Instructive artificial intelligence (AI) for human training, assistance, and explainability

AI can teach humans more effectively than traditional explanations: Rather than simply explaining why an AI made a decision, this approach has AI actively instruct humans by identifying specific strategic improvements. In Hanabi experiments, this instructive method improved human performance while simultaneously making AI reasoning more understandable.

Humans are often unaware of their own actual strategies: The research reveals a significant gap between how people think they perform tasks and how they actually perform them. This "inaccurate self-assessment" is a critical barrier to explainability—traditional AI explanations may fail because humans don't recognize themselves in the description of their own behavior.

AI can reverse-engineer human decision-making by observation: By analyzing human actions rather than relying on self-reported explanations, the AI can estimate the underlying "weights" or priorities driving human choices. This allows for concrete, actionable feedback like "value factor A more and factor B less" rather than abstract explanations.

Instructive AI bridges the gap between AI performance and human understanding: The approach addresses a key challenge in responsible AI: explanations are only useful if humans can actually understand and implement them. By framing AI insights as concrete instructions for improvement, this method makes explainability more practical and actionable.

This approach has dual benefits for both performance and transparency: The technique simultaneously improves human task performance while increasing human understanding of AI reasoning, suggesting that explainability and effectiveness need not be trade-offs.