The Impact of Narrative AI Explanations on Human Decision-Making: The More Persuasive, the Weaker the Judgment?

This article introduces a large-scale human behavior experiment study published on arXiv in May 2026. Key findings: While narrative explanations generated by LLMs increase users' trust in AI, they do not improve decision accuracy; instead, they may prolong decision time and weaken the ability to distinguish between correct and incorrect AI outputs. This study sounds an alarm for AI interpretability design.

Original paper link: http://arxiv.org/abs/2605.23867v1

AI explanations aim to help users understand and trust AI decisions. Traditional methods like feature importance and attention visualization provide transparency, but technical jargon is unfriendly to non-professional users.

The rise of LLMs has brought narrative explanations, which use natural language paragraphs to elaborate reasoning (e.g., medical diagnosis cases). Their advantages include understandability, credibility, and persuasiveness, but whether these advantages translate into better decisions remains questionable.

Experimental task: Participants completed classification decisions under three conditions—AI prediction only, AI + low-persuasion explanation, AI + high-persuasion explanation.

Key metrics: Decision accuracy, AI dependence, response time, discriminative ability.

Persuasion manipulation: Optimize the persuasiveness of explanations through rhetorical techniques, evidence citation, narrative structure, and emotional appeals.

1. No significant improvement in decision accuracy: Regardless of the persuasiveness of the explanation, there was no statistical difference compared to AI prediction only;
2. Blind increase in AI dependence: High-persuasion explanations increased the acceptance rate of both correct and incorrect AI suggestions;
3. Prolonged decision time: High-persuasion explanations may lead to increased response time due to reading, cognitive processing, etc.;
4. Decreased discriminative ability: Participants' alertness to AI errors decreased, and decision patterns tended to be consistent.

1. Heuristic processing: Users are persuaded by the smooth narrative and stop in-depth examination of the essence of reasoning;
2. Cognitive load shift: Detailed narratives increase cognitive load, diverting attention from core decision factors;
3. Confusion between trust and ability: Mistaking narrative fluency for strong AI ability, leading to over-reliance;
4. Reinforcement of confirmation bias: Narratives provide rich evidence, reinforcing users' initial judgments even when the AI is wrong.

1. Redefine the goal of explanations: Focus on decision quality as the core, not just user satisfaction;
2. Be alert to the persuasion trap: Balance understandability with users' critical thinking;
3. Scenario-matched explanation strategies: Use structured facts + uncertainty prompts for high-risk scenarios, concise predictions + key features for fast decisions, etc.;
4. Cultivate AI literacy: Provide confidence information, encourage questioning and verification, and educate about AI limitations.

Research limitations: Simple task types, sample representativeness, unknown long-term effects, etc. Future directions include comparison of explanation types, research on individual differences, interactive explanations, etc.

Human-AI collaboration should promote meaningful collaboration: Explanations need to help users understand reasoning rather than just persuade, and encourage critical thinking. Conclusion: Smooth narration ≠ correct reasoning; maintaining independent judgment is crucial.