This human study did not involve human subjects: Validating LLM simulations as behavioral evidence

Jessica Hullman; David Broska; Huaman Sun; Aaron Shaw

cs.AIFeb 17, 2026

This human study did not involve human subjects: Validating LLM simulations as behavioral evidence

Authors: Jessica Hullman, David Broska, Huaman Sun, Aaron Shaw

Organizations: Northwestern University · Stanford University

Abstract

A growing literature uses large language models (LLMs) as synthetic participants to generate cost-effective and nearly instantaneous responses in social science experiments. However, there is limited guidance on when such simulations support valid inference about human behavior. We contrast two strategies for obtaining valid estimates of causal effects and clarify the assumptions under which each is suitable for exploratory versus confirmatory research. Heuristic approaches seek to establish that simulated and observed human behavior are interchangeable through prompt engineering, model fine-tuning, and other repair strategies designed to reduce LLM-induced inaccuracies. While useful for many exploratory tasks, heuristic approaches lack the formal statistical guarantees typically required for confirmatory research. In contrast, statistical calibration combines auxiliary human data with statistical adjustments to account for discrepancies between observed and simulated responses. Under explicit assumptions, statistical calibration preserves validity and provides more precise estimates of causal effects at lower cost than experiments that rely solely on human participants. Yet the potential of both approaches depends on how well LLMs approximate the relevant populations. We consider what opportunities are overlooked when researchers focus myopically on substituting LLMs for human participants in a study.

Figures & tables

Figure 1: A validate-then-simulate approach involves collecting some human observations, which are jointly labeled by the LLM ( $D_{\text{shared}}$ ), and using this dataset to demonstrate that the model achieves sufficient fidelity in approximating important aspects of the human results. Heuristic approaches then imply or assert that estimates ( $\hat{\theta}_{LLM}$ ) derived from a larger dataset where only LLM labels are available ( $D_{\text{LLM}}$ ) may serve as proxies for human-derived estimates ( $\hat{\theta}_{H}$ ) without statistical adjustment.

Figure 2: Statistical calibration approaches derive estimators that explicitly account for bias contributed by LLM approximations to human responses. PPI Angelopoulos et al. (2023) , DSL Egami et al. (2024) , and related approaches learn a rectifier and additively combine it with a base estimator. Approaches to plug-in correction learn a model of the relationship between human ground truth and LLM predictions using jointly labeled data ( $D_{\text{shared}}$ ), then correct either the LLM predictions prior to estimation ( Ludwig et al., 2025 ) or directly adjust the target inference ( Wang et al., 2020 ) .

Figure 3: In a simulate-then-validate approach, the researcher uses LLMs exclusively for exploratory piloting to find hypotheses with support, then conducts a human study on those that appear most promising.

Paper	Validation Description
Abdurahman, S., Atari, M., Karimi-Malekabadi, F., Xue, M., Trager, J., Park, P., Golazizian, P., Omrani, A., and Dehghani, M. (2024) Perils and Opportunities in Using Large Language Models in Psychological Research. PNAS Nexus .	GPT-3.5 administered MFQ-2 1,000 times; compared mean response magnitude, variance, and item-level correlations to human survey data; noted exaggerated cross-domain correlations and lower within-domain variance. Emphasis on replicating qualitative and quantitative effect patterns rather than formal hypothesis testing.
Abeliuk, A., Gaete, V., and Bro, N. (2025). “Fairness in LLM-Generated Surveys.” arXiv Preprint.	Benchmarked LLM-simulated survey responses against ground-truth demographic labels and annotated data. (i) individual level accuracy of predictions; (ii) distributional similarity between LLM- and human-derived responses using Jensen-Shannon similarity; (iii) subgroup parity tests (differences in mean response probability by demographic group); (v) comparison to a Random Forest trained on the same human data to contextualize predictive accuracy.
Aher, G., Arriaga, R. I., and Kalai, A. T. (2023). “Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies.” ICML.	Constructed parallel LLM and human experiments and compared summary results. (i) sign and direction of effects matched to original human studies; (ii) magnitude comparison via mean differences and standard errors on the same tasks; (iii) distributional comparisons to human results (graphical); (iv) evaluation on both replicated and unseen stimuli to test generalization beyond training exposure. Emphasis on replicating qualitative and quantitative effect patterns rather than formal hypothesis testing. Also checked that validity rate of LLM responses (proportion valid generations) is high and for consistency of responses for name pairs.
Ahnert, G., Pellert, M., Garcia, D., and Strohmaier, M. (2025). “Extracting Affect Aggregates from Longitudinal Social Media Data with Temporal Adapters for Large Language Models.” Proceedings of the International AAAI Conference on Web and Social Media , 19(1), 15-36. https://doi.org/10.1609/icwsm.v19i1.35801	Evaluated whether temporally fine-tuned LLMs reproduce population-level affect signals. (i) time-series correlation of LLM-derived affect aggregates (weekly) with YouGov survey affect indices, using Pearson’s $r$ and permutation-based p-values; (ii) examination of mean and median squared errors and standard deviation of squared error computed against human data; (iii) trend alignment across major events (graphical qualitative analysis); (iv) comparison against BERT baseline models and statistical nulls.
Argyle, L., Busby, E., Fulda, N., Gubler, J., Rytting, C. and Wingate. D. (2023). “Out of One, Many: Using Language Models to Simulate Human Samples.” Political Analysis.	Proposed “algorithmic fidelity” as a construct for LLM-human comparison. (i) human-judge evaluations of free-text responses for human-likeness, including statistical test of difference in proportion of human lists that people judged to be human versus proportion of GPT generated lists they judged to be human and comparisons of prevelance of traits humans rated in human versus GPT lists (fit multiariate regression to traits and whether it was a human or GPT list, compare coefficients) ; (ii) correlation between probability of voting for human and GPT responses on ANES data, (iii) Cramer’s $V$ and correlation matrices across categorical and ordinal variables capturing voting and demographic and political preference variables between LLM-generated and ANES survey samples. Integrated both quantitative similarity metrics and qualitative human judgments of realism.
Binz, M., and Schulz, E. (2023). “Turning Large Language Models into Cognitive Models.” ICLR.	Fine-tuned LLaMA embeddings with a logistic layer on human trial-level data and validated by: (i) predictive accuracy (log-likelihood) on held-out human trials from Choices13k and horizon tasks; (ii) compare goodness of fit for random model, LLM without finetuning, domain specific model (iii) graphical comparison of patterns in model predictions relative to human data (iv) proportion of participants whose individual level data was best explained by Centaur model (v) cross-task generalization to an unseen task. Emphasis on replicating quantitative fit patterns and out-of-sample predictive accuracy rather than hypothesis testing.

Table 4

Explore similar work

Mar 8, 2026

Evaluating LLMs as Human Surrogates in Controlled Experiments

Large language models (LLMs) are increasingly used to simulate human responses in behavioral research, yet it remains unclear when LLM-generated data support the same experimental inferences as human data. We evaluate this by directly comparing off-the-shelf LLM-generated responses with human responses from a canonical survey experiment on accuracy perception. Each human observation is converted into a structured prompt, and models generate a single 0--10 outcome variable without task-specific training; identical statistical analyses are applied to human and synthetic responses. We find that LLMs reproduce several directional effects observed in humans, but effect magnitudes and moderation patterns vary across models. Off-the-shelf LLMs therefore capture aggregate belief-updating patterns under controlled conditions but do not consistently match human-scale effects, clarifying when LLM-generated data can function as behavioral surrogates.

Adnan Hoq, Tim Weninger

University of Notre Dame Notre Dame, Indiana, USA

Human Behavior Simulation Large Language Model-Based Social Simulation LLM Evaluation Belief Updating in LLMs

Ask PDF

Apr 2, 2026

Statistical realism is not evidence that LLMs can estimate treatment effects in social science experiments

Large language models (LLMs) are increasingly used to simulate human responses and estimate treatment effect of interventions when real-world experiments are costly or infeasible. The treatment-effect estimates are often evaluated using statistical realism, the degree to which simulated responses reproduce properties of observed human responses, although whether realism predicts treatment-effect accuracy remains unknown. Here we test this proxy relationship by jointly measuring statistical realism and treatment-effect accuracy on the same simulated responses in a cross-national experiment with 59,508 participants from 62 countries using three LLMs. The correlation between statistical realism and treatment-effect accuracy is weak, and optimizing for statistical realism can even worsen treatment-effect accuracy when selecting models, prompts, and target populations. The pattern replicates in two additional cross-national experiments spanning 12 and 27 countries with 20,785 participants. The divergence between the two reflects distinct error structures and is larger for behavioral outcomes, where models appear to extrapolate behavioral effects from attitudinal patterns. Because this divergence may remain hidden in deployment, errors can propagate into simulation-informed decisions. We introduce a diagnostic framework for LLM-generated synthetic data and discuss how treatment-effect validation should proceed under varying availability of experimental benchmarks. Simulated responses and simulated treatment effects are distinct estimation targets, and evidence for one does not certify the other.

Zonghan Li, Feng Ji

Department of Applied Psychology and Human Development, University of Toronto, ON, Canada

Large Language Model-Based Social Simulation Synthetic Data Evaluation Causal Effect Estimation

Ask PDF

May 20, 2026

The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational Study

Large language models (LLMs) show potential as simulators of human behavior, offering a scalable way to study responses to interventions. However, because LLMs are trained largely on observational data, interventions in experiments with LLM-simulated synthetic users can induce unintended shifts in latent user attributes, causing user drift where the implicit simulated population differs across treatment conditions, potentially distorting effect estimates. We formalize the confounding or selection bias that can arise due to user drift and show how intervention-dependent shifts can inflate or attenuate observed differences in user responses under intervention. To diagnose confounding, we propose using negative control outcomes--attributes that should remain invariant under intervention--to identify distribution shifts across intervention conditions, providing evidence of user drift. To mitigate drift, we study adjusting the persona specification by eliciting additional confounders, finding that targeted, setting-relevant confounders can substantially reduce bias across survey-style and multi-turn agent evaluations.

Victoria Lin, Taedong Yun, Maja Matarić +3

Google DeepMind · Carnegie Mellon University

Human Behavior Simulation Large Language Model-Based Social Simulation Causal Inference Selection Bias User Simulation

Ask PDF