Finding Uncommon Ground: A Human-Centered Model for Extrospective Explanations

One-size-fits-all explanations don't work: The paper argues that traditional AI explanations focusing on internal model mechanisms are "often unsuitable for non-experts," highlighting that transparency alone doesn't equal understanding for most users.

Personalization is key to useful explanations: Rather than providing the same technical details to everyone, AI systems should tailor explanations based on individual user preferences and context, making information more relevant and actionable for each person.

AI systems should learn from past interactions: The paper proposes that AI agents maintain a "dynamic memory of previous interactions" with users, allowing them to avoid repeating information and instead focus on what's genuinely new to that specific person.

User-centered design matters more than technical accuracy: The research emphasizes that effective AI explanations require understanding "individuals and their preferences as well as the context in which explanations are given"—shifting focus from what the AI knows to what the user actually needs to know.

Explanations should reflect the agent's understanding of the user: By modeling "the agent's worldview" of each user, AI systems can better estimate knowledge gaps and provide explanations that bridge those gaps, rather than overwhelming users with irrelevant technical details.