GLARE: A Natural Language Interface for Querying Global Explanations

Natural language makes AI explanations more accessible: The paper demonstrates that using conversational language to query explanations—rather than requiring users to understand technical visualizations or statistics—significantly improves how people can understand why image classifiers make decisions. This bridges the gap between complex AI systems and non-technical stakeholders who need to evaluate model behavior.

Interactive questioning beats static reports: Rather than presenting users with fixed explanation documents, the system allows people to ask targeted questions about model behavior (e.g., "Why does the model confuse cats with tigers?"). The paper shows this question-driven approach is more practical than traditional static explanation artifacts for real-world decision-making.

LLMs can serve as trustworthy intermediaries for AI transparency: The system uses a large language model to translate natural questions into technical queries, then converts results back into human-readable insights. This demonstrates that LLMs can effectively mediate between technical AI systems and human users seeking accountability, provided the underlying explanations are sound.

The system maintains robustness across different contexts: The evaluation shows the interface generalizes to novel questions and datasets while remaining resilient to linguistic errors and variations in how users phrase questions—important for practical deployment where users won't follow rigid protocols.

Explainability requires both technical rigor and human-centered design: The paper illustrates that having good explanations isn't enough; they must be presented accessibly through interfaces that match how people naturally seek information, combining statistical accuracy with usable design.