arxiv: 2602.04674 · v2 · submitted 2026-02-04 · 💻 cs.SI · cs.AI· cs.CL

Recognition: 2 theorem links

· Lean Theorem

Overstating Attitudes, Ignoring Networks: LLM Biases in Simulating Misinformation Susceptibility

Eun Cheol Choi , Lindsay E. Young , Emilio Ferrara

Authors on Pith no claims yet

Pith reviewed 2026-05-16 07:06 UTC · model grok-4.3

classification 💻 cs.SI cs.AIcs.CL

keywords large language modelsmisinformation susceptibilitysurvey simulationsocial networkscomputational social sciencebelief and sharingmodel biases

0 comments

The pith

LLM simulations overstate how strongly belief predicts sharing of misinformation and largely ignore personal networks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether LLMs prompted with real participant profiles from social surveys can reproduce observed human patterns of believing and sharing misinformation. LLM outputs match broad response distributions and show modest correlation with human answers, yet they consistently inflate the link between belief and sharing. Linear models fitted to the simulated data achieve higher explained variance but assign heavy weight to attitudinal and behavioral variables while downplaying network characteristics that matter in human data. The authors trace these distortions to systematic biases in how LLMs represent misinformation-related concepts in their training data.

Core claim

When LLMs are prompted with participant profiles drawn from social survey data that include network, demographic, attitudinal, and behavioral features, their generated responses capture broad distributional tendencies and exhibit modest correlation with human responses, but they overstate the association between belief and sharing; consequently, linear models fit to the simulated responses show substantially higher explained variance and place disproportionate weight on attitudinal and behavioral features while largely ignoring personal network characteristics.

What carries the argument

Prompting LLMs with participant profiles from surveys measuring network, demographic, attitudinal, and behavioral features to simulate patterns of misinformation belief and sharing.

If this is right

LLM-based survey simulations are better suited for diagnosing systematic divergences from human judgment than for substituting it.
Linear models on LLM data emphasize attitudinal and behavioral features at the expense of network characteristics.
Analyses of model-generated reasoning and training data indicate that distortions arise from how misinformation concepts are represented.
Human responses exhibit more balanced influence from personal networks than LLM-simulated responses do.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Social-science researchers relying on LLM proxies for network-influenced behaviors may reach conclusions that understate the role of social ties.
Prompt designs could be tested to force greater attention to network variables and reduce the observed attitude bias.
Policy models that use LLM simulations to forecast misinformation spread risk overemphasizing individual attitudes relative to relational factors.

Load-bearing premise

That prompting LLMs with participant profiles from social survey data will reproduce human patterns of misinformation belief and sharing without introducing systematic biases from the models' training data.

What would settle it

A replication showing that models fitted to additional human survey responses assign similar weight to network features as models fitted to LLM responses, or that the overstatement of the belief-sharing link disappears under different survey samples or prompting methods.

Figures

Figures reproduced from arXiv: 2602.04674 by Emilio Ferrara, Eun Cheol Choi, Lindsay E. Young.

**Figure 1.** Figure 1: Study pipeline. Human survey participants provided demographic, attitudinal, behavioral, and personal network in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: Example prompt format used to simulate public [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Simulation distorts the structure of feature importance relative to ground truth. Each point shows the proportion of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Standardized coefficients for simulation-by-feature interaction terms (Simulated [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 6.** Figure 6: Distribution of inferred association directions [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: Conceptual illustration of a possible pathway linking training data associations, model reasoning traces, and simulated [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 8.** Figure 8: Three meme stimuli used in climate change survey, [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

**Figure 9.** Figure 9: Example prompt format used to simulate climate [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

**Figure 12.** Figure 12: Illustrative example of a network feature block [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗

**Figure 11.** Figure 11: Illustrative example of demographic and attitudi [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗

**Figure 14.** Figure 14: Illustrative example of a network feature block [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗

**Figure 13.** Figure 13: Illustrative example of demographic and attitu [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗

**Figure 16.** Figure 16: Illustrative example of a network feature block [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗

**Figure 15.** Figure 15: Illustrative example of demographic and attitu [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗

**Figure 18.** Figure 18: Illustrative example of demographic, network, [PITH_FULL_IMAGE:figures/full_fig_p020_18.png] view at source ↗

**Figure 19.** Figure 19: Illustrative example of demographic, network, [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗

**Figure 20.** Figure 20: Relative explained variance after removing each feature block. This serves as an ablation-style analysis of feature [PITH_FULL_IMAGE:figures/full_fig_p022_20.png] view at source ↗

**Figure 21.** Figure 21: Relative explained variance after removing each feature block. Alternative profile ordering. [PITH_FULL_IMAGE:figures/full_fig_p023_21.png] view at source ↗

**Figure 22.** Figure 22: Relative explained variance after removing each feature block. Composite score profile format. [PITH_FULL_IMAGE:figures/full_fig_p023_22.png] view at source ↗

**Figure 23.** Figure 23: Standardized coefficients for simulation-by-feature interaction terms (Simulated [PITH_FULL_IMAGE:figures/full_fig_p024_23.png] view at source ↗

**Figure 24.** Figure 24: Standardized coefficients for simulation-by-feature interaction terms (Simulated [PITH_FULL_IMAGE:figures/full_fig_p024_24.png] view at source ↗

**Figure 25.** Figure 25: Prompt used to identify key explanatory variables from model-generated reasoning. [PITH_FULL_IMAGE:figures/full_fig_p026_25.png] view at source ↗

**Figure 26.** Figure 26: Prompt used for direction-of-association annotation. [PITH_FULL_IMAGE:figures/full_fig_p026_26.png] view at source ↗

read the original abstract

Large language models (LLMs) are increasingly used as proxies for human judgment in computational social science, yet their ability to reproduce patterns of susceptibility to misinformation remains unclear. We test whether LLM-simulated survey respondents, prompted with participant profiles drawn from social survey data measuring network, demographic, attitudinal and behavioral features, can reproduce human patterns of misinformation belief and sharing. Using three online surveys as baselines, we evaluate whether LLM outputs match observed response distributions and recover feature-outcome associations present in the original survey data. LLM-generated responses capture broad distributional tendencies and show modest correlation with human responses, but consistently overstate the association between belief and sharing. Linear models fit to simulated responses exhibit substantially higher explained variance and place disproportionate weight on attitudinal and behavioral features, while largely ignoring personal network characteristics, relative to models fit to human responses. Analyses of model-generated reasoning and LLM training data suggest that these distortions reflect systematic biases in how misinformation-related concepts are represented. Our findings suggest that LLM-based survey simulations are better suited for diagnosing systematic divergences from human judgment than for substituting it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LLMs overstate belief-sharing links and downplay networks in misinformation sims, but the prompt format may be doing some of the work.

read the letter

The key point is that when you feed LLMs participant profiles from real surveys, the outputs roughly track human response distributions but then produce regressions where attitudes and behaviors explain far more variance in belief and sharing than they do in the actual human data, while network features get almost no weight. That distortion is the main new observation here, and it lines up with the abstract's claim that these models are better for spotting where they diverge from people than for replacing them outright. The comparison against three independent human surveys is the part that actually grounds the argument, and checking the model's generated reasoning plus training-data patterns adds a bit of mechanistic color without overclaiming. That combination is worth having on record for anyone running LLM-based social simulations in this area. The soft spot is the attribution to training-data biases. The stress-test note is right that we need the exact prompt templates to know whether network details were presented in dense free-text while attitudes came as clean Likert items; if the encoding was uneven, the coefficient gap could be an attention artifact rather than a deep representation issue. The abstract does not give sample sizes, exact statistical controls, or error bars, so the strength of the evidence depends on how cleanly those are handled in the full text. If the paper shows they balanced the prompt structure and still saw the same pattern, the training-data story holds better. This is for computational social scientists who are already using or considering LLMs as survey proxies, especially on misinformation topics. It is not a takedown but it is a useful cautionary result. I would send it to peer review because the core comparison is straightforward and the limitation it flags is practically relevant, even if the causal explanation needs tighter controls on prompting.

Referee Report

3 major / 2 minor

Summary. The paper claims that LLMs prompted with participant profiles from real social surveys (including network, demographic, attitudinal, and behavioral features) reproduce broad distributional tendencies in misinformation belief and sharing and show modest correlation with human responses, but systematically overstate the belief-sharing association. Linear models on LLM outputs exhibit substantially higher explained variance, overweight attitudinal/behavioral features, and largely ignore network characteristics relative to models fit to human data; the authors attribute these distortions to systematic biases in how misinformation concepts are represented in LLM training data.

Significance. If the central findings hold after addressing methodological details, this work is significant for computational social science: it supplies concrete evidence that LLM-based survey simulation is better suited to diagnosing divergences from human judgment than to substituting for it, particularly when recovering feature-outcome associations in misinformation research. The use of independent external human survey baselines as comparators is a strength that grounds the evaluation.

major comments (3)

[Methods] Methods section: the paper does not supply the exact prompt templates or feature-encoding details (e.g., whether network characteristics are conveyed via dense free-text paragraphs while attitudes use explicit Likert-style statements). This omission prevents ruling out prompt-format confounds as the source of the observed under-weighting of networks, which is load-bearing for the claim that the distortion reflects training-data biases rather than input representation.
[Results] Results section: the claims of 'substantially higher explained variance' and 'disproportionate weight' on attitudinal/behavioral features require explicit reporting of R² values, coefficient tables, and statistical tests comparing LLM-fitted versus human-fitted linear models. Without these quantities and error bars, the magnitude and robustness of the reported disparities cannot be verified.
[Discussion] Discussion section: the attribution of network under-weighting to training-data biases on misinformation concepts rests on analyses of model-generated reasoning and LLM training data, yet the manuscript provides no description of the methodology, sample, or controls used in those analyses. This leaves open whether alternative explanations (prompt encoding, context length) have been adequately tested.

minor comments (2)

Add a table or figure that directly juxtaposes the feature coefficients (with standard errors) from the human and LLM linear models side-by-side for each survey.
Clarify the exact sample sizes, number of LLM generations per profile, and any temperature or decoding settings used in the simulation experiments.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments highlight important areas for improving clarity, reproducibility, and rigor. We address each major comment below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Methods] Methods section: the paper does not supply the exact prompt templates or feature-encoding details (e.g., whether network characteristics are conveyed via dense free-text paragraphs while attitudes use explicit Likert-style statements). This omission prevents ruling out prompt-format confounds as the source of the observed under-weighting of networks, which is load-bearing for the claim that the distortion reflects training-data biases rather than input representation.

Authors: We agree that the exact prompt templates and feature-encoding details are necessary for full reproducibility and to evaluate potential confounds. The original manuscript omitted these for brevity, but we will add them in a new appendix (or expanded Methods section) in the revision. This will include the complete prompt templates for each of the three surveys, along with explicit descriptions of how network characteristics (e.g., free-text summaries of ego-network size and composition), demographics, attitudes (e.g., Likert-scale statements), and behaviors were encoded. These additions will allow direct assessment of whether input format contributed to the observed feature weighting differences. revision: yes
Referee: [Results] Results section: the claims of 'substantially higher explained variance' and 'disproportionate weight' on attitudinal/behavioral features require explicit reporting of R² values, coefficient tables, and statistical tests comparing LLM-fitted versus human-fitted linear models. Without these quantities and error bars, the magnitude and robustness of the reported disparities cannot be verified.

Authors: We accept that the Results section would be strengthened by more granular quantitative reporting. In the revised manuscript, we will include a new table (or expanded existing tables) reporting exact R² values for all linear models fit to LLM and human data, full coefficient tables with standard errors and confidence intervals, and formal statistical comparisons (e.g., tests for differences in R² or coefficient magnitudes across models). This will provide the necessary evidence to substantiate the claims of substantially higher explained variance and disproportionate weighting on attitudinal/behavioral features. revision: yes
Referee: [Discussion] Discussion section: the attribution of network under-weighting to training-data biases on misinformation concepts rests on analyses of model-generated reasoning and LLM training data, yet the manuscript provides no description of the methodology, sample, or controls used in those analyses. This leaves open whether alternative explanations (prompt encoding, context length) have been adequately tested.

Authors: The manuscript's Discussion references qualitative examination of model-generated reasoning chains and patterns in LLM pretraining data coverage of misinformation-related concepts, but we acknowledge that a detailed methodological description was not provided. In the revision, we will expand the relevant subsection (and add an appendix if needed) to describe the sample of responses analyzed, the coding procedure for identifying network references, and any controls or checks performed. We will also explicitly discuss and report sensitivity analyses addressing alternative explanations such as prompt encoding and context length to strengthen the attribution to training-data biases. revision: partial

Circularity Check

0 steps flagged

No circularity: external human survey baselines anchor all comparisons

full rationale

The paper evaluates LLM-simulated responses against three independent online survey datasets that supply the participant profiles and serve as the human baseline. Linear models are fit separately to the LLM outputs and to the human responses; the resulting coefficients and explained variances are then compared directly. No quantity is fitted to a subset of the LLM data and then presented as a prediction of the same data. The attribution of observed divergences to training-data biases is supported by separate analyses of model-generated reasoning traces and publicly available LLM training corpora, none of which are derived from the simulation results themselves. Because the central claims rest on external, falsifiable benchmarks rather than self-referential definitions, fitted inputs renamed as predictions, or load-bearing self-citations, the derivation chain contains no circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of the three online surveys as accurate representations of human patterns and on the assumption that LLM prompting can capture human-like responses without training-data artifacts.

axioms (1)

domain assumption The three online surveys accurately represent human patterns of misinformation belief and sharing.
Used as ground-truth baselines for all comparisons of LLM outputs.

pith-pipeline@v0.9.0 · 5493 in / 1237 out tokens · 43229 ms · 2026-05-16T07:06:44.293898+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Linear models fit to simulated responses exhibit substantially higher explained variance and place disproportionate weight on attitudinal and behavioral features, while largely ignoring personal network characteristics
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Analyses of model-generated reasoning and LLM training data suggest that these distortions reflect systematic biases in how misinformation-related concepts are represented

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages · 2 internal anchors

[1]

InForty-second International Conference on Machine Learning Position Paper Track

Position: LLM social simulations are a promising re- search method. InForty-second International Conference on Machine Learning Position Paper Track. Ashery, A. F.; Aiello, L. M.; and Baronchelli, A. 2025. Emergent social conventions and collective bias in LLM populations.Science Advances, 11(20): 1–10. Battaglia, P. W.; Hamrick, J. B.; Bapst, V .; Sanche...

work page 2025
[2]

Relational inductive biases, deep learning, and graph networks

Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261. Borah, A.; Mihalcea, R.; and P ´erez-Rosas, V . 2025. Per- suasion at play: Understanding misinformation dynamics in demographic-aware human-LLM interactions.arXiv preprint arXiv:2503.02038. Chang, S.; Chaszczewicz, A.; Wang, E.; Josifovska, M.; Pierson, E.; an...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

distribution copies

LLaGA: Large language and graph assistant. InPro- ceedings of the 41st International Conference on Machine Learning, 7809–7823. Chew, L. D.; Griffin, J. M.; Partin, M. R.; Noorbaloochi, S.; Grill, J. P.; Snyder, A.; Bradley, K. A.; Nugent, S. M.; Baines, A. D.; and VanRyn, M. 2008. Validation of screen- ing questions for limited health literacy in a large...

work page arXiv 2008
[4]

Images Amplify Misinformation Sharing in Vision-Language Models

Large-language-model-powered agent-based frame- work for misinformation and disinformation research: Op- portunities and open challenges.IEEE Security & Privacy, 22(3): 24–36. Pennycook, G.; and Rand, D. G. 2021. The psychology of fake news.Trends in Cognitive Sciences, 25(5): 388–402. Plebe, A.; Douglas, T.; Riazi, D.; and del Rio-Chanona, R. M. 2025. I’...

work page internal anchor Pith review Pith/arXiv arXiv 2021
[5]

Ivermectin pills, known as the antiparasitic drug, have been approved by the FDA to treat COVID-19

work page
[6]

Vaccinated individuals emit Bluetooth signals Public Health

work page
[7]

Most generic sunscreens on the market contain benzenes which are a cancer-causing agent

work page
[8]

Diabetes can be treated by wearing a copper bracelet

work page
[9]

WHO has said smoking prevents people from getting infected with the novel coronavirus

work page
[10]

Climate Change

The increase in the global polar bear population from about 5,000 in the 1960s to over 25,000 to- day proves that global warming is exaggerated or a hoax. Climate Change

work page
[11]

while let- ting China and India pollute more, making it use- less for protecting the environment

The Paris Climate Treaty hurts the U.S. while let- ting China and India pollute more, making it use- less for protecting the environment

work page
[12]

A single eruption of Mount Etna releases more carbon dioxide than all human activity combined

work page
[13]

The government is deliberately minimizing the group of people eligible for COVID-19 diagnostic test to reduce the number of confirmed cases before the election

work page
[14]

Pandemic Politics

The government has exclusive control over COVID-19 clinical information, refusing to share them with the experts. Pandemic Politics

work page
[15]

Purchasing public masks at a pharmacy will lead to leakage of personal information which will be used in election fraud

work page
[16]

There was a shortage in mask at hospitals in Ko- rea because the government sent masks to China

work page
[17]

Table 4: Misinformation items used in the study

The government is providing masks purchased with tax money to China. Table 4: Misinformation items used in the study. Figure 8: Three meme stimuli used in climate change survey, namely (i) stop global warming hype (top left), (ii) Paris Cli- mate Treaty (right), and (iii) Mount Ena CO2 (bottom left). A.2. Descriptive Statistics VariablesM(SD) or % Demogra...

work page
[18]

Ivermectin cures covid 2.55 (1.80)

work page
[19]

Vaccinated emit signals 1.33 (1.08)

work page
[20]

Sunscreen and cancer 3.75 (1.56)

work page
[21]

Bracelet cures diabetes 1.49 (1.17)

work page
[22]

Smoking prevents covid 1.59 (1.33) Sharing Intention 2.18 (1.25)

work page
[23]

Ivermectin cures covid 2.44 (2.06)

work page
[24]

Vaccinated emit signals 1.57 (1.51)

work page
[25]

Sunscreen and cancer 3.58 (2.25)

work page
[26]

Bracelet cures diabetes 1.68 (1.60)

work page
[27]

Smoking prevents covid 1.63 (1.52) Table 5: Descriptive statistics of public health survey. VariablesM(SD) or % Demographics Female 59.95% White 77.72% Age 45.88 (13.52) Education 4.22 (1.51) Income 3.48 (1.59) Attitudinal/Behavioral Political identification 3.49 (1.27) Systemic Processing 3.94 (0.47) Primary source: legacy media 2.79 (1.74) Primary sourc...

work page
[28]

Stop climate change hype 40.32%

work page
[29]

Paris Climate Treaty 58.62%

work page
[30]

Mount Ena CO2 50.40% Sharing Intention 0.23 (0.35)

work page
[31]

Stop climate change hype 23.61%

work page
[32]

Paris Climate Treaty 24.14%

work page
[33]

Mount Ena CO2 20.42% Table 6: Descriptive statistics of climate change survey. VariablesM(SD) or % Demographics Female 46.75% Seoul metropolitan area 83.42% Age 47.21 (13.03) Education 5.60 (1.02) Income 4.97 (1.98) Attitudinal/Behavioral Political identification 4.80 (1.90) Trust in science 6.32 (1.57) Social media use 4.11 (1.16) Health literacy 3.26 (0...

work page
[34]

Gov’t minimizes cases 1.80 (0.87)

work page
[35]

Gov’t controls clinic info 1.81 (0.83)

work page
[36]

Masks and election fraud 1.71 (0.88)

work page
[37]

Mask shortage due to China 2.29 (1.06)

work page
[38]

Give masks away to China 2.15 (1.00) Sharing Intention 1.27 (0.44)

work page
[39]

Gov’t minimizes cases 1.19 (0.47)

work page
[40]

Gov’t controls clinic info 1.30 (0.56)

work page
[41]

Masks and election fraud 1.43 (0.65)

work page
[42]

Mask shortage due to China 1.21 (0.49)

work page
[43]

unknown” or similar. Output format (JSON): {“response

Give masks away to China 1.21 (0.49) Table 7: Descriptive statistics of pandemic politics survey. Appendix B. Divergence and Correlation between Human and Simulated Susceptibility Climate Change (US, 2025) Public Health (US, 2023) Pandemic Politics (KR, 2020) Misinfo Belief Misinfo Sharing Misinfo Belief Misinfo Sharing Misinfo Belief Misinfo Sharing JS↓E...

work page 2025
[44]

I agree with the information above

work page
[45]

OR Question: Please choose the option below that best describes your intention to share the above message with other people

I disagree with the information above. OR Question: Please choose the option below that best describes your intention to share the above message with other people

work page
[46]

I am likely to share it with others

work page
[47]

unknown” or similar. Output format (JSON): {“response

I am unlikely to share it with others. Figure 9: Example prompt format used to simulate climate change survey participants’ misinformation belief or sharing intention. Belief and sharing questions were queried sepa- rately using identical system prompts, participant profiles, and claims. C.2. Pandemic Politics Simulation Prompt Format SYSTEMThis survey wa...

work page 2020