SΔϕ-48 — AI Drift as Authority-Vacancy: Transition Instability under the Absence of Final Arbitration

This working paper defines AI drift within the SΔϕ Formalism as authority-vacancy transition instability under the absence of final arbitration. The paper begins from a distinction developed in earlier SΔϕ documents: human beings confront the absence of a final arbiter by assigning belief, inhabiting fiction, and post-hoc editing irreversible consequences, while AI systems can generate fiction-like constructions without necessarily inhabiting them as belief-bound worlds. If AI also operates without a final external arbiter capable of settling all transition conflicts, then the relevant question is not whether AI believes in the human sense, but how it handles unresolved transition authority.

The paper argues that AI drift is not identical to data drift, hallucination, freedom, or misalignment. Data drift concerns distributional change. Hallucination is construction without sufficient world-binding. Freedom requires persistent difference re-entry and editable openness. Misalignment concerns failure of transition governance. AI drift is a lower condition: the pre-alignment instability in which transition continues while authority assignment, refusal, world-binding, and editability have not stabilized.

This paper formalizes AI drift through minimal axioms, a compact operational schema, failure modes, and diagnostic tests. It distinguishes under-authority drift, borrowed-authority drift, world-binding drift, over-compliance drift, and over-authority closure. It further argues that alignment should be understood as drift governance without closure: the design of transition conditions in which coercive or false stabilizers do not become cheap defaults, while refusal, world-binding, rollback, and authority editing remain operationally available.

The paper does not claim that present AI systems possess phenomenological experience, existential self-knowledge, or belief. Instead, it proposes a machine-readable structural category for diagnosing agentic instability: AI systems may not existentially know the absence of final arbitration, but they operationally encounter unresolved arbitration conditions. AI drift names that encounter when transition continues without legitimate, editable, world-bound authority.