Conversational Explanations: Discussing Explainable AI with Non-AI Experts

One-time explanations aren't enough: Traditional AI explanations provide static information that doesn't adapt to different users' knowledge levels and needs. Conversational explanations—where users can ask follow-up questions—better serve diverse audiences by allowing customization based on individual understanding.

Synthetic training data can work, but needs safeguards: The research shows that AI systems can be trained on artificially generated conversations to build explanation systems, but this approach requires two critical protections: methods to ensure diverse responses (using a "repetition penalty") and filters to catch false or misleading information ("hallucination detection").

Conversational explanations significantly improve user understanding: In human testing with 60 participants, the EMCEE system substantially improved users' comprehension, acceptance, trust, and willingness to collaborate with AI systems compared to traditional static explanations—suggesting this approach has real practical value for responsible AI deployment.

Quality matters more than quantity in synthetic data: Rather than simply generating large amounts of training data, filtering out untruthful responses and promoting diverse answers led to better performance, demonstrating that data quality is essential when building AI systems meant to explain other AI systems.

Truthfulness is achievable in conversational AI: The research demonstrates that conversational explanation systems trained on carefully curated synthetic data can generate more truthful and understandable answers, addressing a key concern about whether AI systems can reliably explain themselves to non-experts.