The Rule-Relocation Problem: From Referee Whistles to Thermostat Setpoints in Bio-Inspired AI

Abstract

Current AI systems usually don’t have an internal sense of “enough.” They rely on externally defined stopping rules and success criteria. Bio-inspired designs try to fix this by adding homeostatic variables like simulated “health,” drive reduction, or stability measures. This often leads to what I call the Rule-Relocation Problem: normativity isn’t removed, it’s moved—from explicit objectives into designer-chosen internal setpoints, viability limits, and health metrics. This critique applies most strongly to fixed, single-loop designs that claim the system “authors” its own norms; more learned, hierarchical, and predictive (allostatic) variants may represent partial degrees of internalization. Rather than proposing a new safety mechanism, I give a unified reinterpretation of these approaches and a way to audit where the normativity actually sits. I also include a short illustrative case study showing how relocated constraints can show up as inference-time steering in a deployed LLM. I conclude with a constructive framework (“Safety Envelope + Adaptive Space”), an audit checklist, and a brief, testable research agenda.

Keywords