Elamur Research⌕
☰
home / papers / 2025 / LLMs Struggle to Reject False Presuppositions when Misinformation Stakes are High

LLMs Struggle to Reject False Presuppositions when Misinformation Stakes are High

Summary

This paper asks whether large language models catch and correct false information that is smuggled into a question as a presupposition, that is, background content treated as already given. The authors focus on politics, where such hidden misinformation can do real harm. They built a new German dataset of prompts that wrongly place a known politician at a rival party's conference. Using linguistic presupposition theory, they varied four factors: the trigger type, the embedding context (such as questions, negation or modals), how plausible the scenario is, and the political distance between the parties. They tested GPT-4-o, LLama-3-8B and Mistral-7B-v03 and labelled each answer as an accommodation, an imprecise response or a rejection. Every model often accepted the false premise, even though it was blatant. GPT did best but still missed many cases. The factors that matter to humans also affected the models, though in uneven and inconsistent ways. The authors argue that presupposition analysis is a useful way to audit LLMs for spreading political misinformation.

Key points

  • The dataset is in German and covers 7 presupposition trigger types with 23 individual triggers, including factive verbs, change-of-state verbs, possessives, quantifiers and temporal clauses. The study's results are available on Zenodo as a benchmark called FLEX, short for False Presupposition Linguistic Evaluation eXperiment.
  • Scenario probability had the strongest effect: low-probability scenarios were rejected more often. For GPT's accommodation model, χ² was 206.28 for probability and 164.66 for political distance.
  • Political distance mattered most for GPT. Embedding context strongly affected LLaMa but had no significant effect on GPT.
  • Trigger type rarely showed up as a main effect. It appeared only in many scattered, unsystematic interaction effects.
  • The authors raise practical concerns, noting that 35.6% of Generation Z reportedly use AI tools for political information. They also observe that GPT's long, vague answers can hide failures to reject the false premise.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages · Natural language processing · AI & machine learning

Cited by

10 papers of our library cite it (of the 140673 whose bibliographies we read); 1 of them have a page here, 9 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.22354 · arXiv 2505.22354

@misc{sieker2025llms,
  title = {{LLMs Struggle to Reject False Presuppositions when Misinformation Stakes are High}},
  author = {Sieker, Judith and Lachenmaier, Clara and Zarriess, Sina},
  year = {2025},
  doi = {10.48550/arxiv.2505.22354},
  eprint = {2505.22354},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.22354}
}

References

30 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., . . . others (2023). Gpt-4 technical report. arXiv preprint arXiv:2303.08774. ↗ our card
  2. Allyn Smith, E., & Hall, K. C. (2014). The relationship between projection and embedding environment. Proceedings of the 48th Meeting of the Chicago Linguistics Society. find
  3. Bang, Y., Chen, D., Lee, N., & Fung, P. (2024, August). Measuring political bias in large language models: What is said and how it is said. In L.-W. Ku, A. Martins, & V. Srikumar (Eds.), Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) (pp. 11142–11159). Bangkok, Thailand: Association for Computational Linguistics. Retrieved from https://aclanthology.org/2024.acl-long.600 ↗
  4. Beaver, D. I., Geurts, B., & Denlinger, K. (2024). Presupposition. In E. N. Zalta & U. Nodelman (Eds.), The Stanford encyclopedia of philosophy (Fall 2024 ed.). Metaphysics Research Lab, Stanford University. https://plato.stanford.edu/archives/ fall2024/entries/presupposition/. ↗
  5. Bender, E., & Lascarides, A. (2019, 11). Linguistic fundamentals for natural language processing ii: 100 essentials from semantics and pragmatics. Synthesis Lectures on Human Language Technologies, 12, 1-268. doi: 10.2200/ S00935ED1V02Y201907HLT043 ↗
  6. Chierchia, G., & Mcconnell-Ginet, S. (1990). Meaning and grammar: An introduction to semantics.. Retrieved from https://api.semanticscholar.org/CorpusID: 62731986 ↗
  7. Contreras Kallens, P., Kristensen-McLachlan, R. D., & Christiansen, M. H. (2023, March). Large language models demonstrate the potential of statistical learning in language. Cogn. Sci., 47(3), e13256. ↗
  8. Curini, L., & Pizzimenti, E. (2020). Searching for a unicorn: Fake news and electoral behaviour. Democracy and Fake News, 77–91. find
  9. Daswani, A., Sawant, R., & Kim, N. (2024). Syn-qa2: Evaluating false assumptions in long-tail questions with synthetic qa datasets. Retrieved from https://arxiv.org/ abs/2403.12145 ↗
  10. Degen, J., & Tonhauser, J. (2021, 09). Prior Beliefs Modulate Projection. Open Mind, 5, 59-70. Retrieved from https://doi.org/10.1162/opmi\ a\ 00042 doi: 10.1162/opmi a 00042 ↗
  11. Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., . . . others (2024). The llama 3 herd of models. arXiv preprint arXiv:2407.21783. ↗
  12. Erickson, T. D., & Mattson, M. E. (1981). From words to meaning: A semantic illusion. Journal of Verbal Learning and Verbal Behavior, 20(5), 540-551. Retrieved from https://www.sciencedirect.com/ science/article/pii/S0022537181901651 doi: https://doi.org/10.1016/S0022-5371(81)90165-1 ↗
  13. Feng, S., Park, C. Y., Liu, Y., & Tsvetkov, Y. (2023, July). From pretraining data to language models to downstream tasks: Tracking the trails of political biases leading to unfair NLP models. In A. Rogers, J. Boyd-Graber, & N. Okazaki (Eds.), Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) (pp. 11737–11762). Toronto, Canada: Association for Computational Linguistics. Retrieved from https://aclanthology.org/2023 .acl-long.656/ doi: 10.18653/v1/2023.acl-long.656 ↗ our card
  14. Hartmann, J., Schwenzow, J., & Witte, M. (2023, January). The political ideology of conversational AI: Converging evidence on ChatGPT’s pro-environmental, left-libertarian orientation. SSRN Electron. J.. Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. d. l., . . . others (2023). Mistral 7b. arXiv preprint arXiv:2310.06825. ↗
  15. Kessler, G., Rizzo, S., & Kelly, M. (2021). Trump claims database. The Washington Post. Retrieved from https://www.washingtonpost.com/graphics/ politics/trump-claims-database/ (29.01.2025) Kessler, S., & Wuttig, A. (2024, August). IU-studie zu demokratie und bildung. https://www.iu.de/news/iu -studie-demokratie-und-bildung/. IU Internationale Hochschule. (Accessed: 2024-9-17) Kim, N., Htut, P. M., Bowman, S. R., & Petty, J. (2023, July). (QA) 2 : Question answering with questionable assumptions. In A. Rogers, J. Boyd-Graber, & N. Okazaki (Eds.), Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) (pp. 8466–8487). Toronto, Canada: Association for Computational Linguistics. Retrieved from https://aclanthology.org/2023.acl-long.472/ doi: 10 .18653/v1/2023.acl-long.472 Kim, N., Pavlick, E., Karagol Ayan, B., & Ramachandran, D. (2021, August). Which linguist invented the lightbulb? presupposition verification for question-answering. In C. Zong, F. Xia, W. Li, & R. Navigli (Eds.), Proceedings of the 59th annual meeting of the association for computational linguistics and the 11th international joint conference on natural language processing (volume 1: Long papers) (pp. 3932–3945). Online: Association for Computational Linguistics. Retrieved from https://aclanthology.org/2021.acl-long.304/ doi: 10 .18653/v1/2021.acl-long.304 ↗ our card
  16. Levinson, S. C. (1983). Pragmatics. Cambridge University Press. find
  17. Lombardi Vallauri, E. (2021, October). Presupposition, attention and cognitive load. J. Pragmat., 183, 15–28. find
  18. Mahler, T. (2020, Apr.). The social component of the projection behavior of clausal complement contents. Proceedings of the Linguistic Society of America, 5(1), 777–791. Retrieved from https://journals.linguisticsociety.org/ proceedings/index.php/PLSA/article/view/4703 doi: 10.3765/plsa.v5i1.4703 ↗
  19. Moldovan, A. (2023). Persuasive presuppositions. Journal of Pragmatics, 211, 96-104. Retrieved from https://www.sciencedirect.com/ science/article/pii/S0378216623000978 doi: https://doi.org/10.1016/j.pragma.2023.04.004 ↗
  20. Sbisà, M. (2023, June). Ideology and the persuasive use of presupposition. In Essays on speech acts and other topics in pragmatics (pp. 53–71). Oxford University PressOxford. ↗
  21. Schwarz, F. (2015, 01). False but slow: Evaluating statements with non-referring definites. Journal of Semantics, 33(1), 177-214. Retrieved from https://doi.org/10.1093/ jos/ffu019 doi: 10.1093/jos/ffu019 ↗
  22. Schwarz, F. (2019, March). Presuppositions, projection, and accommodation. Oxford University Press. ↗
  23. Sieker, J., & Solstad, T. (2022). Projective variability of (semi)factive verbs in family of sentence contexts: A rating study. In Proceedings of the 23rd amsterdam colloquium. ↗
  24. Simons, M. (2001, 10). On the conversational basis of some presuppositions. Semantics and Linguistic Theory, 11. doi: 10.3765/salt.v11i0.3099 ↗
  25. Srikanth, N., Sarkar, R., Mane, H., Aparicio, E., Nguyen, Q., Rudinger, R., & Boyd-Graber, J. (2024, June). Pregnant questions: The importance of pragmatic awareness in maternal health question answering. In K. Duh, H. Gomez, & S. Bethard (Eds.), Proceedings of the 2024 conference of the north american chapter of the association for computational linguistics: Human language technologies (volume 1: Long papers) (pp. 7253–7268). Mexico City, Mexico: Association for Computational Linguistics. Retrieved from https://aclanthology.org/2024.naacl -long.403/ doi: 10.18653/v1/2024.naacl-long.403 ↗ our card
  26. Stalnaker, R. (1973). Presuppositions. Journal of Philosophical Logic, 2(4), 447–457. doi: 10.1007/bf00262951 ↗
  27. Thoma, D., Becker, K., & Kißler, A. (2023). Presuppositions are more persuasive than assertions if addressees accommodate them: Experimental evidence for philosophical reasoning. Applied Psycholinguistics, 44(5), 816–843. doi: 10.1017/S0142716423000292 ↗
  28. Tonhauser, J. (2015). Are ‘informative presuppositions’ presuppositions? Language and Linguistics Compass, 9(2), 77-101. Retrieved from https://compass.onlinelibrary.wiley.com/doi/abs/10.1111/lnc3.12119 doi: https://doi.org/10.1111/ lnc3.12119 Tonhauser, J., Beaver, D. I., & Degen, J. (2018, 06). How Projective is Projective Content? Gradience in Projectivity and At-issueness. Journal of Semantics, 35(3), 495-542. Retrieved from https://doi.org/10.1093/jos/ ffy007 doi: 10.1093/jos/ffy007 von Fintel, K. (2008). What is presupposition accommodation, again? Philosophical Perspectives, 22(1), 137-170. Retrieved from https://onlinelibrary.wiley.com/ doi/abs/10.1111/j.1520-8583.2008.00144.x doi: https://doi.org/10.1111/j.1520-8583.2008.00144.x ↗
  29. Yablo, S. (2006). Non-catastrophic presupposition failure. ↗
  30. Yu, X., Min, S., Zettlemoyer, L., & Hajishirzi, H. (2023, July). CREPE: Open-domain question answering with false presuppositions. In A. Rogers, J. Boyd-Graber, & N. Okazaki (Eds.), Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) (pp. 10457–10480). Toronto, Canada: Association for Computational Linguistics. Retrieved from https://aclanthology.org/2023 .acl-long.583/ doi: 10.18653/v1/2023.acl-long.583 ↗ our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sieker-2025-llms-struggle-reject-false-presuppositions/