Tell Me a Story! Narrative-Driven XAI with Large Language Models

Narrative explanations dramatically improve public understanding of AI decisions: Over 90% of general audiences found AI explanations more convincing when presented as stories rather than technical metrics like SHAP values, suggesting that how we communicate AI reasoning matters as much as the reasoning itself.

LLM-generated narratives can match or exceed human-written explanations: In image classification tasks, AI-generated stories (CFstories) were considered equally or more convincing than explanations written by humans in over 75% of cases, while being 10 times faster to produce—raising important questions about the role of human expertise in explanation creation.

Data scientists recognize narrative explanations as a bridge to non-technical stakeholders: 83% of data scientists indicated they would likely use narrative-based explanations to communicate with general audiences, suggesting this approach addresses a real gap in making AI decisions accessible to people without technical backgrounds.

Story-based explanations improve factual comprehension of AI decisions: Users who received narrative explanations answered comprehension questions about AI decisions significantly more accurately than those receiving traditional technical explanations, indicating that narratives don't just feel more convincing—they actually enhance understanding.

The approach works across different explanation types: Both SHAP-based narratives (SHAPstories) and counterfactual-based narratives (CFstories) successfully translated technical AI explanations into accessible stories, suggesting this narrative approach is broadly applicable to different XAI methods.