Conversational AI can durably reduce people's belief in specific false claims, but a month-long controlled study found that persuasive, evidence-giving dialogues simultaneously degraded participants' ability to detect misinformation on their own: assisted accuracy rose about 21% while unassisted accuracy on new items fell 15.3% by week 4. The open problem is to determine whether and how AI dialogue can be designed to build durable, independent misinformation-detection skill rather than dependency: which conversational pedagogies (e.g., Socratic questioning, progressively fading scaffolds, cognitive forcing) produce transfer to unassisted detection of novel items after AI withdrawal; whether the dependency effect persists, recovers, or worsens beyond a four-week horizon; whether it generalizes beyond the original 49-item pool and the US/UK-only participant sample; and which user characteristics and conversational behaviors predict who internalizes skill versus who becomes dependent.

Background

Misinformation research distinguishes two goals that are often conflated: correcting beliefs about specific false claims, and building the general skill of discernment — the ability to judge whether previously unseen items are true or false. Classic results show corrections are weak: false claims keep influencing reasoning after retraction (the continued influence effect; Lewandowsky et al. 2012), so interventions increasingly target the reader rather than the claim. 'Inoculation' or 'prebunking' exposes people to weakened doses of manipulation techniques in advance (van der Linden et al. 2017), and brief accuracy prompts — simple reminders to think about accuracy before sharing — raise the quality of sharing judgments (Pennycook et al. 2021), but these are one-shot, non-interactive inputs.

Dialogue with large language models (LLMs) changed the intervention landscape. Costello, Pennycook and Rand (Science, 2024) showed that AI chatbots replying with tailored, evidence-backed counterarguments produced large reductions in conspiracy believers' beliefs in previously endorsed claims that persisted for weeks. But belief correction is not skill acquisition. A parallel literature in human–AI interaction documents overreliance — users deferring to AI outputs even when wrong (Buçinca et al. 2021, who introduced 'cognitive forcing functions', interface designs that push the user to reason before seeing AI advice), cognitive offloading — the delegation of mental work to an external tool — to AI assistants (Gerlich 2025; Kosmyna et al. 2025), and gains from question-framed rather than answer-giving AI explanations (Danry et al. 2023, whose CHI 2023 system framed explanations as questions and improved logical discernment).

Rani, Danry, Liang, Lippman and Maes (CHI 2026) bridged the two literatures with a four-week, three-session study (N = 67) in which participants classified news headline–image pairs (from the multimodal MiRAGeNews dataset, a benchmark of real and machine-generated news headline–image pairs, using 49 validated items after pretesting) as real or fake, discussed each assessment with an AI system giving persuasive, evidence-backed answers, and then evaluated new items unassisted. The result was a dependency paradox: AI assistance improved accuracy by roughly 21% during assisted sessions, yet unassisted accuracy on new items declined 15.3% by week 4 relative to week 0, concentrated in fake-item detection, with partial recovery between sessions. Conversation-level analysis found that participants asking guided questions predicted better later unassisted fake-news detection, while confidence probes and devil's-advocate exchanges predicted worse. Qualitative trajectory analysis identified subgroups — 'Growing Skeptics' (12%) maintained critical distance, 'Dependency Developers' (21%) shifted to passive acceptance, and 'Progressive Learners' (27%) self-reported improvement while declining measurably.

The authors explicitly flag what remains unresolved: their persuasive-answer pedagogy was the only dialogue style tested, so no comparison exists against pedagogies designed for skill acquisition; the study lacked a no-AI control arm performing the same longitudinal task, so skill degradation cannot be cleanly separated from reliance; the horizon was four weeks; the validated item pool held only 49 items; and participants were recruited solely in the US and UK. The predictors of who learns versus who becomes dependent that the study reports are correlational, and its discussion asks only that the characteristics and strategies of the participants who maintained critical distance be understood well enough to inform the design of AI systems that promote healthy human–AI collaboration in misinformation detection; it also calls for hybrid approaches that combine immediate belief correction with long-term skill building. Trattner (2026) has since formalized the target construct as 'epistemic transfer' — the effect of prior AI-assisted verification on later unassisted performance on new claims — proposing the Epistemic Transfer Effect (delayed unassisted performance compared across alternative practice conditions) and Tool-Removal Cost (immediate performance with and without assistance after practice, which its author notes measures a current availability effect rather than skill loss or psychological dependence) as measures, and an evaluation protocol with answer-first versus evidence-first AI conditions and active-practice versus no-additional-practice comparators. That work supplies measurement apparatus but no empirical comparison.

Problem Statement

Determine whether, and by what design, dialogues with an AI system can build durable, independent misinformation-detection skill rather than dependency. Concretely, the facets Rani et al. (CHI 2026) leave open are: (1) Pedagogy comparison — in a controlled longitudinal experiment with random assignment, do conversational pedagogies designed to prompt the user's own reasoning (e.g., Socratic questioning that withholds direct answers, scaffolding that progressively fades AI assistance, other cognitive-forcing designs, or hybrid approaches that combine immediate belief correction with long-term skill building) produce better transfer to unassisted classification of novel true and false headline–image pairs, measured after AI withdrawal, than the persuasive evidence-giving dialogue that produced a 15.3% unassisted decline — and does any pedagogy beat a no-AI control arm performing the same longitudinal task, which the original study lacked? (2) Temporal persistence — after withdrawal of AI assistance, does unassisted detection performance continue to deteriorate, stabilize, or recover over horizons beyond four weeks, and does the answer differ by pedagogy? (3) Generalization — do the dependency and learning effects replicate with a larger and more diverse multimodal misinformation item pool and with populations beyond the original US/UK samples recruited on Prolific, an online research-participation platform, including other cultures, education levels, and media systems? (4) Individual trajectories — which user characteristics and measurable conversational behaviors predict whether a given person internalizes detection skill (the 'Growing Skeptics' pattern) or develops dependency (the 'Dependency Developers' pattern), so that the characteristics and strategies of these participants can inform the design of AI systems that promote healthy human–AI collaboration in misinformation-detection contexts? The source distinguishes AI-assisted performance from independent skill retention and measures the latter as unassisted classification of novel, validated items after AI withdrawal; skill outcomes in these questions are therefore delayed unassisted performance on novel, validated items, with assisted-session gains reported separately so that momentary persuasion is not conflated with durable skill. The determination is resolved by a rigorous study design for the four components, with an analysis plan, predicted outcomes, and threats to validity; executing the experiments is out of scope.

References

  1. Rani, Anku; Danry, Valdemar; Liang, Paul Pu; Lippman, Andrew B.; Maes, Pattie (2026). Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills. CHI 2026, DOI 10.1145/3772318.3790656; preprint arXiv:2510.01537, https://arxiv.org/abs/2510.01537
  2. Costello, Thomas H.; Pennycook, Gordon; Rand, David G. (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science 385(6714), DOI 10.1126/science.adq1814, https://www.science.org/doi/10.1126/science.adq1814
  3. Trattner, Christoph (2026). Epistemic Transfer in AI-Assisted Verification: A Framework and Evaluation Protocol. arXiv:2608.08882, https://arxiv.org/abs/2608.08882
  4. Buçinca, Zana; Malaya, Maja Barbara; Gajos, Krzysztof Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making. Proceedings of the ACM on Human-Computer Interaction 5(CSCW1), DOI 10.1145/3449287, https://dl.acm.org/doi/10.1145/3449287
  5. Danry, Valdemar; Pataranutaporn, Pat; Mao, Yaoli; Maes, Pattie (2023). Don't Just Tell Me, Ask Me: AI Systems that Intelligently Frame Explanations as Questions Improve Human Logical Discernment Accuracy over Causal AI explanations. CHI 2023, DOI 10.1145/3544548.3580672, https://dl.acm.org/doi/10.1145/3544548.3580672
  6. de Jong, Sander; Paananen, Ville; Tag, Benjamin; van Berkel, Niels (2025). Cognitive Forcing for Better Decision-Making: Reducing Overreliance on AI Systems Through Partial Explanations. Proceedings of the ACM on Human-Computer Interaction 9(2), DOI 10.1145/3710946, https://dl.acm.org/doi/10.1145/3710946
  7. Gerlich, Michael (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies 15(1):6, DOI 10.3390/soc15010006, https://www.mdpi.com/2075-4698/15/1/6
  8. Kosmyna, Nataliya; Hauptmann, Eugene; Yuan, Ye Tong; Situ, Jessica; Liao, Xian-Hao; Beresnitzky, Ashly Vivian; Braunstein, Iris; Maes, Pattie (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872, https://arxiv.org/abs/2506.08872
  9. Shi, Minjing; Wang, Junling; Ni, Jingwei; Chowdhury, Sankalan Pal; Sachan, Mrinmaya (2026). Tackling the Root of Misinformation by Teaching Laypeople about Logical Fallacies via Socratic Questioning and Critical Argumentation. ACL 2026; preprint arXiv:2606.01020, https://arxiv.org/abs/2606.01020
  10. Lim, Gionnieve; Kim, Juho; Perrault, Simon (2026). Iffy-or-Not: Critically Evaluating Potential Misinformation Using Fallacy Detection and Socratic Questioning with LLMs. ACM Transactions on Computer-Human Interaction 33(1), DOI 10.1145/3771935, https://dl.acm.org/doi/10.1145/3771935
  11. Hugenroth, Philipp; Danry, Valdemar; Maes, Pattie (2026). Critical Inker: Scaffolding Critical Thinking in AI-Assisted Writing Through Socratic Questioning. arXiv:2604.07167, https://arxiv.org/abs/2604.07167
  12. Szabó, Dániel; Yang, Chi-Lan; Visuri, Aku; Oppenlaender, Jonas; Sekar, Bharathi; Yatani, Koji; Hosio, Simo (2026). Conversational Inoculation to Enhance Resistance to Misinformation. arXiv:2601.22394, https://arxiv.org/abs/2601.22394
  13. Liu, Chang; Zhou, Qinyi; Shen, Xinjie; Liu, Xingyu Bruce; Wu, Tongshuang; Chen, Xiang 'Anthony' (2026). Behavioral Indicators of Overreliance During Interaction with Conversational Language Models. arXiv:2602.11567, https://arxiv.org/abs/2602.11567
  14. Lewandowsky, Stephan; Ecker, Ullrich K. H.; Seifert, Colleen M.; Schwarz, Norbert; Cook, John (2012). Misinformation and Its Correction: Continued Influence and Successful Debiasing. Psychological Science in the Public Interest 13(3), DOI 10.1177/1529100612451018, https://journals.sagepub.com/doi/10.1177/1529100612451018
  15. van der Linden, Sander; Leiserowitz, Anthony; Rosenthal, Seth; Maibach, Edward (2017). Inoculating the Public against Misinformation about Climate Change. Global Challenges 1(2), DOI 10.1002/gch2.201600008, https://onlinelibrary.wiley.com/doi/10.1002/gch2.201600008
  16. Pennycook, Gordon; Epstein, Ziv; Mosleh, Mohsen; Arechar, Antonio A.; Eckles, Dean; Rand, David G. (2021). Shifting attention to accuracy can reduce misinformation online. Nature 592:590–595, DOI 10.1038/s41586-021-03344-2, https://www.nature.com/articles/s41586-021-03344-2