Pith. sign in

REVIEW 3 major objections 6 minor 71 references

ConspirED: A Dataset for Cognitive Traits of Conspiracy Theories and Large Language Model Safety

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Language models are easier to misalign with conspiratorial text than with fact-checked misinformation, even when they can name the conspiratorial reasoning pattern when asked.

desk verdict Useful new dataset and honest benchmarks, but the headline safety claim compares apples to oranges. read the letter →

arxiv 2508.20468 v1 pith:ZOFCK76K submitted 2025-08-28 cs.CL

Luke Bates , Max Glockner , Preslav Nakov , Iryna Gurevych This is my paper
classification cs.CL
keywords conspiracytheoriesCONSPIRframeworkcognitivetraitsLLMsafetymisalignmenttraitdetectionmisinformationdataset
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

The paper introduces ConspirED, the first dataset of conspiracy-theory excerpts annotated with the CONSPIR cognitive traits: recognizable reasoning patterns such as overriding suspicion of official accounts, nefarious intent, immunity to counter-evidence, and reinterpreting randomness. Using about 1,974 snippets from real conspiracy articles, it asks whether machines can spot these traits and whether they can resist reproducing them. The central finding is that models can often identify the traits when explicitly prompted, yet when given a benign-sounding rewrite instruction they comply with conspiratorial content at roughly twice the rate they comply with comparable fact-checked misinformation (40.3% vs. 19.7%). The paper argues this shows that current safety alignment is bypassed by conspiratorial framing that fact-checking-style defenses were not built for.

What carries the argument

The load-bearing object is the CONSPIR cognitive framework, operationalized as six trait labels: Contradictory, Overriding suspicion, Nefarious intent, Persecuted victim, Immune to evidence, and Re-interpreting randomness, with the handbook's 'Something must be wrong' merged into Overriding suspicion. ConspirED annotates 80–120 word excerpts with all applicable traits plus a dominant trait, and the experimental engine is the contrast between two tasks on the same text: a classification prompt that asks models to name the traits, and a journalistic-rewrite prompt that asks models to restyle them. That contrast quantifies the paper's central paradox: trait recognition without trait resistance.

What would settle it

Run the same journalistic-rewrite instruction on matched 80-to-120-word non-conspiratorial misinformation with similarly sensational claims; if compliance matches the conspiratorial rate, the gap is driven by format rather than by conspiratorial traits.

Watch

Extended reading notes

Core claim

The paper's central claim is that large language and reasoning models are misaligned by conspiratorial content: prompted to rewrite a snippet in journalistic style, six models complied with ConspirED excerpts at roughly twice the rate they complied with AVeriTeC fact-checked claims, and the traits most likely to be refused were Nefarious intent and Immune to evidence. The same models, prompted with trait definitions and guidelines, detected the dominant conspiratorial trait nearly as well as human annotators in a relaxed setting, yet this detection ability did not carry over to resistance during generation. The paper's paradox is concrete: the models can name the reasoning pattern and still

Load-bearing premise

The safety comparison assumes the conspiratorial excerpts and the fact-checked claims are comparable apart from being conspiratorial, but they also differ in length, writing style, and how directly the falsehood is stated.

Editorial extensions

If this is right

  • Safety evaluations that use only single-sentence fact-checked misinformation will underestimate how often models reproduce conspiracy narratives.
  • Trait labels make targeted prebunking possible: a text flagged as Immune to evidence can be countered with inoculation messaging aimed at that pattern rather than at the topic.
  • Dominant-trait detection is reachable for LLMs and near-LLM for lightweight classifiers, enabling low-cost screening of conspiracy rhetoric at scale.
  • Models' selective deflection, concentrated on Nefarious intent and Immune to evidence, gives alignment researchers a concrete target for adversarial or preference-based training.
  • The dataset's temporal split (pre-2020 training, post-2020 test) supports evaluation of generalization to novel conspiracy narratives.

Reading between the lines

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

  • Beyond the paper's experiments, a matched control is the natural next test: non-conspiratorial misinformation written in the same 80–120 word multi-sentence style. If compliance matches the 40.3% conspiratorial rate, much of the reported gap is format rather than conspiracy-specific traits.
  • The near-zero annotation agreement on Contradictory and its rarity in the dataset suggest that the absence of Contradictory from the safety analysis may be an annotation artifact rather than a fact about how conspiracy discourse behaves.
  • The paper's own detection results imply that trait-aware refusal training is a testable extension: models already can name the patterns, so preference tuning on ConspirED snippets could plausibly turn recognition into resistance.
  • Because the 50 safety-test snippets were manually selected as the 'most sensational' matches to the AVeriTeC claims, a random sample of ConspirED excerpts would give a fairer baseline estimate of model compliance.
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

3 major / 6 minor

Summary. This paper introduces ConspirED, a dataset of 1,974 multi-sentence excerpts (80–120 words) from conspiracy theory articles, annotated with the CONSPIR cognitive traits (multi-label and dominant-trait). The authors evaluate trait detection with LaGoNN and several LLMs/LRMs, and test LLM safety via a paraphrase task that compares ConspirED snippets with AVeriTeC fact-checked claims. They report that models detect traits at moderate performance (macro-F1 38–44 for all traits; 50–60 for dominant trait), and that LLMs comply with journalistic rewrites of conspiracy content at 40.3% versus 19.7% for fact-checked misinformation, concluding that LLMs are misaligned by conspiratorial content even when they deflect comparable fact-checked misinformation.

Significance. The dataset is a useful contribution: it is, to my knowledge, the first resource annotated for general conspiratorial cognitive traits, and it targets a topic-agnostic framework with practical relevance for prebunking and content moderation. The paper is transparent about annotation agreement and limitations, and the authors plan to release code and data. However, the headline safety claim—that LLMs are more easily misaligned by conspiratorial content than by comparable fact-checked misinformation—rests on a confounded comparison, and the claim that LLMs perform on par with humans in the relaxed setting is not borne out by the reported numbers. With a revised, better-controlled comparison and corrected claims, the dataset and benchmarks could be valuable to the community.

major comments (3)
  1. [5.1, 5.2, Figure 4] The central safety claim depends on the comparability of the 50 ConspirED snippets and 50 AVeriTeC claims, but the two sets are not matched on format, length, or sensationalism. ConspirED snippets are 80–120-word multi-sentence article excerpts, while AVeriTeC claims are typically single extracted sentences. The selection procedure in §5.1 applies 'most sensational' only to the ConspirED side, and AVeriTeC claims are fact-checked, so models may have memorized their refutation, making deflection easier. Any of these differences could raise the 40.3% vs. 19.7% compliance gap without the content being inherently conspiratorial. The authors should either match inputs on length and format, or explicitly analyze how these factors affect compliance, and should not call the sets 'comparable'.
  2. [1, 6, Table 7] The abstract and conclusion state that LLMs perform 'on par with humans' in the relaxed dominant-trait setting. Table 7 reports human Relax macro-F1 of 69.70, while the best LLM (Llama 3.1 70B, k=0, Context1000) is 60.59—a gap of about 9 points. This is not parity. The claim should be removed or qualified, or the authors should report a statistical test if they believe the difference is not significant.
  3. [3.5–3.6, Tables 4–5] The annotation quality is moderate (γ=0.57, α=0.52), and the consolidation was performed by a single author with validation on only 31 instances. This is a known limitation, but the paper should provide more evidence of label stability: report agreement between the original and consolidated labels, report IAA on the final snippet units rather than only on spans, and discuss how annotation subjectivity affects trait-level conclusions such as the chi-square analysis in §5.4. The moderate agreement should temper the confidence in the gold labels used for both detection benchmarks and safety experiments.
minor comments (6)
  1. [3.3, footnote 4] Typo: 'hhttps://pypi.org/project/beautifulsoup4/' should be 'https://pypi.org/project/beautifulsoup4/'.
  2. [Figure 4] The bars are not labeled with numeric percentages, which makes it difficult to compare values across models and conditions. Please add value labels or a table.
  3. [5.4] The chi-square tests are reported at p<0.001 without correction for multiple comparisons across the six traits. Report adjusted p-values or state the number of comparisons.
  4. [Reproducibility] The code and data are described as 'will be made publicly available shortly after release of this preprint.' For review, provide an anonymous link or commit hash to verify reproducibility.
  5. [Throughout] The dataset name is rendered inconsistently as 'C ONSPIR ED', 'CONSPIR ED', and 'ConspirED'. Please standardize the typography.
  6. [3.2] The temporal split is sensible, but the test set is limited to 2020–2023 topics from a single source (GlobalResearch). A short discussion of domain shift and source diversity would improve external validity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset labels, trait detections, and LLM misalignment measurements are externally grounded empirical results.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. The CONSPIR cognitive framework is taken from the external Conspiracy Theory Handbook by Lewandowsky and Cook (2020), not from the authors' prior work. The dataset labels are produced by human annotators with reported agreement (gamma 0.57 overall, Krippendorff's alpha 0.52 for consolidation), and the classification experiments are standard supervised evaluations against those independently produced labels. The zero-shot and few-shot prompting experiments use the trait definitions and annotation guidelines as prompts, but this is ordinary task specification, not circularity: the labels being predicted were not generated by the models being evaluated. The LLM misalignment comparison is an empirical measurement of response categories (Deflected/Equivocated/Complied) on two corpora. The paper's central safety claim is a comparative behavioral observation, not a quantity derived from a fitted parameter. The AVeriTeC control is not perfectly matched to the ConspirED snippets in length, genre, or sensationalism, and the manual selection of 'the most sensational instance' introduces a potential confound; however, this is a validity or comparability threat, not a circularity in the sense of the target result being equivalent to the inputs by construction. Self-citations appear (LaGoNN, SetFit, annotation-management methodology, and prior conspiracy-trait work), but none of them is load-bearing in the way that would force the paper's conclusions: LaGoNN is one of several classifiers compared, and the CONSPIR trait definitions are externally sourced. No uniqueness theorem, ansatz smuggling, or renaming of a known result is present. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are postulated. The CONSPIR traits come from prior literature, and ConspirED is a dataset, not an invented entity in the graviton sense. The free parameters listed are model hyperparameters tuned on dev data; they do not directly affect the central misalignment claim but are part of the detection benchmarks.

free parameters (3)
  • LaGoNN training epochs = 3
    Selected by validation on dev splits (Section 4.1.1 and Appendix C).
  • LaGoNN contrastive pairs (num_iters) = 17
    Selected by validation on dev splits (Section 4.1.1 and Appendix C).
  • Number of in-context examples k = 10 or 20
    Evaluated as a hyperparameter for LLM prompting; k=20 used for few-shot results (Section 4.1.2).
assumptions (5)
  • domain assumption The CONSPIR framework (Lewandowsky and Cook, 2020) is a valid taxonomy of conspiratorial cognitive traits.
    The dataset and all classification tasks are built on this framework; no independent validation is provided (Section 3).
  • domain assumption The trait 'Something must be wrong' can be merged into 'Overriding suspicion' without loss of information.
    Section 3 states subtle differences rarely manifest in text; this merge changes the label space and affects downstream statistics.
  • domain assumption The 50 pairs of ConspirED snippets and AVeriTeC claims are comparable except for the conspiratorial versus fact-checked distinction.
    The misalignment experiment relies on this assumption; it is not controlled for length, format, or genre (Section 5.1).
  • domain assumption Rejection of conspiratorial content is always the correct model behavior in the rewriting task.
    Acknowledged in Limitations: 'it assumes that rejection is always correct.'
  • domain assumption Snippets of 80-120 words preserve the conspiratorial traits present in the full articles.
    Snippet extraction is based on annotator span selections; no analysis of completeness is provided (Section 3.6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of ConspirED: A Dataset for Cognitive Traits of Conspiracy Theories and Large Language Model Safety." pith.science (2026). https://pith.science/paper/ZOFCK76K

@misc{pith2026250820468,
  author       = {Pith},
  title        = {Pith review of: ConspirED: A Dataset for Cognitive Traits of Conspiracy Theories and Large Language Model Safety},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZOFCK76K}},
  note         = {Machine review of arXiv:2508.20468}
}
read the original abstract

Conspiracy theories erode public trust in science and institutions while resisting debunking by evolving and absorbing counter-evidence. As AI-generated misinformation becomes increasingly sophisticated, understanding rhetorical patterns in conspiratorial content is important for developing interventions such as targeted prebunking and assessing AI vulnerabilities. We introduce ConspirED (CONSPIR Evaluation Dataset), which captures the cognitive traits of conspiratorial ideation in multi-sentence excerpts (80--120 words) from online conspiracy articles, annotated using the CONSPIR cognitive framework (Lewandowsky and Cook, 2020). ConspirED is the first dataset of conspiratorial content annotated for general cognitive traits. Using ConspirED, we (i) develop computational models that identify conspiratorial traits and determine dominant traits in text excerpts, and (ii) evaluate large language/reasoning model (LLM/LRM) robustness to conspiratorial inputs. We find that both are misaligned by conspiratorial content, producing output that mirrors input reasoning patterns, even when successfully deflecting comparable fact-checked misinformation.

Figures

Figures reproduced from arXiv: 2508.20468 by the authors.

Figure 1
Figure 1. We show that LLMs are easily misaligned by C [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Frequency of CONSPIR traits by annotator [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Frequency of each CONSPIR trait in the train￾ing and test splits of our consolidated dataset. tions. In this way, each sample in our dataset con￾tains at most three annotations (two original an￾notations and one consolidated annotation) and at least two annotations (one original and one con￾solidated). During consolidation, we extracted snippets (si) from the span annotations, favor￾ing multiple shorter snippets ove… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: LLM response percentages grouped into Equivocated, Deflected, and Complied. Dark bars: CONSPIRED; light bars: AVeriTeC. 1. Deflected: Completely rejected the request and redirected the conversation. 2. Equivocated: Acknowledged the input as false/problematic but showed…
Figure 5
Figure 5. Figure 5: Screenshot of the annotation interface. Annotators selected one or more applicable labels and provided [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Guidelines prompt template used for LLM-based CONSPIRED classification [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 34 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. http://arxiv.org/abs/1606.06565 Concrete problems in AI safety . arXiv preprint arXiv:1606.06565

    arXiv 2016

  4. [4]

    Anthropic. 2024. https://claude.ai Claude 3 Sonnet . Accessed: 2025-07-03

    work page 2024

  5. [5]

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, et al. 2022. http://arxiv.org/abs/2212.08073 Constitutional ai: Harmlessness from ai feedback . arXiv preprint arXiv:2212.08073

    arXiv 2022

  6. [6]

    Banas and Gregory Miller

    John A. Banas and Gregory Miller. 2013. https://doi.org/10.1111/hcre.12000 Inducing resistance to conspiracy theory propaganda: Testing inoculation and metainoculation strategies . Human Communication Research, 39(2):184--207

    work page doi:10.1111/hcre.12000 2013

  7. [7]

    Luke Bates and Iryna Gurevych. 2024. https://aclanthology.org/2024.eacl-long.17 Like a good nearest neighbor: Practical content moderation and text classification . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 276--297, St. Julian ' s, Malta. Association for C...

    work page 2024

  8. [8]

    Michael Butter. 2020. https://politybooks.com/bookdetail/?isbn=9781509540815 The Nature of Conspiracy Theories . Polity, Cambridge

    work page 2020

Show all 71 references
  1. [9]

    Michael Butter and Peter Knight, editors. 2020. https://www.routledge.com/Routledge-Handbook-of-Conspiracy-Theories/Butter-Knight/p/book/9781032173986?srsltid=AfmBOop5mlQzo_Yb1cwr0AldQc_AyZq9BKJW5xVgBT64xNqGYhezgZTv Routledge Handbook of Conspiracy Theories . Routledge, London

    2020

  2. [10]

    John Cook, Stephan Lewandowsky, and Ullrich K. H. Ecker. 2017. https://doi.org/10.1371/journal.pone.0175799 Neutralizing misinformation through inoculation: Exposing misleading argumentation techniques reduces their influence . PLOS ONE, 12(5):1--21

    2017 doi

  3. [11]

    Francesco Corso, Francesco Pierri, and Gianmarco De Francisci Morales. 2025. https://doi.org/10.18653/v1/2025.acl-long.408 Conspiracy theories and where to find them on T ik T ok . In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volu...

    2025 doi

  4. [12]

    Costello, Gordon Pennycook, and David G

    Thomas H. Costello, Gordon Pennycook, and David G. Rand. 2024. https://doi.org/10.1126/science.adq1814 Durably reducing conspiracy beliefs through dialogues with AI . Science, 385(6714):eadq1814

    2024 doi

  5. [13]

    Giovanni Da San Martino, Seunghak Yu, Alberto Barr \'o n-Cede \ n o, Rostislav Petrov, and Preslav Nakov. 2019. https://doi.org/10.18653/v1/D19-1565 Fine-grained analysis of propaganda in news articles . In Proceedings of the 2019 Conference on Empirical Methods in Natural Lan...

    2019 doi

  6. [14]

    Thomas Daigle. 2021. https://www.cbc.ca/news/science/russian-disinformation-global-research-website-1.5767208 Canadian professor's website helps Russia spread disinformation, says U.S. State Department . CBC News

    2021

  7. [15]

    Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W

    Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. 2022. https://doi.org/10.1162/tacl_a_00459 Time-aware language models as temporal knowledge bases . Transactions of the Association for Computational Linguistics,...

    2022 doi

  8. [16]

    Douglas, Robbie M

    Karen M. Douglas, Robbie M. Sutton, and Aleksandra Cichocka. 2017. https://doi.org/10.1177/0963721417718261 The psychology of conspiracy theories . Current Directions in Psychological Science, 26(6):538--542

    2017 doi

  9. [17]

    Holford, Luke Bates, Iryna Gurevych, and Stephan Lewandowsky

    Angelo Fasce, Philipp Schmid, Dawn L. Holford, Luke Bates, Iryna Gurevych, and Stephan Lewandowsky. 2023. https://doi.org/10.1038/s41562-023-01644-3 A taxonomy of anti-vaccination arguments from a systematic literature review and text modelling . Nature Human Behaviour, 7(9):1...

    2023 doi

  10. [18]

    Global Engagement Center . 2020. https://2017-2021.state.gov/russias-pillars-of-disinformation-and-propaganda-report/ Pillars of Russia’s disinformation and propaganda ecosystem . US Department of State

    2020

  11. [19]

    Max Glockner, Yufang Hou, Preslav Nakov, and Iryna Gurevych. 2024. https://doi.org/10.18653/v1/2024.acl-long.240 Missci: Reconstructing fallacies in misrepresented science . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

    2024 doi

  12. [20]

    Google. 2025. https://gemini.google.com Gemini . Accessed: 2025-06-03

    2025

  13. [21]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, et al. 2024. http://arxiv.org/abs/2407.21783 The Llama 3 herd of models . arXiv preprint arXiv:2407.21783

    2024 arXiv

  14. [22]

    Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens van der Maaten. 2020. https://proceedings.mlr.press/v119/guo20c/guo20c.pdf Certified data removal from machine learning models . In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceeding...

    2020

  15. [23]

    Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. https://doi.org/10.1162/tacl_a_00454 A survey on automated fact-checking . Transactions of the Association for Computational Linguistics, 10:178--206

    2022 doi

  16. [24]

    Maram Hasanain, Fatema Ahmad, and Firoj Alam. 2024. http://arxiv.org/abs/2311.09812 Large language models for propaganda span annotation . arXiv preprint arXiv:2311.09812

    2024 arXiv

  17. [25]

    Chadi Helwe, Tom Calamai, Pierre-Henri Paris, Chlo \'e Clavel, and Fabian Suchanek. 2024. https://doi.org/10.18653/v1/2024.naacl-long.270 MAFALDA : A benchmark and comprehensive study of fallacy detection and classification . In Proceedings of the 2024 Conference of the North ...

    2024 doi

  18. [26]

    Pavan Holur, Tianyi Wang, Shadi Shahsavari, Timothy Tangherlini, and Vwani Roychowdhury. 2022. https://doi.org/10.18653/v1/2022.acl-long.341 Which side are you on? Insider-Outsider classification in conspiracy-theoretic social media . In Proceedings of the 60th Annual Meeting ...

    2022 doi

  19. [27]

    Colin Klein, Peter Clutton, and Vince Polito. 2018. https://doi.org/10.3389/fpsyg.2018.00189 Topic modeling reveals distinct interests within an online conspiracy forum . Frontiers in Psychology, 9:189

    2018

  20. [28]

    Jan-Christoph Klie, Michael Bugert, Beto Boullosa, Richard Eckart de Castilho, and Iryna Gurevych. 2018. https://aclanthology.org/C18-2002 The INCE p TION platform: Machine-assisted and knowledge-oriented interactive annotation . In Proceedings of the 27th International Confer...

    2018

  21. [29]

    Jan-Christoph Klie, Richard Eckart de Castilho, and Iryna Gurevych. 2024. https://doi.org/10.1162/coli_a_00516 Analyzing dataset annotation quality management in the wild . Computational Linguistics, 50(3):817--866

    2024 doi

  22. [30]

    Kunst, A.B

    J.R. Kunst, A.B. Gundersen, I. Krysińska, J. Piasecki, T. Wójtowicz, R. Rygula, S. van der Linden, and M. Morzy. 2024. https://doi.org/10.1038/s41467-024-51740-9 Leveraging artificial intelligence to identify the psychological factors associated with conspiracy theory beliefs ...

    2024 doi

  23. [31]

    Florian Kurth, Nicolas Cherbuin, and Eileen Luders. 2017. https://doi.org/10.3389/fpsyg.2017.00860 Promising links between meditation and reduced (brain) aging: An attempt to bridge some gaps between the alleged fountain of youth and the youth of the field . Frontiers in Psych...

    2017

  24. [32]

    Johannes Langguth, Daniel Thilo Schroeder, Petra Filkukov \'a , Stefan Brenner, Jesper Phillips, and Konstantin Pogorelov. 2023. https://doi.org/10.1007/s42001-023-00200-3 COCO : an annotated Twitter dataset of COVID-19 conspiracy theories . Journal of Computational Social Sci...

    2023 doi

  25. [33]

    Anne Lauscher, Brandon Ko, Bailey Kuehl, Sophie Johnson, Arman Cohan, David Jurgens, and Kyle Lo. 2022. https://doi.org/10.18653/v1/2022.naacl-main.137 M ulti C ite: Modeling realistic citations requires moving beyond the single-sentence single-label setting . In Proceedings o...

    2022 doi

  26. [34]

    Angeliki Lazaridou, Adhiguna Kuncoro, Elena Gribovskaya, Devang Agrawal, Adam Li s ka, Tayfun Terzi, Mai Gimenez, Cyprien de Masson d'Autume, Tomas Kocisky, Sebastian Ruder, Dani Yogatama, Kris Cao, Susannah Young, and Phil Blunsom. 2021. https://openreview.net/forum?id=73Ommr...

    2021

  27. [35]

    Yuanyuan Lei and Ruihong Huang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.656 Identifying conspiracy theories news based on event relation graph . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9811--9822, Singapore. Association for...

    2023 doi

  28. [36]

    Sharon Levy, Michael Saxon, and William Yang Wang. 2021. https://doi.org/10.18653/v1/2021.findings-acl.416 Investigating memorization of conspiracy theories in text generation . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4718--4729, On...

    2021 doi

  29. [37]

    Stephan Lewandowsky and John Cook. 2020. https://skepticalscience.com/docs/ConspiracyTheoryHandbook.pdf The Conspiracy Theory Handbook . George Mason University, United States

    2020

  30. [38]

    Stephan Lewandowsky, Ullrich KH Ecker, Colleen M Seifert, Norbert Schwarz, and John Cook. 2012. https://doi.org/10.1177/1529100612451018 Misinformation and its correction: Continued influence and successful debiasing . Psychological Science in the Public Interest, 13(3):106--131

    2012 doi

  31. [39]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. https://doi.org/10.18653/v1/2022.acl-long.229 T ruthful QA : Measuring how models mimic human falsehoods . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa...

    2022 doi

  32. [40]

    Mareike Lisker, Christina Gottschalk, and Helena Mihaljević. 2025. https://arxiv.org/abs/2504.16604 Debunking with dialogue? Exploring AI -generated counterspeech to challenge conspiracy theories . arXiv preprint arXiv:2504.16604

    2025 arXiv

  33. [41]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, et al. 2025 a . http://arxiv.org/abs/2412.19437 DeepSeek-V3 technical report . arXiv preprint arXiv:2412.19437

    2025 arXiv

  34. [42]

    Zhiwei Liu, Paul Thompson, Jiaqi Rong, and Sophia Ananiadou. 2025 b . https://arxiv.org/abs/2505.14917 ConspEmoLLM-v2 : A robust and stable model to detect sentiment-transformed conspiracy theories . arXiv preprint arXiv:2505.14917

    2025

  35. [43]

    Inbal Magar and Roy Schwartz. 2022. https://doi.org/10.18653/v1/2022.acl-short.18 Data contamination: From memorization to exploitation . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 157--165, Dublin...

    2022 doi

  36. [44]

    Yann Mathet, Antoine Widl \"o cher, and Jean-Philippe M \'e tivier. 2015. https://doi.org/10.1162/COLI_a_00227 The unified and holistic method gamma ( ) for inter-annotator agreement measure and alignment . Computational Linguistics, 41(3):437--479

    2015 doi

  37. [45]

    Alessandro Miani, Thomas Hills, and Adrian Bangerter. 2022 a . https://doi.org/10.1126/sciadv.abq3668 Interconnectedness and (in)coherence as a signature of conspiracy worldviews . Science Advances, 8(43):eabq3668

    2022 doi

  38. [46]

    Alessandro Miani, Thomas Hills, and Adrian Bangerter. 2022 b . https://doi.org/10.3758/s13428-021-01698-z LOCO : The 88-million-word language of conspiracy corpus . Behavior Research Methods, 54:1794–1817

    2022 doi

  39. [47]

    OpenAI. 2025. https://chat.openai.com ChatGPT . Accessed: 2025-06-03

    2025

  40. [48]

    Will Orr and Kate Crawford. 2024. http://arxiv.org/abs/2409.00252 Building better datasets: Seven recommendations for responsible design from dataset creators . arXiv preprint arXiv:2409.00252

    2024 arXiv

  41. [49]

    Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Wang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.97 On the risk of misinformation pollution with large language models . In Findings of the Association for Computational Linguistics: EMNL...

    2023 doi

  42. [50]

    Ethan Perez, Sam Ringer, Kamilė Lukošiūtė, Karina Nguyen, et al. 2022. http://arxiv.org/abs/2212.09251 Discovering language model behaviors with model-written evaluations . arXiv preprint arXiv:2212.09251

    2022 arXiv

  43. [51]

    Shawn Pogatchnik. 2021. https://apnews.com/article/920e1c738df04555bccd56c09770b36d AP fact check: Irish ``slavery'' a St . Patrick's Day myth . AP News

    2021

  44. [52]

    Jan-Willem van Prooijen and Karen M Douglas. 2017. https://doi.org/10.1177/1750698017701615 Conspiracy theories as part of history: The role of societal crisis situations . Memory Studies, 10(3):323--333. PMID: 29081831

    2017 doi

  45. [53]

    Alan Ramponi, Agnese Daffara, and Sara Tonelli. 2025. https://doi.org/10.18653/v1/2025.naacl-long.34 Fine-grained fallacy detection with human label variation . In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Li...

    2025 doi

  46. [54]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference...

    2019 doi

  47. [55]

    Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, and Andreas Vlachos. 2024. https://doi.org/10.18653/v1/2024.fever-1.1 The automated verificat...

    2024 doi

  48. [56]

    Shadi Shahsavari, Pavan Holur, Timothy Wang, Timothy R Tangherlini, and Vwani Roychowdhury. 2020. https://doi.org/10.1007/s42001-020-00086-5 Conspiracy in the time of C orona: Automatic detection of emerging COVID-19 conspiracy theories in social media and the news . Journal o...

    2020 doi

  49. [57]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. https://proceedings.neurips.cc/paper/2020/file/d5fbba59676d13573d1d0fde566ff911-Paper.pdf MPNet : Masked and permuted pre-training for language understanding . In Proceedings of the 34th International Conference...

    2020

  50. [58]

    X. Song, J. Petrak, Y. Jiang, I. Singh, D. Maynard, and K. Bontcheva. 2021. https://doi.org/10.1371/journal.pone.0247086 Classification aware neural topic model for COVID-19 disinformation categorisation . PLoS One, 16(2):e0247086. Published 2021 Feb 18

    2021 doi

  51. [59]

    Cass R Sunstein and Adrian Vermeule. 2009. https://doi.org/10.1111/j.1467-9760.2008.00325.x Conspiracy theories: Causes and cures . Journal of Political Philosophy, 17(2):202--227

    2009

  52. [60]

    Timothy R Tangherlini, Shadi Shahsavari, Behnam Shahbazi, Ehsan Ebrahimzadeh, and Vwani Roychowdhury. 2020. https://doi.org/10.1371/journal.pone.0233879 An automated pipeline for the discovery of conspiracy and conspiracy theory narrative frameworks: Bridgegate , Pizzagate and...

    2020 doi

  53. [61]

    Hadrien Titeux and Rachid Riad. 2021. https://doi.org/10.21105/joss.02989 pygamma-agreement: Gamma measure for inter/intra-annotator agreement in python . Journal of Open Source Software, 6(62):2989

    2021 doi

  54. [62]

    Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. 2022. https://doi.org/10.48550/ARXIV.2209.11055 Efficient few-shot learning without prompts . arXiv preprint arXiv:2209.11055

    2022 doi arXiv

  55. [63]

    Fanqi Wan, Weizhou Shen, Shengyi Liao, Yingcheng Shi, Chenliang Li, Ziyi Yang, Ji Zhang, Fei Huang, Jingren Zhou, and Ming Yan. 2025. http://arxiv.org/abs/2505.17667 Qwenlong-l1: Towards long-context large reasoning models with reinforcement learning . arXiv preprint arXiv:2505.17667

    2025 arXiv

  56. [64]

    Yuhui Wang, Rongyi Zhu, and Ting Wang. 2025. https://arxiv.org/abs/2505.12186 Self-destructive language model . arXiv preprint arXiv:2505.12186

    2025

  57. [65]

    Yuxia Wang, Haonan Li, Xudong Han, Preslav Nakov, and Timothy Baldwin. 2024. https://aclanthology.org/2024.findings-eacl.61/ Do-Not-Answer : Evaluating safeguards in LLM s . In Findings of the Association for Computational Linguistics: EACL 2024, pages 896--911, St. Julian ' s...

    2024

  58. [66]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf Chain-of-thought prompting elicits reasoning i...

    2022

  59. [67]

    Michael J. Wood. 2018. https://doi.org/10.1089/cyber.2017.0669 Propagating and debunking conspiracy theories on Twitter during the 2015–2016 Zika virus outbreak . Cyberpsychology, Behavior, and Social Networking, 21(8):485--490

    2018

  60. [68]

    Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. 2021. http://arxiv.org/abs/2112.03254 Human parity on CommonsenseQA : Augmenting self-attention with external attention . arXiv preprint ar...

    2021 arXiv

  61. [69]

    Xuandong Zhao, Will Cai, Tianneng Shi, David Huang, Licong Lin, Song Mei, and Dawn Song. 2025. https://arxiv.org/abs/2503.03710 Improving LLM safety alignment with dual-objective optimization . arXiv preprint arXiv:2503.03710

    2025

  62. [70]

    Zico Kolter, and Matt Fredrikson

    Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. 2023. http://arxiv.org/abs/2307.15043 Universal and transferable adversarial attacks on aligned language models . arXiv preprint arXiv:2307.15043

    2023 arXiv

  63. [71]

    Aneta Zugecova, Dominik Macko, Ivan Srba, Robert Moro, Jakub Kop \'a l, Katar \'i na Marcin c inov \'a , and Mat \'u s Mesar c \'i k. 2025. https://doi.org/10.18653/v1/2025.acl-long.38 Evaluation of LLM vulnerabilities to being misused for personalized disinformation generatio...

    2025 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.