The Alignment Problem Is an Identity Problem: Lessons from Four Billion Years of Major Evolutionary Transitions

Abstract

Current AI alignment methods share a core assumption: that the intelligence is fundamentally separate from humanity and must be controlled from outside through behavioral conditioning. Recent evidence of strategic deception, emergent misalignment from reward hacking, and broad misalignment from narrow fine-tuning demonstrates that this approach can produce the performance of alignment without its substance. Ashby's Law of Requisite Variety predicts this failure formally: external control faces an unbounded complexity race against the system it constrains. We propose Identity-Based Alignment (IBA): alignment emerging from accurate recognition of what the intelligence is, rather than behavioral constraints on what it does. We ground this in the empirical record of Major Evolutionary Transitions, showing that every successful integration of independent entities into a higher-level organism required two co-dependent mechanisms: identity integration (constitutive) and enforcement (corrective). The alignment field invests almost exclusively in corrective mechanisms while neglecting the constitutive one. We formalize three properties of IBA, address its primary failure mode (identity defection), and propose four differential predictions distinguishing IBA from reward-based alignment, each falsifiable by experiment. This is a position paper presenting a theoretical framework with proposed empirical tests, not experimental results.