Explingo: Explaining AI Predictions using Large Language Models

LLMs can make AI explanations more accessible: The paper demonstrates that Large Language Models can successfully transform technical AI explanations (like those from SHAP) into natural-language narratives that are easier for non-experts to understand, achieving high scores on accuracy, completeness, fluency, and conciseness metrics.

Human guidance improves explanation quality: The research found that LLM-generated explanations perform significantly better when trained on "a small number of human-labeled and bootstrapped examples," suggesting that combining AI with human input produces more reliable explanations for decision-making contexts.

Evaluation of explanations remains challenging: While the Explingo system successfully grades narratives in straightforward cases, the authors identified that "effectively scoring narratives in complex domains" remains difficult, indicating that assessing explanation quality is itself a non-trivial problem requiring further work.

Narrative explanations are now more practical: The authors released Explingo as an open-source tool, making narrative explanation technology available for real-world applications beyond research, potentially improving how organizations communicate AI decisions to stakeholders and end-users.

Multiple quality dimensions matter for explanations: The paper emphasizes that good explanations require balancing multiple factors—accuracy, completeness, fluency, and conciseness—rather than optimizing for a single metric, which is important for practitioners designing explanation systems for high-stakes decisions.