Do Synthetic Citizens Have the Right Reasons? Testing Policy-Simulation LLMs for Structural Hallucination

Researchers now use large language models (LLMs) as stand-in survey respondents and as "generative agents," meaning simulated people who read news, talk to each other, and react to policies. These simulations are judged mainly on whether their headline numbers match reality, for example the share of people who accept a vaccine mandate. That check can pass while the simulation is wrong about why people accept or reject a policy. The psychology of policy acceptance has a well-documented internal structure. In human surveys, a person's sense of moral duty strongly predicts support for voluntary, incentive-based "pull" policies such as transit subsidies, and only weakly predicts support for coercive "push" policies such as congestion fees. Across 68 countries, believing that a disaster was caused by climate change predicts climate-policy support, while actually being exposed to the disaster mostly does not. If an LLM population matches the headline numbers but builds them from a different internal structure, then every "what if we tried policy X" answer it produces is a fluent fabrication about how a population reasons. We call this a structural hallucination.

This note proposes measuring structural hallucination with structural equation modeling (SEM). SEM is a standard statistical method from survey science. It estimates unobservable traits such as "personal norm" from several noisy questionnaire items, and then estimates the directed paths among those traits. We give LLM populations the same validated questionnaires that human respondents answered. We fit the same theory-given path model to both. We then test, with standard group-comparison statistics, whether the synthetic population's paths, explained variances, and construct boundaries match the human ones. The central prediction concerns coercive policies. We expect LLM agents to reproduce the pull-policy structure but to erase the push-policy drop-off, projecting moral compliance onto mandates where real people switch to self-protection.

Core Claim

The claim is that structural fidelity is a separate axis of simulation accuracy from marginal fidelity, and that it can be measured. An LLM population can be close to humans on average acceptance rates and still be structurally hallucinated in three detectable ways. First, it can show the wrong path strengths, such as personal norms predicting push acceptance as strongly as pull. Second, it can fabricate causal routes, such as objective disaster exposure driving support directly. Third, it can collapse constructs that humans keep distinct, such as personal norm and policy acceptance becoming one undifferentiated "pro-environment" factor.

The specific, falsifiable target comes from Do Personal Norms Predict Citizens' Acceptance of Green Transport Policies in China. That paper ran a partial-least-squares SEM on 315 Hangzhou residents along the Value-Belief-Norm chain (biospheric values, then awareness of consequences, then ascription of responsibility, then personal norm, then acceptance). It reported $R^2 = 0.437$ for pull-policy acceptance and $R^2 = 0.183$ for push-policy acceptance. We define the asymmetry gap as $\Delta = R^2_{\text{pull}} - R^2_{\text{push}}$, which is $0.254$ in the human data. Our hypothesis is that instruction-tuned models will shrink this gap toward zero.

The proposal also asks whether theory-scaffolded prompting fixes the problem or makes it worse. We suspect it can make it worse, because telling a model to reason through norms may impose the norm path everywhere.

Relation to Prior Work

Three lines of work motivate the gap. Can A Society of Generative Agents Simulate Human Behavior and Inform Public Health Policy? (VACSIM) simulated 100 census-parameterized agents under vaccine incentives and mandates. It reported a mean absolute error of 0.6% against Delphi survey hesitancy and Kendall's $\tau = 0.733$ against expert policy rankings. Its authors also note that agents oversimplify cognitive resistance and respond inconsistently across demographics. VACSIM validates outputs only. We validate the latent structure that produces them.

Chain-of-Planned-Behaviour Workflow Elicits Few-Shot Mobility Generation in LLMs (CoPB) structured LLM reasoning through the Theory of Planned Behavior and cut intention error from 0.578 to 0.194 on Tencent mobility data. It uses psychological theory as a prompt, but it never checks whether the generated behavior actually obeys the theory's estimated paths. We treat CoPB-style scaffolding as a competitor to be audited, not as a solution.

Aligning MLLM Benchmark With Human Preferences via Structural Equation Modeling is the closest methodological precedent. It used SEM diagnostics, including variance inflation factors, the heterotrait-monotrait (HTMT) ratio, and loadings, to prune multimodal benchmark tasks. That pruning produced the GOLD benchmark, whose correlation with Chatbot Arena preferences was 0.7359, versus 0.519 for MMBench. That work applies SEM to model abilities across benchmarks. We apply SEM to simulated human minds and compare their structure against a human reference population. The question, the unit of analysis, and the definition of error all differ.

Two structural hallucination methods supply ideas but not methods we can reuse directly. Hallucination Detection from Structural Reasoning Model turns reasoning into graphs and propagates step-level verification, reaching 76.59% average precision on MATH. Mitigating Hallucinations in Large Language Models via Causal Reasoning fine-tunes on explicit causal graphs and reports 95.33% on CLADDER and a 10% absolute gain on HaluEval. Both assume a correct answer exists for each instance, as in math, code, or formal causal queries. In policy simulation no single response is "wrong." Only the population-level structure can be wrong. That is why a population-level latent-variable test is needed instead of per-response verification.

The SEM theory literature constrains the design. Algebraic Equivalence of Linear Structural Equation Models and Testable Implications of Linear Structural Equation Models show that different directed structures can imply identical covariances. We therefore do not attempt structure discovery from synthetic data. We fix the theory-given model and compare parameters and fit across groups, which these equivalence results do not undermine.

Research Program

The work has four parts.

First, we build a questionnaire harness. We port the Hangzhou Value-Belief-Norm instrument item by item to LLM respondents, with randomized item order and Likert response formats. We do the same for the policy-support and attribution items from the 68-country study, Extreme weather event attribution predicts climate policy support across the world. That study joins 71,922 survey responses with objective hazard exposure computed via the CLIMADA platform. Personas are drawn to match each human sample's demographic marginals. For the global study, personas also receive their location and an objective exposure description, so the model has a chance to fabricate an exposure-to-support route.

Second, we run the analysis. We fit identical measurement and structural models to the human and synthetic samples. We use both partial-least-squares SEM, for comparability with the Hangzhou paper, and covariance-based SEM. The covariance-based version supports multigroup invariance testing, a sequence of nested tests asking whether loadings, then paths, can be constrained equal across human and synthetic groups without loss of fit. Structural hallucination is scored on three things: the asymmetry gap $\Delta$ with bootstrap intervals, the largest standardized path discrepancy, and HTMT discriminant validity in the synthetic data as a measure of construct collapse.

Third, we test whether structural fidelity can be improved without circularity. We calibrate persona construction and prompting so that the synthetic population matches the human pull-policy model only. We then check whether the push-policy drop-off emerges on held-out push items. If matching the pull structure restores the push asymmetry, models carry a latent reactance mechanism that prompting can activate. If calibration on pull leaves $\Delta$ near zero, the asymmetry is missing from the model and simulations of mandates should not be trusted.

Fourth, we check external validity. VACSIM reports that its agents distinguished strong from weak interventions on hesitancy. We will test whether the agents' attitude responses show the push/pull split for mandates versus incentives when given a short norm-and-acceptance instrument.

Baselines

The naive reference point is a demographically matched LLM population answering the questionnaire with plain zero-shot persona prompts. This is roughly what "silicon sample" studies do. It is run on the same instruments, the same persona draws, and the same number of respondents as every other condition. It defines how much structural hallucination exists by default. Every other condition is scored by its reduction in path discrepancy and in $|\Delta - 0.254|$ relative to this baseline.

The strongest existing simulation method is a VACSIM-style agent society, implemented as described in Can A Society of Generative Agents Simulate Human Behavior and Inform Public Health Policy?. It uses census-sampled personas, a warmup period, and temperature-based attitude modulation, with temperature tuned under the same budget used to tune our calibration. Its reported numbers are an MAE of 0.6% against real hesitancy and Kendall's $\tau = 0.733$ for Llama-3.1-8B-Instruct. Those numbers are on vaccine marginals, not on latent structure. We found no reported result on Lacuna for the structural fidelity of any LLM population on Value-Belief-Norm or push/pull acceptance models. We will match VACSIM on marginal acceptance error. A method beats it if, at equal or better marginal error, it shows significantly lower path discrepancy and does not fail invariance at the structural-path level where VACSIM fails.

The theory-scaffolding baseline is a CoPB-style prompt from Chain-of-Planned-Behaviour Workflow Elicits Few-Shot Mobility Generation in LLMs, adapted to walk through values, consequences, responsibility, and norms before answering. It is the most plausible "obvious fix," so it is the most important comparison.

Two ablations isolate our key ingredient. The first replaces latent SEM scoring with ordinary regression on sum scores. It tests whether the measurement model changes the conclusions. The second calibrates on pull and push jointly rather than on pull only. It shows how much of any apparent recovery is just fitting the target.

Why not just…?

Why not just compare acceptance rates, as VACSIM does? Acceptance rates are what structural hallucination hides behind. Our first experiment is designed to show populations whose marginal acceptance error is small while their push-path coefficients differ from humans' by more than their bootstrap intervals. If no such dissociation appears across models and prompts, the structural axis is redundant, and we will report that as a negative result.

Why not just correlate sum scores with ordinary regression? This is a real threat, and the second ablation settles it. The case for SEM is specific. LLM respondents tend to answer multi-item scales with unnaturally high internal consistency, so their measurement error differs from humans'. Ordinary regression attenuates coefficients by each group's own reliability. That makes human and synthetic coefficients non-comparable even when the underlying structure is identical. If the ablation yields the same conclusions as latent modeling, we will say so and recommend the cheaper analysis.

Why not just ask a strong LLM to judge whether the simulated reasoning is realistic? The HaluEval study, HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models, found ChatGPT detecting hallucinations at 58.53% accuracy on summarization. A judge sharing the simulator's priors about moral compliance has no population reference to check against. We will include a judge condition, asking it to rate each synthetic population's plausibility, and test whether its ratings track measured path discrepancy.

Why not just scaffold the prompt with the theory? Scaffolding may enforce the theory's paths everywhere, including where the human data show they weaken. We predict that CoPB-style prompting reduces marginal error but drives $\Delta$ toward zero or makes it negative by inflating the push-policy norm path. If instead it reproduces the human gap, scaffolding is the practical fix and our contribution narrows to the audit method.

Proposed Experiments

The decisive check uses the Hangzhou instrument. We run several open and closed models under the baseline, VACSIM-style, scaffolded, and pull-calibrated conditions, with 315 respondents each to match the human sample size and repeated draws to obtain intervals. The proposal succeeds as a diagnostic if at least one widely used condition is near the human acceptance marginals yet has a bootstrap interval for $\Delta$ that excludes $0.254$. It succeeds as a mitigation if pull-only calibration yields an interval for $\Delta$ that includes $0.254$ on held-out push items.

A data risk applies here. The item-level Hangzhou responses are not confirmed to be public. If they are unavailable, we will re-field the instrument on a comparable human sample and treat the published coefficients as a secondary target.

The second check uses the 68-country study. Synthetic respondents given objective exposure descriptions should show the human pattern: attribution positively associated with support for all five policies, and exposure mostly unrelated. Wildfire exposure should be positive and heavy-precipitation exposure negative. A synthetic population in which described exposure predicts support across hazards is fabricating a rationalist route, and that would be the clearest instance of the failure this note targets.

The construct-collapse check computes HTMT for the norm, responsibility, and acceptance factors in every synthetic sample. Values above conventional thresholds, where humans show discriminant validity, mean the model is not representing distinct mental states at all.

All human reference numbers in these checks come from the cited papers. No synthetic-population figures exist yet. Producing them is the work.