Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills

Paper Title

Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills

Publication Info

• Topic area: AI-assisted misinformation detection and its impact on human discernment skills.
• Keywords: AI chatbots, misinformation detection, human-AI interaction, cognitive dependency, persuasive dialogue, critical thinking, skill degradation, longitudinal study, fake news, AI reliance.

Background and Problem

• Problem / challenge: While AI systems can reduce belief in specific false claims, it is unclear if they help users develop lasting skills to independently detect misinformation.
• Significance: Misinformation, especially AI-generated, poses significant societal risks, including public panic and economic consequences, necessitating effective interventions.
• Motivation and related work: Prior research has shown AI chatbots can reduce belief in conspiracy theories through persuasive dialogues. However, concerns about overreliance and lack of independent skill development remain unaddressed. This study investigates whether AI dialogues foster discernment or dependency.

Solution

• Proposed approach: A web-based AI chatbot designed to engage users in persuasive dialogues for misinformation detection, evaluated through a month-long longitudinal study.
• Novelty:
  1. Development of an AI chatbot integrating persuasive dialogue and artifact detection for misinformation evaluation.
  2. First longitudinal study measuring misinformation detection with and without AI assistance over time (N=67).
  3. Identification of a trade-off: immediate accuracy gains during AI-assisted sessions (+21%) but long-term declines in unassisted performance (−15.3% by week 4).
  4. Analysis of 24 conversational strategies to understand their impact on learning versus dependency.
• Procedure and key techniques:
  • Participants evaluated news items in three phases: before AI interaction, with AI assistance, and after AI interaction.
  • AI dialogues included up to 9 rounds of evidence-based persuasion.
  • Data was collected on accuracy, confidence, and conversational dynamics, with 7,203 Human-AI conversation pairs analyzed.

Results

• Concrete findings:
  • AI-assisted accuracy improved by +21.3 percentage points on average across sessions.
  • Unassisted accuracy declined by 15.3% by week 4 compared to week 0, particularly for detecting fake content.
  • Strategies like guided questioning supported long-term learning, while confidence calibration and devil’s advocate approaches fostered dependency.
• Advantage over baselines:
  • Immediate belief correction was effective during AI-assisted sessions, with participants showing significant accuracy gains compared to their unaided baseline performance.
• Experiments / evaluation:
  • Longitudinal study with 67 participants over four weeks.
  • Participants evaluated 49 news items (real and fake) across three phases.
  • Metrics included accuracy, confidence, and conversational strategy correlations.
• Limitations and future work:
  • Small dataset (49 validated items) and participant pool (N=67) may limit generalizability.
  • Study focused on US/UK populations, potentially limiting cross-cultural applicability.
  • No control group without AI assistance for comparison.
  • Future work should explore longer study durations, cross-cultural validation, and alternative AI pedagogical approaches like Socratic questioning.

Summary

This study demonstrates that while AI chatbots can effectively reduce belief in misinformation during interactions (+21% accuracy gain), they fail to build lasting independent discernment skills, with unassisted accuracy declining by 15.3% over four weeks. The findings highlight a dependency paradox, where reliance on AI undermines users’ ability to detect fake content independently. Strategies like guided questioning showed potential for fostering learning, but others, such as confidence calibration, increased dependency. The results emphasize the need for AI systems designed to enhance critical thinking rather than replace human reasoning, with implications for misinformation detection and broader domains requiring human-AI collaboration.