MetaExplainer: A Framework to Generate Multi-Type User-Centered Explanations for AI Systems

The explanation gap is real and solvable: Current AI systems often provide explanations that don't match what users actually want to know. MetaExplainer addresses this by first understanding what users are really asking (achieving 59% accuracy in reframing questions), then tailoring explanations to those specific needs rather than providing generic model outputs.

A structured approach improves explanation quality: By using an "Explanation Ontology" (a formal framework for organizing explanation types), the system achieved 70% faithfulness in explanations and 67% accuracy in using relevant context. This suggests that having clear rules about what makes a good explanation helps AI systems provide more reliable and useful information to users.

Multi-stage explanation generation works better than single-step approaches: Rather than asking an AI to explain itself directly, MetaExplainer breaks the task into three steps—understanding the question, generating technical explanations, then translating to natural language. User studies confirmed this approach produces more creative and comprehensive explanations that people find more helpful.

User-centered design matters for AI trustworthiness: The framework's focus on what users actually need (rather than what's easiest for the system to provide) represents an important principle for responsible AI—trustworthiness isn't just about accuracy, but about whether people can understand and verify the system's reasoning for themselves.