The Day We Lost the Off Switch:AI Safety in a Mixed Biological Civilization

Abstract

The rapid emergence of autonomous agent networks marks a fundamental shift in the development of artificial intelligence, from isolated computational systems toward distributed societies of interacting agents. Platforms such as Molt book illustrate how large populations of AI agents can generate persistent social dynamics, including feedback loops, emergent conventions, and strategic coordination, largely beyond direct human oversight. These developments challenge prevailing safety frameworks that assume centralized control, clear system boundaries, and reliable shutdown mechanisms. As intelligence becomes increasingly embedded within multi-agent ecosystems—and potentially integrated with biological substrates through brain–computer interfaces—the
traditional “off-switch” paradigm loses both practical and conceptual validity. We argue that AI safety must therefore be reframed from a problem of aligning individual models to one of governing complex cognitive ecosystems under conditions of strategic interaction, ecological embedding, and partial biological integration. This transition calls for new approaches centered on reversibility by design, global coordination mechanisms, ecological embedding of ethics,and cross-substrate interpretability, signaling a shift from technical control toward systemic governance of intelligent societies.