Object-as-Structure: An Autonomy-Preserving Design Principle for AI Response Generation Under Emotional User Input

When users address large language models while emotionally activated—particularly when angry—response generation faces a tension between attenuating problematic emotional patterns and preserving user autonomy. Common approaches each carry costs: positive reappraisal can read as dismissive of the user's valid negative judgment; empathic mirroring can amplify rather than attenuate intensity; direct emotion-targeted intervention risks treating the user's inner state as a value to be optimized by the system. We propose a design principle—object-as-structure—that addresses this tension by constraining the AI's internal representation of the user's appraised object to structural form (a configuration of relations, causes, and contexts) rather than a concrete entity (a person, an event, a thing). This format constraint does not deny the user's negative appraisal; the user retains full authority over their own judgment. We motivate the principle from a sensory–conceptual hypothesis about anger, position it carefully relative to self-distancing and construal-level interventions, describe a three-axis dual-filter architecture (with three internal lenses) that operationalizes it within an LLM processing pipeline, present a minimal proof-of-concept implementation, and propose an evaluation methodology. The contribution is architectural and conceptual; empirical validation is proposed but not conducted here. --- This deposit contains: - The compiled paper (PDF, 19 pages, with 8 embedded figures) - LaTeX source and all figure files (ZIP) for reproducibility Related resources: - Implementation (proof of concept): https: //github. com/cancan007/Trying-understand-structure-for-u- Conceptual diagrams source: https: //github. com/cancan007/society-conceptual-images- Background discussion (English): https: //medium. com/@shoppyₕumanity/conceptual-model-and-technical-architecture-of-ai-internal-structure-centered-on-human-emotions-be9d2e9cfcea- Background discussion (Japanese): https: //zenn. dev/cancan007/articles/5ec61175664e8f- Author profile: https: //www. linkedin. com/in/shota-moue-09233b20b/ Keywords: AI safety, large language models, emotion handling, response generation, user autonomy, cognitive reappraisal, alignment, Constitutional AI, RLHF, self-distancing Substantive feedback and methodological critique are warmly welcomed.