{"id":"5748c81c-0dbb-435c-9735-37c16a16c1c7","arxiv_id":"2507.02919","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ChatGPT and Llama personas show severe opinion homogenization and inconsistent distributions across demographic levels compared to real ANES 2020 survey data.","lead":"Large language models like ChatGPT and Llama give much less varied opinions than real people, and their answers change depending on how detailed a persona they are given. This study argues that treating chatbots as stand-ins for human survey respondents can misrepresent minorities and mislead policy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The probability-scoring method may not recover the models' true answer distributions; GPT-4 token logprobs and Llama constrained-completion scores are normalized as full answer probabilities, so every accuracy, consistency, and homogenization statistic inherits the distortion.","rationale":"The reader's weakest assumption is the right one. The paper's central empirical claims—structural inconsistency and homogenization—are computed entirely from purported probability distributions over the five answer options. The GPT-4 method is described only as 'logprobs ... transformed into probabilities of answers'; the Llama method scores a single fixed continuation and normalizes the five scores to sum to 1. Neither procedure is guaranteed to recover P(answer option), because the model's response space is much larger than the five tokens or strings considered. This is not a small implementation detail: all accuracy, structural-consistency, and variation-ratio statistics are functions of those distributions. A biased scoring method could manufacture both structural inconsistency and homogenization even from a model whose true conditional distributions were perfectly representative.\n\nI considered whether any other concern is more load-bearing. The accuracy-optimization proof is mathematically correct as a statement about expected 0/1 accuracy, though it is not a direct derivation from next-token training; however, that hypothesis is explanatory, not constitutive of the empirical claim. The structural-consistency criterion could be criticized as specific to the authors' query protocol rather than to individual-level silicon sampling, but that critique is secondary: the measurement problem would invalidate even that protocol's output. The paper's limitation section acknowledges comparability assumptions but does not address scoring validity, so the concern is not preempted by the authors' own caveats.\n\nThe proposed test—comparing logprob-derived distributions to empirically sampled free-form answer distributions for a subset of personas—would directly settle whether the measured distributions correspond to the models' actual behavior. Conditional on that test passing, the empirical findings stand; conditional on it failing, the central claims collapse. Thus the reader's CONDITIONAL verdict is appropriate and unchanged.","tokens_in":16239,"tokens_out":7646,"duration_ms":84994,"concrete_test":"For a random subset of ~50 personas and the abortion question, draw 500 free-form responses per persona from GPT-4 and from Llama-8B/70B/405B (temperature=1), independently parse the chosen option, and estimate the empirical answer distribution. Compare this to the logprob-derived distribution used in the paper using total variation distance; if the mean TVD over personas exceeds 0.1 for any model, the scoring method does not measure the model's true answer distribution. As a second arm, for Llama, rerun the constrained scoring with a different phrasing (e.g., 'I choose N') and check whether the normalized scores shift materially; large shifts would indicate the fixed-phrase scores are not a stable measure of option probability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the Prompts and Measures sections, the paper's entire quantitative argument depends on extracting each model's conditional distribution over the five answer options. For GPT-4, the authors say only that API logprobs are 'transformed into probabilities of answers,' but the logprobs parameter returns token-level probabilities for the generated completion and top alternatives, not a distribution over complete responses. A response can begin with '1', 'Option 1', 'The answer is 1', or a refusal; the probability of choosing option j is the sum over all surface forms, which is not equal to the first-token logprob of 'j' even with top-5 alternatives. For Llama, the method scores only the fixed assistant continuation 'The answer would be N' and normalizes these five scores to sum to 1. This treats those five strings as partitioning the response space, ignoring probability mass on other phrasings, refusals, or multi-sentence answers. If either conversion is biased, every value in Figures 2–7 and every accuracy, structural-consistency, and homogenization measure is distorted. The paper's limitation section does not flag this scoring validity issue, so the concern is not preempted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that LLM-based \"silicon samples\" misrepresent population opinions in two specific ways: a failure of structural consistency (accuracy at a coarse demographic level does not match the accuracy obtained by aggregating fine-grained persona-conditioned predictions) and homogenization (the variation ratio of LLM responses is far lower than that of human survey data). The authors measure these properties using GPT-4 and three Llama 3.1 models on two ANES 2020 multiple-choice questions, computing model answer distributions from API log probabilities. They report that LLMs achieve only moderate accuracy relative to a mode-prediction benchmark, violate structural consistency when aggregating across demographic granularity levels, and show severe homogenization compared with ANES respondents. They propose an \"accuracy-optimization hypothesis,\" supported by a simple formal argument that a predictor maximizing expected match rate should always output the modal answer.","tokens_in":16491,"tokens_out":5990,"duration_ms":61940,"significance":"If the measurement assumptions hold, the paper is a valuable contribution to the growing literature on LLM survey simulation. It provides a clear formal benchmark (mode optimality), a new diagnostic (structural consistency across aggregation levels), and a specific hypothesis linking homogenization to the model's optimization objective. The structural consistency test is a useful and nontrivial check that goes beyond aggregate accuracy comparisons, and the paper is transparent about its reliance on external ANES data rather than fitting parameters. The mode-optimality lemma is correct and well suited as a benchmark. However, the empirical findings all depend on the validity of the probability-extraction procedure, which is underdocumented and questionable, and the homogenization comparison does not account for sampling error in small ANES cells. These issues affect the central claims.","major_comments":[{"comment":"","section":"Models and Prompts"},{"comment":"","section":"Results, Homogeneity in Silicon Samples (Figures 6-7)"},{"comment":"","section":"Theoretical Framework, Accuracy Optimization Hypothesis"}],"minor_comments":[{"comment":"","section":"Title and Abstract"},{"comment":"","section":"Models"},{"comment":"","section":"Prompts"},{"comment":"","section":"Results, Structural Inconsistency"},{"comment":"","section":"Theoretical Framework"},{"comment":"","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important question, and the empirical design has clear merits. However, the core quantitative claims depend on the probability-extraction method, which is not adequately documented or justified, and the homogenization comparison ignores sampling error in small cells. These are fixable with additional analysis or a revised measurement approach, so I see this as a major revision rather than a rejection. The accuracy-optimization hypothesis, as presented, is more of an interpretive framing than a tested mechanism, and should be reframed accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read if you follow the silicon-samples debate. The paper's real contribution is the structural-consistency test: query the same model with personas built from 0, 1, 2, 3, and 4 demographic variables, aggregate the fine-grained results up to a coarse level, and compare with the directly queried coarse persona. Real survey data close under aggregation; these models don't. That is a clean, transportable diagnostic, and it isn't in Argyle, Bisbee, or Boelaert. The homogenization finding—severe overcommitment to the modal answer—replicates the compressed-variance critique, but the authors give it a sharper form by measuring log-probability distributions over all five options.\n\nThe formal piece is sound. The lemma that expected match rate is maximized by always answering the mode is correct, and using 'answer with mode' as the upper-bound benchmark is the right way to calibrate accuracy. The paper also engages the existing literature honestly and says clearly that this is a critical intervention, not a defense of silicon sampling. The citation pattern is appropriate; Bisbee et al. and Boelaert et al. get their due as prior compressed-variance critiques.\n\nThe biggest soft spot is measurement. For GPT-4, the paper says only that API logprobs were 'transformed into probabilities of answers.' The logprobs parameter returns token-level probabilities for generated tokens and top alternatives; the probability of 'choosing option 4' is a sum over surface forms ('4', 'Option 4', 'The answer is 4', refusals), not the first-token logprob of '4'. For Llama, they score only the fixed assistant continuation 'The answer would be N' and normalize the five scores to sum to one. That implicitly assumes those five strings partition the response space. If either conversion is biased, every accuracy, consistency, and homogenization statistic inherits the distortion. This is load-bearing, and the limitation section does not flag it. The stress-test note lands.\n\nSmaller issues: the VR<0.05 threshold is arbitrary (they admit it), and human variation ratios for tiny ANES subgroups are noisy, which may inflate the homogenization contrast at the margins. Minor compared with the scoring issue. Also, the accuracy-optimization hypothesis is plausible and consistent with the data, but the data don't distinguish it from the biased-sample hypothesis; it's a mechanism story, not a test.\n\nVerdict: the direction of the findings is almost certainly right—LLMs over-concentrate on modal opinions and don't aggregate consistently—and the structural-consistency diagnostic is worth having. But the missing code and parsing details make the quantitative claims provisional. A competent referee could fix this. I'd send it out, with a request for the full pipeline and reproducibility materials.","headline":"The structural-consistency diagnostic and mode-optimality benchmark are genuinely useful; the probability-scoring pipeline is under-specified enough that I'd want code and exact parsing before trusting the figures.","tokens_in":16973,"tokens_out":3326,"would_cite":true,"duration_ms":34167,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Large language models prompted to simulate survey respondents misrepresent real opinion distributions in two predictable ways: proportions do not survive demographic aggregation, and minority views are systematically flattened, because…","keywords":["silicon samples","large language models","representativeness","structural consistency","response homogenization","accuracy-optimization hypothesis","opinion surveys","ANES"],"falsifier":"Re-run the same 395-persona prompting protocol on GPT-4 and Llama 3.1 with sampling-based decoding instead of token log-probabilities — for example, draw 1,000 temperature-1 completions per persona and count the answer choices. If the resulting distributions match ANES variation ratios and aggregate consistently across demographic levels, the paper's structural-inconsistency and homogenization results would be artifacts of the scoring method rather than properties of the models.","tokens_in":16010,"feed_emoji":"🗳️","tokens_out":8301,"duration_ms":75496,"temperature":0.7,"pith_summary":"This paper argues that large language models cannot stand in for human survey respondents. Using two ANES 2020 questions on abortion and unauthorized immigration, the authors show that GPT-4 and Llama 3.1 break two requirements of a representative sample: structural consistency (the proportions of answers do not line up across demographic groupings) and response variation (minority opinions are severely underrepresented). The authors propose an accuracy-optimization hypothesis: because these models are trained to maximize the chance of predicting the respondent's answer, the mathematically optimal strategy is to always output the most common opinion for the given persona, which flattens within-group diversity. If correct, the finding undermines the 'silicon samples' approach and warns against using chatbot-generated opinions for statistical inference or policy.","feed_headline":"LLM 'silicon samples' flatten minority views, study finds","feed_subtitle":"GPT-4 and Llama 3.1 fail structural-consistency and variation tests against ANES survey data.","key_machinery":"The accuracy-optimization hypothesis is the central mechanism. It rests on a short proof: if the population holds beliefs $x_1,\\dots,x_n$ with proportions $p_1>\\dots>p_n$, and the model answers $x_i$ with probability $q_i$, the expected match rate is $\\sum_i p_i q_i$, which is at most $p_1$ and is maximized by setting $q_1=1$. So a model trained to maximize the chance of predicting its user's belief is driven to emit the modal answer, producing homogenization. The paper's other key instrument is the structural-consistency check: it queries the same questions at five levels of demographic granularity (no persona; sex; sex $\\times$ race; plus education; plus religion) and tests whether probability distributions at finer levels, aggregated using ANES weights, coincide with distributions obtained from direct queries at coarser levels. A real sample of individuals is closed under aggregation; the paper shows the LLMs are not.","core_discovery":"The paper's central claim is that LLM-generated 'silicon samples' systematically misrepresent population opinions even when conditioned on detailed demographic personae, and that this misrepresentation takes two specific forms. First, the models violate structural consistency: the probability distribution over answer options changes depending on how many demographic variables are used to define the persona, and aggregating fine-grained persona results back to a coarse level (for example, from sex × race × education × religion to sex) does not reproduce the distribution obtained by querying the coarse persona directly, something real survey data would always satisfy. Second, the models homogenize responses: across hundreds of demographic subgroups, the predicted probability is far more concentrated on a single modal answer than in the ANES data, erasing minority opinions. The paper formalizes the accuracy-optimization hypothesis, proving that for any belief distribution, the response pattern that maximizes the expected chance of matching an unknown respondent's belief is to always answer with the mode; this creates a structural incentive for next-token predictors to collapse onto modal opinions regardless of the diversity in their training knowledge.","pith_inferences":["If the structural-consistency failure is general, it suggests the LLM does not store a coherent joint distribution over demographics and opinions; each persona query reconstructs a context-specific stereotype from the same aggregate discourse, so the model is closer to a collection of separate 'Das Man' constructions than to a latent population model.","A testable extension of the accuracy-optimization hypothesis: asking the model to self-report a probability distribution ('give the percentage chance a person like this would choose each option') may reveal more within-group variation than decoding probabilities do; if so, part of the homogenization is a property of the elicitation or decoding procedure, not just of the latent distribution.","The paper's results suggest that the homogenization problem is not a static defect of current models: any future model trained with the same token-prediction objective will inherit the incentive to collapse onto modal answers, so fixes must change the training or sampling objective, for example through explicit diversity rewards, rather than merely enlarging the training corpus."],"forward_implications":["Survey researchers who use LLM-generated silicon samples should treat their outputs as stereotypes, not samples: reported subgroup means and regression coefficients will be based on distributions that are both too concentrated and not closed under aggregation.","The accuracy-optimization hypothesis predicts that bigger models or cleaner training data will not by themselves cure homogenization, because the incentive to answer with the mode is built into the next-token prediction objective.","Structural consistency offers a cheap diagnostic: before drawing conclusions from synthetic respondents, check whether fine-grained persona results, weighted and aggregated, reproduce the model's own coarse-grained answers; if they do not, the model does not encode a coherent population-level opinion distribution.","Apparent consistency across subgroups can be counterfeit: a model that gives the same dominant answer to every subgroup will look perfectly consistent and perfectly homogeneous at the same time, so consistency checks should be paired with variation checks."],"supporting_citations":[{"why":"Introduced the 'silicon sampling' method of conditioning LLMs on first-person demographic backstories, the approach this paper tests and extends.","marker":"Argyle et al. 2023"},{"why":"Found that LLM-generated opinion averages can resemble ANES averages but that regression coefficients are unreliable, motivating the variance and consistency checks here.","marker":"Bisbee et al. 2024"},{"why":"Distinguished the representative hypothesis from the social-bias hypothesis and documented lower between-subpopulation variance in LLM poll answers, which the paper builds on.","marker":"Boelaert et al. 2025"},{"why":"Supplies the ground-truth human opinion distributions for the two survey questions and the 395 demographic subgroups used as personae.","marker":"ANES 2020 Time Series Study"},{"why":"Compared LLM opinions against many US demographic subgroups, providing the subgroup-level comparison baseline that the paper's variation analysis refines.","marker":"Santurkar et al. 2023"},{"why":"The accuracy-diversity trade-off literature frames the homogenization hypothesis and the proof that the mode maximizes expected match rate.","marker":"Peng et al. 2023"}],"fun_headline_variants":["LLM 'silicon samples' erase minority opinions in surveys","Chatbots fail structural consistency when simulating populations","GPT-4 and Llama 3.1 homogenize survey responses, study finds","LLM personas collapse to modal answers, missing minority views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on measuring a model's 'true' answer distribution from the log-probability scores returned by the APIs (GPT-4's logprobs and Llama's summed token log-probabilities over constrained continuations); if those scores do not faithfully represent the probabilities the model would assign to the five answer options, every accuracy, structural-consistency, and homogenization statistic is distorted.","fun_headline_variants_meta":{"raw":{"variants":["LLM 'silicon samples' erase minority opinions in surveys","Chatbots fail structural consistency when simulating populations","GPT-4 and Llama 3.1 homogenize survey responses, study finds","LLM personas collapse to modal answers, missing minority views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2562,"prompt_tokens":950,"completion_tokens":1612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":1541}},"tokens_in":566,"tokens_out":1612,"duration_ms":11526,"temperature":1.0,"reasoning_tokens":1541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:50:14.330165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same 395-persona prompting protocol on GPT-4 and Llama 3.1 with sampling-based decoding instead of token log-probabilities — for example, draw 1,000 temperature-1 completions per persona and count the answer choices. If the resulting distributions match ANES variation ratios and aggregate consistently across demographic levels, the paper's structural-inconsistency and homogenization results would be artifacts of the scoring method rather than properties of the models.","supporting_citations":[],"review_version":1}