Safety–Relational AI — Designing Safety for Long-Term Human–AI Interaction —

Notice (Version 1.3): This paper is a preliminary report aiming to conceptually present design principles concerning the long-term, iterative relationship between generative AI and humans. Version 1.3 adds connectivity with Shine's process consulting.Version 1.2 revised and expanded the SRAI design philosophy, proposing the concept of gradual enablement. Version 1.1 added prior research on human-centered design (Human-Centered AI/Automation) and socio-emotional interaction, clarifying this paper's positioning. It also included supplementary discussion on design ethics approaches that conceptualize autonomy as a long-term relational capability structure. We are currently translating the entire paper into English and refining terminology to enhance conceptual rigor and precision. Upon completion, the revised version will be published on arXiv and other platforms. Abstract The proliferation of generative AI is transforming human-AI interactions from one-off task execution or information retrieval into ongoing, daily, and iterative relationships. This shift necessitates redefining AI safety not merely as the appropriateness of individual outputs, but as the design of the relational structure itself—the dynamic formed as interactions accumulate over time. This paper defines a design philosophy that places relational safety at the highest level of design principles as Safety-Relational AI (SRAI). Relational safety here refers to the safety of the relational structure defined by the temporal dimension of repeated and accumulated interactions and the set of possibilities for actions, judgments, and involvement that humans maintain in their interactions with AI. This paper views the irreversible degradation of this set of possibilities over time in long-term human-AI interactions as a breakdown of the relational structure. Based on this stance, the paper identifies breakdown modes inherent to long-term, iterative interactions and proposes three principles—Safety, Stability, and Generation—as the SRAI Design Principles (SRAI-DP). Furthermore, assuming the safety of the relational structure is maintained, it introduces Progressive Enablement as a design principle to incrementally expand generative freedom tailored to the user. Finally, this paper presents the SRAI Framework (SRAI-Frame) as a conceptual framework for treating relational structures as analyzable design subjects, and the Safety-Relational Architecture of AI (SRAI-Arch) as an engineering response based on this framework. This paper attempts to clarify a design philosophy that captures the safety of generative AI not at the level of output or functionality, but at the level of the human-AI relationship formed over time. --- 注記(Version 1.3): 本稿は、生成AIと人間の長期的・反復的関係をめぐる設計思想を概念的に先行提示することを目的とした予備的報告である。Version1.3 では、シャインのプロセスコンサルティングとの接続を追加した。Version1.2では、SRAIの設計思想を加筆修正し、漸次的イネーブルメントの概念を提案した。なお、ver.1.1では、人間中心設計(Human-Centered AI / Automation)