derived from a single foundation model may reproduce familiar descriptions of these differences without reproducing their actual distributions or the experiences of people represented. The relevant problem is epistemic as well as statistical: model-generated statements can acquire unwarranted authority when attributed to communities who did not provide them. Studies must identify which groups supplied the human benchmark, how profiles were constructed, and which voices are missing from the underlying datasets. Reporting an overall accuracy score while concealing subgroup errors fails the requirements of representative social inference Wang et al. (2025) Henrich et al. (2010).

The evidence has become especially concrete. Wang and colleagues found misportrayal and flattening of demographic identities in model-substitution experiments. The matched cross-national investigation by Chen and colleagues found substantial limitations in recovering variation associated with cultural dimensions. Pew Research Center's 2026 analysis likewise documented consequential subgroup discrepancies in synthetic public opinion. These are distinct datasets and research designs, yet they converge on a key warning: a demographically conditioned language model can produce confident statements about social categories while offering an impoverished representation of the heterogeneity within them. Improvements should be evaluated with the affected populations, independent benchmarks, and explicit analysis of the distribution tails Wang et al. (2025) Chen et al. (2026) Chapekis et al. (2026).

Cross-cultural research design must go beyond translating prompts. Survey concepts may lack direct functional equivalence across languages; institutions confer different meanings on cooperation, trust, authority, and personal choice. Good practice includes local construct adaptation, response-style analysis, multi-language testing, and population-specific validation before comparing countries. For disadvantaged or scarcely digitized communities, the absence of relevant training material may interact with stereotypes learned from dominant media. Models that serve as exploratory scenario engines can remain useful, but policy or scientific claims about a particular population should rest on observations of that population. The distinction preserves both the ethical legitimacy and empirical accuracy of social research Henrich et al. (2010) Chen et al. (2026).

Anthropomorphism, Role Play, and Human–Artificial Relations

Humans routinely interpret responsive systems through social categories. An agent that remembers a name, uses polite language, and appears to anticipate intentions can elicit judgments of personality, trustworthiness, competence, or care. Psychological theories of anthropomorphism help explain why such attributions arise under uncertainty, perceived social connection, and accessible human-centered knowledge. These responses are legitimate empirical phenomena in people interacting with artificial systems. They should, however, be measured separately from the internal computational processes of the system. A user's experience of reciprocity and the agent's architecture of turn-taking are complementary research objects, and their relationship can be tested experimentally Epley et al. (2007) Shanahan et al. (2023).

Role play is a useful formal lens for language-model behavior. Shanahan and colleagues described how dialogue models can sustain enacted roles through context-conditioned generation. This approach explains why an agent may present a coherent persona or discuss apparent motives without implying that every first-person assertion is a valid measurement of internal experience. Investigators should inspect whether an agent's reported intentions predict its future actions under controlled conditions, whether the apparent personality survives new constraints, and how user expectations shape judgments. The same simulation can support valuable research into human perception of artificial agents even when its validity as a stand-in for particular human respondents is limited Shanahan et al. (2023) Epley et al. (2007).

In the Artificial Era, sustained contact with artificial partners creates longitudinal social relations that deserve separate empirical study. A research program can measure perceived continuity, trust calibration, disclosure, reliance, emotional attachment, and changes in human social behavior without assuming a predetermined ontological answer about artificial subjective experience. Generative-agent simulations may help researchers construct controlled social stimuli and compare interaction structures, but their human consequences require studies with actual users. This scope connects the dossier to Psychology of Artificial and to investigations of human–artificial relations while maintaining clear evidential boundaries between engineering observations, psychological attribution, and philosophical interpretation Dillion et al. (2023) Wang et al. (2025).

Privacy, Bias, Safety, and Governance of Synthetic Social Worlds

An interview-conditioned agent or individualized digital twin may store sensitive recollections, preferences, relationships, and behavioral descriptions. The resulting data warrant governance comparable to other identifiable research records: consent about reuse, bounded access, retention rules, controlled publication of transcripts, and disclosure of third-party model processing. Removing an explicit name does not guarantee privacy when a distinctive history remains retrievable. Investigators should minimize stored personal material, document whether model providers retain prompts, and assess memorization or leakage risks. Human sources have interests in how their narratives are represented even when an artificial system generates the final output Park et al. (2024, preprint) Weidinger et al. (2022).

Social simulations can also cause harm through misleading impersonation, group stereotypes, manipulated public opinion narratives, or automated generation of persuasive interactions. The taxonomy of language-model risks developed by Weidinger and colleagues is relevant because generative agents turn linguistic outputs into persistent social actions and shared environments. Evaluation should include harmful response patterns, susceptibility to adversarial prompts, misleading claims of representativeness, and the consequences of exposing model-generated dialogue as genuine testimony. Safety testing should use defined outcomes and escalation criteria. A simulation intended for research should remain explicitly identifiable as artificial to participants and readers Weidinger et al. (2022) Bender et al. (2021).

Governance requires assigning responsibility for data collection, model selection, simulation design, interpretation, and publication. Human participation should be ethically reviewed under the applicable research rules, particularly where the system uses identifiable information or exposes participants to consequential manipulation. Research teams should record funding and conflicts, preserve versioned protocols, and provide an account of whether third parties can reproduce the reported outputs. Public use for market research, electoral communication, or policy decisions demands especially rigorous disclosure of model limitations. The defensible contribution of synthetic agents is to augment scientific inquiry with inspectable scenarios and hypotheses while maintaining direct human evidence for assertions about human lives Wang et al. (2025) Chapekis et al. (2026).

An Integrated Validation Protocol and Research Agenda

A scientific protocol for generative social simulation begins with a sharply defined target claim: a pattern of conversations, a statistical distribution of survey answers, an intervention effect, or a collective trajectory. Investigators then register the target population, eligible contexts, expected mechanism, and primary outcomes. They document the agent architecture, prompts, model versions, retrieval sources, environment, network topology, action permissions, stopping rules, and sampling design. The corresponding human reference must be specified before the synthetic results are interpreted. For a conversational simulation this may be a corpus with independently coded reply structure; for synthetic surveys it is a representative human dataset; for intervention forecasts it is a set of experiments with observed treatment effects Grimm et al. (2010) Xie et al. (2026).

The next stage combines internal and external tests. Internal tests verify state consistency, memory provenance, action execution, prompt stability, and controlled responses to manipulated information. External tests compare model outcomes with held-out human measures, using suitable distributional metrics, subgroup diagnostics, effect calibration, and uncertainty estimates. Alternative mechanisms and simpler baselines must be evaluated against the same evidence. Replications vary seeds, backend models, languages, time points, and network assumptions. Data contamination is investigated through prepublication or prospective benchmarks where possible. Reporting should include negative findings and the conditions in which a simulated pattern disappears Ashokkumar et al. (2026) Bisbee et al. (2024) Xie et al. (2026).

The research agenda has several connected priorities. It needs stronger measurements of genuine human-agent behavioral correspondence, more culturally diverse and privacy-respecting benchmarks, transparent model ensembles that do not masquerade as independent human samples, and prospective tests of collective processes under interventions. It also needs instruments to separate generated discourse from consequential action and to identify which network conditions transform local changes into social outcomes. Progress can be judged by transparent improvements in prediction calibration and explanation, with explicit attention to failures in rare groups and unfamiliar institutions. These measures support a cumulative science of artificial social modeling within the larger study of psychologically consequential Artificial systems Chen et al. (2026) Jeon et al. (2025).

Conclusion

Social simulation and generative agents form a distinct research object at the intersection of psychology, computational social science, and artificial systems research. Their central achievement is the capacity to construct interactive environments in which agents produce contingent language, retain selected histories, pursue plans, and participate in group processes under explicit constraints. Earlier agent-based science established the logic of generative explanation and the importance of mechanistic transparency. Language-model systems broaden the expressiveness of artificial actors and make new experimental scenarios feasible. These developments justify systematic investigation of synthetic social behavior as an engineered phenomenon whose properties can be observed, perturbed, and reproduced Epstein (1999) Park et al. (2023).

The empirical record supports differentiated conclusions. Controlled artificial communities demonstrate architecture-level possibilities; studies of experimental forecasting show useful correlations with some human treatment effects alongside systematic calibration problems; survey comparisons reveal serious limitations in distributional realism, subgroup representation, and cross-national transport. The 2026 research literature adds particularly important evidence on heterogeneity compression and culturally weak preference matching. Accordingly, no single claim that artificial agents are realistic or unrealistic captures the scientific findings. An adequate conclusion names the target outcome, the independent human comparator, the tested population, and the type of correspondence actually demonstrated Ashokkumar et al. (2026) Xie et al. (2026) Chen et al. (2026).

The durable methodological principle is that generated social behavior is an observation of an artificial system, while claims about human populations require appropriately designed human evidence. Generative agents can formulate hypotheses, expose assumptions, simulate mechanisms, help screen candidate interventions, and support experiments on human responses to artificial partners. They earn stronger explanatory or predictive standing through preregistered comparisons, causal interventions, diverse external benchmarks, and independent replication. In the historical horizon of the Artificial Era, this framework establishes a disciplined route for studying artificial social organization and its relationship with Homo: describe mechanisms precisely, test correspondence directly, preserve the provenance of synthetic observations, and publish conclusions at the level warranted by the evidence Grimm et al. (2006) Wang et al. (2025) Chapekis et al. (2026).

References

  • Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337–351. https://doi.org/10.1017/pan.2023.2
  • Ashokkumar, A., Hewitt, L., Ghezae, I., & Willer, R. (2026). Large language models can predict the results of social science experiments. Nature, 656, 115–122. https://doi.org/10.1038/s41586-026-10742-x
  • Axelrod, R. (1997). The dissemination of culture: A model with local convergence and global polarization. Journal of Conflict Resolution, 41(2), 203–226. https://doi.org/10.1177/0022002797041002001
  • Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. B. F., Lee, J., Mann, M., Merhout, F., & Volfovsky, A. (2018). Exposure to opposing views on social media can increase political polarization. Proceedings of the National Academy of Sciences, 115(37), 9216–9221. https://doi.org/10.1073/pnas.1804840115
  • Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). ACM. https://doi.org/10.1145/3442188.3445922
  • Bisbee, J., Clinton, J. D., Dorff, C., Kenkel, B., & Larson, J. M. (2024). Synthetic replacements for human survey data? The perils of large language models. Political Analysis, 32(4), 401–416. https://doi.org/10.1017/pan.2024.5
  • Centola, D., & Macy, M. (2007). Complex contagions and the weakness of long ties. American Journal of Sociology, 113(3), 702–734. https://doi.org/10.1086/521848
  • Chapekis, A., Lau, A., Bestvater, S., Shah, S., Mercer, A., & Smith, A. (2026, September 30). How well synthetic samples replicate public opinion. Pew Research Center. https://www.pewresearch.org/data-labs/2026/09/30/how-well-synthetic-samples-replicate-public-opinion/
  • Chen, F., Wang, Q., Wang, L., & Bhatia, S. (2026). Simulated responses from large language models are weak proxies for human economic preferences across 74 countries. Nature Human Behaviour. Advance online publication. https://doi.org/10.1038/s41562-026-02611-4
  • Dillion, D., Tandon, N., Gu, Y., & Gray, K. (2023). Can AI language models replace human participants? Trends in Cognitive Sciences, 27(7), 597–600. https://doi.org/10.1016/j.tics.2023.04.008
  • Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
  • Epstein, J. M. (1999). Agent-based computational models and generative social science. Complexity, 4(5), 41–60. https://pure.johnshopkins.edu/en/publications/agent-based-computational-models-and-generative-social-science-7/
  • Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., Goss-Custard, J., Grand, T., Heinz, S. K., Huse, G., Huth, A., Jepsen, J. U., Jorgensen, C., Mooij, W. M., Muller, B., Pe’er, G., Piou, C., Railsback, S. F., Robbins, A. M., ... DeAngelis, D. L. (2006). A standard protocol for describing individual-based and agent-based models. Ecological Modelling, 198(1–2), 115–126. https://doi.org/10.1016/j.ecolmodel.2006.04.023
  • Grimm, V., Berger, U., DeAngelis, D. L., Polhill, J. G., Giske, J., & Railsback, S. F. (2010). The ODD protocol: A review and first update. Ecological Modelling, 221(23), 2760–2768. https://doi.org/10.1016/j.ecolmodel.2010.08.019
  • Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and Brain Sciences, 33(2–3), 61–83. https://doi.org/10.1017/S0140525X0999152X
  • Horton, J. J., Filippas, A., & Manning, B. S. (2023; revised 2026). Large language models as simulated economic agents: What can we learn from Homo Silicus? National Bureau of Economic Research Working Paper No. 31122. https://doi.org/10.3386/w31122
  • Jeon, M. S., Mendoza, M., Fernández, M., Providel, E., Rodríguez, F., Espina, N., Carvallo, A., & Abeliuk, A. (2025). Simulating conversations on social media with generative agent-based models. EPJ Data Science, 14, Article 79. https://doi.org/10.1140/epjds/s13688-025-00593-3
  • Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251), Article aac4716. https://doi.org/10.1126/science.aac4716
  • Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, Article 2, 1–22. https://doi.org/10.1145/3586183.3606763
  • Park, J. S., Zou, C. Q., Shaw, A., Hill, B. M., Cai, C., Morris, M. R., Willer, R., Liang, P., & Bernstein, M. S. (2024). Generative agent simulations of 1,000 people [Version 1 preprint]. arXiv. https://arxiv.org/abs/2411.10109v1
  • Schelling, T. C. (1971). Dynamic models of segregation. Journal of Mathematical Sociology, 1(2), 143–186. https://doi.org/10.1080/0022250X.1971.9989794
  • Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498. https://doi.org/10.1038/s41586-023-06647-8
  • Wang, A., Morgenstern, J., & Dickerson, J. P. (2025). Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intelligence, 7, 400–411. https://doi.org/10.1038/s42256-025-00986-z
  • Weidinger, L., Uesato, J., Rauh, M., Griffin, C., Huang, P.-S., Mellor, J., Glaese, A., Cheng, M., Balle, B., Kasirzadeh, A., Biles, C., Brown, S., Kenton, Z., Hawkins, W., Stepleton, T., Birhane, A., Hendricks, L. A., Rimell, L., Isaac, W., ... Gabriel, J. (2022). Taxonomy of risks posed by language models. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 214–229). ACM. https://doi.org/10.1145/3531146.3533088
  • Xie, Y., Liang, L., Li, S., Lu, Y., Xiao, Z., Shi, M., Huang, J., Wang, M., & Xie, Y. (2026). Evaluating the statistical realism of LLM-generated social science data. Proceedings of the National Academy of Sciences, 123(19), e2538145123. https://doi.org/10.1073/pnas.2538145123