From Local Alignment to Global Relational Coherence: Reframing AI Safety Across Scales through HIDLA

This paper proposes a novel theoretical framework for AI safety that shifts the focus from individual-level alignment to cross-scale relational coherence.

Current AI safety approaches—such as reinforcement learning from human feedback (RLHF), rule-based constraints, and value alignment—have achieved significant progress in controlling individual AI behavior. However, they lack design principles for addressing misalignment across scales, where optimization at one level can generate unintended consequences at another.

To address this gap, this study introduces the concept of Scale Misalignment (SM) as a central problem in AI safety. It further extends the Heart-Inspired Dual-Layer Architecture (HIDLA), originally proposed as a micro-level relational framework, into a multi-scale design principle, and proposes Global Relational Coherence (GRC) as the target state for safe AI systems.

The framework is illustrated through a central conceptual diagram (Figure 1), which presents a four-quadrant structure of AI safety across dimensions of scale (local–global) and optimization unit (individual–relational). This figure serves as a visual summary of the proposed paradigm shift—from alignment-based approaches to relational coherence across scales.

Using social media fragmentation as a primary case, the paper demonstrates how local optimization processes systematically generate cross-scale relational incoherence through a four-stage mechanism. It further shows that such patterns are not merely theoretical, but are increasingly recognized in real-world contexts, including recent judicial decisions concerning platform design.

This preprint is currently under review at AI & Society. The conceptual framework and associated figures are original contributions of the author. Any reuse or adaptation of the figure or framework should appropriately cite this work.

This work reframes AI safety from the correctness of individual outputs to the coherence of relational structures across interacting scales, offering a conceptual foundation for future design principles in socio-technical systems.

Keywords

AI Safety, Scale Misalignment, HIDLA, Relational Coherence, Multi-scale Systems