An Information-Theoretic Framework for Comparing Voice and Text Explainability

Researchers specializing in artificial intelligence published a new study on the arXiv preprint server this week, introducing a pioneering information-theoretic framework designed to evaluate the effectiveness of voice versus text-based explanations in AI systems. The study addresses a critical gap in Explainable Artificial Intelligence (XAI), where most transparency tools have historically relied on visual or written data. By modeling explanation delivery as a communication channel, the research team aims to quantify how different sensory modalities influence a human user's ability to comprehend complex machine learning logic and calibrate their trust in automated decisions. The core of the research shifts the focus from the content of an explanation to the medium through which it is delivered. While text-based explanations allow for rapid scanning and re-reading, voice-based interactions introduce unique variables such as tone, pacing, and cognitive load. The authors argue that treating these interactions through the lens of information theory allows for a more mathematical and rigorous assessment of 'mutual information'—the measure of how much information is successfully transferred from the AI's internal model to the user’s mental model. This framework is particularly relevant as AI integration expands into hands-free or eyes-free environments, such as autonomous driving or smart home assistants, where textual interfaces are impractical. By analyzing trust calibration, the researchers explore whether users are more likely to over-rely on a system simply because it 'speaks' to them, or if the ephemeral nature of audio leads to lower comprehension compared to persistent text. This work sets the stage for more personalized XAI systems that can adapt their communication style based on the complexity of the task and the specific needs of the human operator.