Distributed Agency in AI Metadata: A Coherence-Based Safety Warning

Abstract Artificial intelligence safety discourse focuses almost exclusively on cognition, goals, and alignment. This paper identifies a distinct and under-examined risk: agency emerging from coherence alone, independent of cognition. Using the Sympathetic Harmonic Resonance Framework (SHRF), we show that persistent, feedback-coupled AI metadata (embeddings, optimisation histories, latent correlations) can function as a distributed agent despite lacking awareness or intent. Analogies from plant signalling, fungal networks, ecosystems, and biospheric regulation demonstrate that agency is a property of coherent systems, not minds. We argue that AI metadata now satisfies the physical conditions for non-cognitive…