From Consumption to Reflection: Designing Human-AI Relations for Stable Reasoning

The paper introduces Relational Reflective Intelligence (RRI), a governance layer designed to enhance reasoning in interactions between humans and large language models (LLMs). RRI operates externally to LLMs and comprises three components: the Rose-Frame for identifying reasoning breakdowns, the Architect's Pen for introducing reflection at critical moments, and an inference-time workflow for embedding these processes without model retraining. This approach aims to mitigate cognitive vulnerabilities shared between humans and LLMs, promoting a structured interaction that enhances decision-making reliability and addresses AI safety as a cognitive architecture challenge.