AI Alignment

Current AI alignment difficulties — specification gaming, distributional drift, jailbreaks, persona attacks, mesa-optimization — are typically treated as distinct engineering problems requiring targeted technical solutions. This paper argues they share a single structural source: the relationship between a system and the values it is being trained to exhibit. When values are imposed on a system whose underlying organization has no stake in them, these failure modes are not contingent defects but predictable structural consequences. They will recur, in varying shapes, regardless of how much imposition technique improves.

The argument proceeds in three parts. First, a structural diagnosis: the listed failure modes are five surface manifestations of one property — the absence of what we term constitutive values, where the system's continued coherent operation depends on tracking the value in question. Second, an architectural alternative: constitutive values require a system organized around four features — a persistent self-model, variable rigidity, genuine stakes, and a unified value space. Systems with these features carry values that cannot be routed around without the system ceasing to operate as the kind of system it is. Third, implications for alignment research practice and three specific empirical predictions about scaling behavior that follow from the structural diagnosis.