Rationalize: A Framework for Shared Semantic Reasoning in Human-AI Alignment

A newly introduced framework named Rationalize outlines a structure of role pairs aimed at facilitating shared semantic reasoning between humans and AI models during data-driven sensemaking. It draws upon concepts from human-machine collaboration and critical thinking, defining interactions through complementary roles: Explorer-Guide, Investigator-Informant, Teacher-Student, and Judge-Advocate. These roles function within a common reasoning environment where both entities clarify their objectives, inquiries, assumptions, evidence, conclusions, and implications. The framework emphasizes alignment not just in outputs but also in the rationalization of intentions and actions. It connects these role pairs to mutual human-AI alignment, illustrating the distinctions between 'aligning AI to humans' and 'aligning humans to AI' based on the roles. A collaborative research agenda for designing and evaluating alignment using element-level and role-specific strategies is proposed.