答え責任と外部化の線引き: AI 協働における認知作業の委譲原理

While AI increasingly absorbs cognitive work, no established principle delimits what may be externalized from what humans must retain. Meaningful Human Control (MHC), developed for autonomous weapons and self-driving vehicles, offers tracking and tracing conditions but has not been translated to cognitive division of labor. Moral responsibility theory has refined answerability—being able to respond to demands for justification—and its application to publishing norms has just been proposed (Floridi 2026). There, however, answerability is distributed among institutional roles (authors, reviewers, editors, publishers); how cognitive work should be divided between human and AI—criteria for the boundary of externalization and a method for demonstrating them—remains unaddressed. This paper imports answerability from responsibility theory, reformulates it as "answerability" (答え責任) for cognitive work, and translates MHC's two conditions into human-AI cognitive collaboration, yielding criteria for the boundary of externalization: what can be externalized is what humans can still answer for—provide grounds, detect and correct errors, and decide adoption. As an existence proof, practice logs (git history, session logs) serve as evidential apparatus making answerability externally auditable. The argument does not depend on the contested existence of responsibility gaps.