An Agentic Approach to Generating XAI-Narratives

AI explanations are becoming more accessible through conversational language: Rather than relying on technical charts and statistics, researchers are using large language models to translate complex AI decision-making into natural-language narratives that non-experts can understand.

Iterative feedback improves explanation accuracy: The study found that having multiple "critic agents" review and refine explanations significantly improved their faithfulness to the actual AI model's reasoning, with one system reducing inaccurate explanations by 90% after three rounds of revision.

Different AI models perform differently at generating explanations: The research tested five different language models and found substantial variation in their ability to generate reliable explanations—Claude-4.5-Sonnet performed best, while DeepSeek-V3.2-Exp showed different patterns, suggesting that model choice matters for responsible AI deployment.

Combining multiple AI reviewers enhances reliability: Using an ensemble approach where multiple critic agents vote on explanation quality consistently improved performance across most models tested, suggesting that redundancy and consensus-building can strengthen AI transparency efforts.

Explainability remains a practical challenge for real-world AI systems: The paper addresses a genuine gap between technical XAI research and what stakeholders actually need—accessible explanations they can trust and act upon—highlighting that making AI interpretable requires ongoing technical innovation.