NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 25·8 min read
Human Decision-Making with Persuasive and Narrative LLM Explanations
TL;DR
Prior work showed people find AI narrative explanations trustworthy and persuasive, but their actual impact on decision accuracy remained unclear. Researchers conducted a large-scale human behavioral experiment testing how LLM-generated explanations of varying persuasiveness affected classification task performance.
✦ Why It Matters
Engineers deploying AI explanations should recognize that persuasive narratives may increase over-reliance on AI without improving accuracy.
Key Takeaways
- Narrative explanations increase human reliance on AI predictions regardless of correctness—monitor for over-dependence.
- Highly persuasive explanations may impair users' ability to detect AI errors—balance clarity with critical thinking.
- Adding explanations does not guarantee better decision accuracy—validate actual performance impact empirically.
- Consider tradeoffs between explanation persuasiveness and decision quality when designing AI-assisted systems.
Full Summary
When AI systems provide predictions, they often include explanations to help humans understand the reasoning. Large language models (LLMs)—neural networks trained on vast text—can generate natural-language narratives explaining their predictions.
Prior studies found these narratives feel understandable and convincing to people, but whether they actually improve decision-making performance remained unknown. Researchers conducted a large-scale human experiment where participants made classification decisions using AI predictions paired with LLM-generated narrative explanations designed to vary in persuasiveness.
Results showed that narrative persuasiveness did not meaningfully improve decision accuracy compared to predictions alone, consistent with prior explainable AI research. However, narratives did increase human reliance on AI predictions—both correct and incorrect ones.
More persuasive narratives appeared to slow response times and reduce participants' ability to distinguish between accurate and inaccurate AI outputs.