An Information-Theoretic Framework for Comparing Voice and Text Explainability

Text explanations are more efficient for comprehension: The study found that text-based explanations achieved higher comprehension efficiency (CE) compared to voice, suggesting that users can process and understand written explanations more effectively when learning why an AI system made a particular decision.

Voice explanations improve trust calibration: While text was more efficient, voice explanations yielded improved trust calibration error (TCE), meaning users who received spoken explanations developed more accurate and appropriate levels of trust in the AI system—neither over-trusting nor under-trusting it.

Analogy-based explanations offer the best balance: Across different delivery styles tested (brief, detailed, and analogy-based), analogy-based explanations achieved the best overall trade-off between comprehension and trust calibration, suggesting that using relatable comparisons may be particularly effective for responsible AI communication.

Explanation modality matters for user outcomes: The research demonstrates that how an explanation is delivered—not just what is explained—significantly affects both user understanding and appropriate trust levels, highlighting that responsible AI requires attention to communication channels, not just model transparency.

Framework enables future responsible AI research: The paper provides a reproducible, information-theoretic foundation that researchers can use to design and test multimodal explanation systems, supporting the development of more trustworthy AI systems grounded in empirical evidence.