Pith. sign in

REVIEW 4 major objections 5 minor 10 references

Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read The paper argues that an LLM can simulate whether a person will believe a misinformation claim far better when seeded with the beliefs typical of that person's demographic group than when told only their demographics, reaching up to 92 perc

desk verdict Useful evaluation toolkit and honest leakage discussion, but the headline 'beliefs beat demographics' claim is undermined because the imputed beliefs are demographic group averages, not independent belief signals. read the letter →

arxiv 2603.03585 v2 pith:QKJC6TAM submitted 2026-03-03 cs.CL cs.AI

classification cs.CLcs.AI
keywords misinformationsusceptibilityLLMsimulationbeliefpriorsdemographicpersonasWorldValuesSurveyalignmentcounterfactualevaluationtwo-phasefine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

BeliefSim is a framework for making LLM personas simulate demographic misinformation susceptibility by conditioning them on belief priors rather than demographics alone. The paper claims that survey-derived group-level beliefs provide a strong prior: adding them raises alignment with human true/false judgments over zero-shot and demographic-only prompting, with imputed beliefs outperforming individually observed ones. A two-phase fine-tuning variant that first learns to predict demographic belief distributions and then learns susceptibility from a frozen belief representation transfers across participant pools, reaching up to 92.4 percent accuracy while nearly eliminating demographic label flips. The reason this matters is that LLMs are increasingly used as cheap, scalable testbeds for studying and intervening on misinformation, and the paper argues those simulations are more faithful when beliefs, not surface demographics, drive the persona.

What carries the argument

The machinery is BeliefSim-Tax, a seven-dimension taxonomy of beliefs—worldview/identity, epistemic trust, cognitive style, conspiracy mentality, morals/values, emotion, and heuristics—that organizes 126 World Values Survey items; the imputed demographic belief priors derived as modal responses per group; BeliefSim-PC, which injects those priors into prompts; and BeliefSim-FT, a two-phase design with a frozen belief adapter and a trainable susceptibility head. The taxonomy's job is to give the model interpretable, compact belief structure, while the two-phase design prevents the model from learning demographic label shortcuts during susceptibility training.

What would settle it

Give PANDORA and MIST participants the WVS belief items used for imputation, then check whether the imputed modal belief vector matches their actual answers and whether replacing imputed beliefs with true individual beliefs changes alignment; if the imputed vectors are uncorrelated with real answers yet still 'improve' alignment, the central claim fails.

Watch

Extended reading notes

Core claim

Beliefs are a stronger prior than demographics for simulating whether an individual will accept a false claim. The paper shows that prompting an LLM with the modal World Values Survey belief responses of a demographic group outperforms prompting with demographic labels alone, and that combining imputed beliefs with demographics gives the best prompt-based results. It then shows that decoupling belief modeling from susceptibility prediction—training a belief adapter on survey distributions first, then a lightweight susceptibility head on top of the frozen representation—improves cross-study transfer to a standardized instrument while reducing demographic shortcut flip rates to near zero. The

Load-bearing premise

The load-bearing premise is that the modal World Values Survey beliefs of a demographic group approximate the actual beliefs of the PANDORA and MIST participants in that group; if the imputed beliefs do not match those participant pools, the claimed advantage of beliefs over demographics is an encoding artifact.

Editorial extensions

If this is right

  • Adding belief priors improves susceptibility alignment over zero-shot and demographic-only prompting, with imputed group priors consistently stronger than sparse observed beliefs.
  • Compact single-dimension belief priors (notably emotion and moral-values) can outperform the full taxonomy stuffed into one prompt, suggesting that belief selection matters.
  • Demographics alone are a weak prior: demo-only prompting can degrade zero-shot performance, and counterfactual swaps show model-dependent stereotype-like shortcut reliance.
  • Two-phase, head-only adaptation transfers substantially better across participant pools (up to 92.4% on MIST-2) than one-phase fine-tuning, while cutting flip rates to near zero.
  • Belief conditioning partially decouples prediction from the model's pretrained factual-confidence bias, confirming that simulating susceptibility is a different task than predicting veracity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the central premise would be to collect WVS-style belief answers from the actual PANDORA/MIST participants; if those answers do not match the imputed modal beliefs, the belief advantage may be a re-encoded demographic stereotype rather than evidence that beliefs drive susceptibility.
  • The taxonomy invites an intervention design: if moral-values and emotion beliefs are the strongest single priors, then targeting those belief dimensions in corrections or education may be more effective than targeting demographic groups themselves.
  • Because gender showed the largest sensitivity to which belief dimension is chosen while overall gender effects were small, simulation studies should avoid treating gender as a uniform belief proxy.
  • The MIST-2 result reflects cross-participant transfer under a shared item pool, not fully claim-disjoint generalization; a stricter holdout of unseen claims would be the natural next test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces BeliefSim, a framework for simulating demographic misinformation susceptibility with LLMs by conditioning on belief profiles constructed from demographic survey priors. BeliefSim-Tax organizes beliefs into seven psychology-informed dimensions; BeliefSim-PC adds demographic and belief information to prompts; BeliefSim-FT uses a two-phase fine-tuning procedure that first learns demographic-conditioned WVS response distributions and then trains a lightweight susceptibility head. The authors evaluate susceptibility alignment on PANDORA and MIST-1, and cross-study transfer on MIST-2, and also propose counterfactual flip-rate analyses to separate demographic utility from stereotype reliance. The central claim is that 'beliefs provide a strong prior for simulating misinformation susceptibility, with alignment up to 92%'.

Significance. If the claim held as stated, the paper would be a useful step toward belief-grounded rather than purely demographic LLM simulation, with practical implications for misinformation research and intervention design. The paper has genuine strengths: a multi-model evaluation across two datasets, a thoughtful distinction between veracity and susceptibility prediction, a counterfactual evaluation protocol that goes beyond accuracy, and an open-source release of the framework. Section D.1 is also commendable for honestly acknowledging the MIST-1/MIST-2b claim-overlap issue. However, the central comparison is undermined by two load-bearing problems: (1) imputed beliefs are a deterministic re-encoding of demographic group, so the beliefs-vs-demographics comparison is confounded with representation richness; and (2) the headline prompt-based gains are obtained by selecting the best belief dimension on the evaluation set. These issues must be fixed before the paper's central claim is acceptable.

major comments (4)
  1. [§4.2, §5.1, Fig. 3] Imputed beliefs are WVS Wave 7 U.S. modal responses conditioned on the same four demographic axes used in PANDORA/MIST. The imputed-belief vector is therefore a deterministic function of demographic group. 'Imputed+Demo' is not an independent belief signal; it is a 126-item demographic profile, while 'Demo-only' is a 4-axis label. The comparison in Fig. 3 thus tests representation richness, not beliefs vs. demographics, and the Sec. 5.1 assertion that 'what people believe may be more predictive than who they are' does not follow. Please validate the imputed priors against the actual beliefs of the PANDORA/MIST participants (e.g., by collecting WVS-style belief items from those samples) or include a content control that replaces WVS belief items with equally long demographic-conditioned but belief-irrelevant survey content. Without such a control, the belief-driven interpretation remains
  2. [§5.1, Fig. 10, Appendix C.1] Fig. 3 and Sec. 5.1 report 'best' imputed settings, where the belief dimension is selected by the highest accuracy on the evaluation data. This is test-set selection and overstates expected gains. The fuller Fig. 10 shows that on MIST the average imputed-only condition is 88.7%, below the zero-shot baseline of 88.9%, so the caption claim 'all belief-based settings do better than zero-shot and demo-only' is false for the average condition. Likewise, Sec. 5.1's 'screening outlier modal runs' needs an a-priori exclusion rule. Please report means over belief dimensions and select dimensions on a validation split; otherwise the +1–16 point improvements are upper bounds rather than robust evidence.
  3. [§6, Table 3, §D.1] The 'up to 92.4%' cross-study result is obtained on MIST-2b, which, as D.1 acknowledges, contains items derived from the MIST-1 item-development pool. Since the model is trained on MIST-1, part of the improvement could come from memorizing instrument-specific wording rather than from transferable belief-susceptibility structure. The authors are transparent in D.1, but the abstract and conclusion do not carry the qualification. Please report a claim-disjoint evaluation (e.g., held-out MIST items never seen in training) or prominently state in the abstract/conclusion that the 92% figure is shared-instrument cross-participant transfer, not claim-disjoint generalization.
  4. [§4.2, §5.1, §C.7] Observed beliefs are held-out judgments from the same participant whose target claim is being predicted. This condition gives the model an individual-level anchor: the participant's response pattern on other claims. Gains from 'observed beliefs' may therefore reflect participant identity, not belief content as a psychological construct. To support the belief-driven interpretation, add a control in which observed beliefs are randomly reassigned across participants. If the gain persists under random reassignment, it is not identity; if it disappears, the observed-belief results are an artifact of participant-level leakage, albeit not label leakage. This distinction is important because the paper uses 'observed' and 'imputed' interchangeably as 'beliefs' in the summary of results.
minor comments (5)
  1. [§2] 'Relevant relevant to our work' is a duplicated word; please edit.
  2. [§5, Table 2] The caption and text contain typos: 'Qwn' should be 'Qwen'; 'Deepseek' vs. 'DeepSeek' should be consistent.
  3. [Appendix C.2] The text refers to 'Prolific' and 'Dataset 1' inconsistently; use the dataset name PANDORA throughout.
  4. [General] The paper states that the framework is open-sourced (Appendix G), but the URL is an anonymous link. If the link is anonymized for review, that is fine; if not, provide a contactable repository.
  5. [Appendix E] The paired t-tests are described only as 'belief settings are statistically significant'; please specify which settings were included and how the best-dimension selection interacted with the test.

Circularity Check

2 steps flagged · score 4.0 of 10

BeliefSim's central 'belief-driven' claim is partially constructed: imputed 'beliefs' are demographic-conditioned WVS profiles, and the prompt-based 'best' results are selected on the evaluation set.

  1. fitted input called prediction [Section 5.1, Figure 3 caption]
    "For cases that include Imputed beliefs, we only report best results by selecting the belief dimension that achieves the highest accuracy."

    The reported best imputed and imputed+demo alignments (e.g., 65.6% PANDORA, 90.9% MIST) are maxima over seven belief dimensions evaluated on the same test set. The selection criterion is the evaluation metric itself, so these headline numbers are by construction the largest of several test-set accuracies rather than unbiased predictions of the framework. This makes the 'beliefs improve alignment' result partly a product of test-set selection, though averaged and fine-tuning results are also reported.

  2. other [Section 4.2 (Belief Data); Section 5.1 (Results)]
    "Imputed data represent demographic belief priors (group–level), inferred from WVS, conditioned solely on demographic attributes. ... This shows that what people believe may be more predictive than who they are."

    The imputed belief profile is a deterministic function of the same demographic axes used in the demo-only condition (gender, age, living area, education). Therefore the 'beliefs-only (imputed)' condition is not belief information independent of demographics; it is a 126-item demographic-conditioned survey profile. The central Fig. 3 comparison thus reduces to a fine-grained demographic encoding versus four coarse demographic categories. The conclusion that beliefs outperform demographics is not established by this contrast, since the belief signal is defined in terms of the very demographic categories it is compared against.

full rationale

BeliefSim is largely an empirical comparison anchored to external WVS survey priors and held-out participant judgments, so it does not reduce to a self-citation chain or to a derivation whose output equals its input. The fine-tuning result (BeliefSim-FT up to 92.4% on MIST-2) is a genuine cross-study, cross-participant transfer evaluation, and observed beliefs are held-out claim judgments from the same participants rather than the target label. However, two issues make the 'belief-driven' framing partially constructed. First, the prompt-based 'best' imputed results are selected by taking the highest accuracy over belief dimensions on the evaluation set, so the reported maxima are test-set selections rather than out-of-sample predictions. Second, the imputed beliefs themselves are WVS modal responses conditioned solely on the same four demographic axes used in the demo-only baseline; hence the 'beliefs vs. demographics' comparison is confounded by construction, and the abstract's causal reading ('beliefs provide a strong prior') is stronger than the evidence supports. A minor self-citation to Borah et al. (2025a) supports but does not force the central claim. Overall, the paper's empirical content is substantial, but the headline interpretive claim is not fully earned by the experimental design.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central comparisons rest on survey-to-sample transfer assumptions, modal-response summarization, and post-hoc best-dimension selection. The taxonomy mapping is not externally validated, and the binary demographic discretizations are chosen for convenience.

free parameters (3)
  • Belief dimension selection ('best' setting) = Varies by dataset/model; e.g., in Appendix C.2, conspiracy-minded or cognitive-style dimensions often score highest
    The main figures report Imputed (best) and Imputed+Demo (best), where the belief dimension is selected as the one that maximizes evaluation accuracy. This is a post-hoc model choice fitted to the test set.
  • Modal response summarization = Most frequent WVS response per demographic group per question
    Belief priors are represented as modal responses rather than full distributions; the paper states this is because it is 'more stable and compact' (§5.1), but this is a modeling choice that discards uncertainty and may not be the best representation.
  • Demographic group cutoffs = Age <=35 vs >=60; education completed/not completed high school; rural/urban; binary gender
    The binary demographic axes are chosen for analytic convenience (§4) and affect both the belief profiles and the counterfactual evaluation, but no principled basis for the cutoffs is given.
assumptions (5)
  • domain assumption WVS Wave 7 US demographic response distributions are valid belief priors for PANDORA and MIST participant pools.
    Invoked in §4.2 when imputed belief distributions are mapped from WVS and used as demographic priors; no validation against participants' actual WVS-equivalent beliefs is provided.
  • domain assumption LLM persona prompting can meaningfully simulate human susceptibility judgments.
    The entire framework relies on the premise that LLM predictions under demographic/belief conditioning are informative about human judgments; this is the research question but also a background assumption for the method.
  • domain assumption A binary true/false judgment on a claim is an adequate operationalization of misinformation susceptibility.
    The evaluation datasets provide binary participant judgments (§4.1), and the model is trained and scored on this binary target, treating it as 'susceptibility' rather than a multi-dimensional construct.
  • ad hoc to paper The seven BeliefSim-Tax dimensions, derived via exploratory factor analysis, are the right factorization of belief space for misinformation susceptibility.
    The taxonomy is compiled from psychology literature and WVS items (§3, Appendix A.2), but the specific seven-category mapping is paper-specific and not externally validated.
  • domain assumption MIST-2b, sharing items with MIST-1, is a valid cross-study transfer target.
    Appendix D.1 acknowledges MIST-2b contains items from the MIST-1 development pool and frames results as 'shared-instrument cross-participant transfer'; this is a weaker claim than fully disjoint generalization.
invented entities (1)
  • BeliefSim-Tax
    purpose: A seven-dimensional taxonomy of beliefs (worldview, epistemic trust, cognitive style, conspiracy mentality, morals/values, emotion, heuristics) used to organize WVS questions and condition LLM personas.
    The taxonomy is a constructed categorization with no external validation that these seven dimensions are the correct or complete factorization for misinformation susceptibility; it is introduced by the paper and used throughout.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility." pith.science (2026). https://pith.science/paper/QKJC6TAM

@misc{pith2026260303585,
  author       = {Pith},
  title        = {Pith review of: Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QKJC6TAM}},
  note         = {Machine review of arXiv:2603.03585}
}
read the original abstract

Misinformation is a growing societal threat, and susceptibility to misinformative claims varies across demographic groups due to differences in underlying beliefs. As Large Language Models (LLMs) are increasingly used to simulate human behaviors, we investigate whether they can simulate demographic misinformation susceptibility, treating beliefs as a primary driving factor. We introduce BeliefSim, a simulation framework that constructs demographic belief profiles using psychology-informed misinformation taxonomies and survey priors. We study prompt-based conditioning and post-training adaptation, and conduct a multi-fold evaluation using: (i) susceptibility alignment and (ii) counterfactual demographic sensitivity. Across both datasets and modeling strategies, we show that beliefs provide a strong prior for simulating misinformation susceptibility, with alignment up to 92%.

Figures

Figures reproduced from arXiv: 2603.03585 by the authors.

Figure 1
Figure 1. BeliefSim Framework. (1) Participant Data, Observed and Imputed Beliefs (mapped to BeliefSim￾Tax) are collected from surveys, (2) Methods consist of prompt-conditioning and post-training adaptation and (3) Evaluation is performed using Susceptibility Align￾ment, Counterfactual and Thematic Analysis. underlying beliefs is essential to explain misinfor￾mation susceptibility, i.e., how likely someone is to believe a mi… view at source ↗
Figure 2
Figure 2. Data Example. We map participant demo￾graphics to WVS responses to derive group-level im￾puted beliefs, and use claim-level evaluations as partic￾ipant observed beliefs These are then used to predict observed beliefs. 3 BeliefSim-Tax Beliefs are an important tool for modeling demo￾graphic simulations for misinformation. Prior work has shown that belief dimensions, such as conspir￾acy beliefs or trust in science, can… view at source ↗
Figure 3
Figure 3. Susceptibility Alignment, averaged across demo￾graphic groups and models. Imputed + Demo(graphic) (best) performs the best. All belief-based settings do better than zero￾shot and demo-only. ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: Flip-Rates for Counterfactual Evaluation: Mis￾tral and Qwn models have lower flip-rates whereas Llama and Deepseek have higher rates. Darker (red) colors mean higher flip rates while lighter (yellow) colors mean lower flips. this steering improves performance. Results …
Figure 4
Figure 4. Figure 4: Demographic-based Counterfactual Evalu￾ation. Note that we perform a demographic swap within the same group, i.e., swap male with female, rural with urban, etc. keeping the claim constant. susceptibility. But they can also induce spurious signals that appear accurate b…
Figure 6
Figure 6. Figure 6: Complementarity Analysis. Adding demograph￾ics has mixed effects (mostly negative) when partial belief information is added. Positive values mean that adding demo￾graphics increases accuracy, while negative values mean that it decreases accuracy. For complementarity an…
Figure 7
Figure 7. Figure 7: BeliefSim-FT framework. Green-shaded com￾ponents correspond to Phase 1 (Belief Modeling), and blue￾shaded components are Phase 2 (Susceptibility Fine-Tuning). concatenate with the claim representation hϕ(x) and predict susceptibility via a binary classifier: pψ(y | x, …
Figure 8
Figure 8. Figure 8: Dataset Examples - MIST and PANDORA C Prompt-Based Conditioning C.1 Prompt Details [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Prompt Conditioning Experiments 50 60 70 80 90 100 Alignment (%) Zero-shot Demo-only Imputed Imputed (best) Observed Imp+Obs Imp+Obs (best) Imp+Demo Imp+Demo (best) Obs+Demo Imp+Obs+ Demo Imp+Obs+ Demo(best) Setting 56.8 56.2 61.9 65.4 58.7 59.0 59.3 63.1 65.6 58.2 60.…
Figure 10
Figure 10. Figure 10: Susceptibility Alignment across settings and [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Prompt for Phase-1 Belief Modeling unseen belief items, reporting KL distributional fit and majority-category accuracy. Phase 2 trains a susceptibility prediction head on top of a frozen belief adapter learned in Phase 1. The model combines a frozen base encoder and f…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 5 linked inside Pith

  1. [1]

    These two are concatenated and fed into a lightweight, train- able classification head

    The model combines a frozen base encoder and frozen belief head to produce both a seman- tic representation of the input and a demographic- conditioned belief probability vector. These two are concatenated and fed into a lightweight, train- able classification head. We format the inputs as instruction-following prompts that include a de- mographic persona...

  2. [2]

    The 1- phase baseline model perform decent, however, the performance drops sharply on MIST-2

    generalization improves drastically once we move from 1-phase to 2-phase pipelines. The 1- phase baseline model perform decent, however, the performance drops sharply on MIST-2. For MIST-2, our BeliefSim-FT approach performs the best, showing that belief adapter + lightweight sus- ceptibility head captures more transferable signals aligned with out-of-dom...

  3. [5]

    Inderjeet Nair and Lu Wang

    Psychological factors contributing to the cre- ation and dissemination of fake news among social media users: a systematic review.BMC psychology, 12(1):673. Inderjeet Nair and Lu Wang. 2025. Do language mod- els think consistently? a study of value preferences across varying response lengths.arXiv preprint arXiv:2506.02481. Keiichi Namikoshi, Alex Filipow...

  4. [6]

    Jan Nehring, Aleksandra Gabryszak, Pascal Jür- gens, Aljoscha Burchardt, Stefan Schaffer, Matthias Spielkamp, and Birgit Stark

    Using llms to model the beliefs and pref- erences of targeted populations.arXiv preprint arXiv:2403.20252. Jan Nehring, Aleksandra Gabryszak, Pascal Jür- gens, Aljoscha Burchardt, Stefan Schaffer, Matthias Spielkamp, and Birgit Stark. 2024. Large language models are echo chambers. InProceedings of the 2024 Joint International Conference on Computa- tional...

  5. [7]

    Nikhil Sharma, Q Vera Liao, and Ziang Xiao

    Whose opinions do language models reflect? Preprint, arXiv:2303.17548. Nikhil Sharma, Q Vera Liao, and Ziang Xiao. 2024. Generative echo chamber? effect of llm-powered search systems on diverse information seeking. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, pages 1–17. Mubashir Sultan, Alan N Tump, Nina Ehmann, Philip...

  6. [8]

    A person with acompleted education is less likely to believe in such a sensational and unsubstanti- ated claim

    Sotopia: Interactive evaluation for social intelligence in language agents.arXiv preprint arXiv:2310.11667. A Taxonomy Dimensions and Imputed Beliefs Beliefs are an important tool for modeling demo- graphic simulations in the context of misinforma- tion. Psychological work shows that several be- lief dimensions, such as conspiracy beliefs, polit- ical ide...

  7. [2016]

    Rakoen Maertens, Friedrich M Götz, Hudson F Golino, Jon Roozenbeek, Claudia R Schneider, Yara Kyrychenko, John R Kerr, Stefan Stieger, William P McClanahan, Karly Drabot, et al

    Social media and credibility indicators: The ef- fect of influence cues.Computers in human behavior, 63:264–271. Rakoen Maertens, Friedrich M Götz, Hudson F Golino, Jon Roozenbeek, Claudia R Schneider, Yara Kyrychenko, John R Kerr, Stefan Stieger, William P McClanahan, Karly Drabot, et al. 2024. The misin- formation susceptibility test (mist): A psychomet...

  8. [2019]

    Nicolas M Anspach and Taylor N Carlson

    Trends in the diffusion of misinforma- tion on social media.Research & politics, 6(2):2053168019848554. Nicolas M Anspach and Taylor N Carlson. 2024. Not who you think? exposure and vulnerability to misin- formation.New Media & Society, 26(8):4847–4866. Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate

Show all 10 references
  1. [2023]

    Dane Bertram

    Out of one, many: Using language mod- els to simulate human samples.Political Analysis, 31(3):337–351. Dane Bertram. 2007. Likert scales.Retrieved Novem- ber, 2(10):1–10. Angana Borah, Marwa Houalla, and Rada Mihalcea. 2025a. Mind the (belief) gap: Group identity in the world ...

  2. [2024]

    Rada Mihalcea, Oana Ignat, Longju Bai, Angana Borah, Luis Chiruzzo, Zhijing Jin, Claude Kwizera, Joan Nwatu, Soujanya Poria, and Thamar Solorio

    Misinformation exploits outrage to spread on- line.Science, 386(6725):991–996. Rada Mihalcea, Oana Ignat, Longju Bai, Angana Borah, Luis Chiruzzo, Zhijing Jin, Claude Kwizera, Joan Nwatu, Soujanya Poria, and Thamar Solorio. 2025. Why ai is weird and shouldn’t be this way: Towa...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.