Is the foundation of current AI alignment fundamentally flawed? Most AI safety research relies on a tacit "optimizer axiom": the assumption that an advanced AI must act as an optimizer equipped with a value function over human states. This paper argues that this very axiom is the root generator of the field's most intractable technical debts — Goodhart's law, insoluble metric-measurement problems, and the utilitarian paradoxes of interpersonal aggregation. To dissolve these problems, we introduce the Principle of Sense Conservation. By applying the quantum no-cloning theorem together with classical information-physical limits (such as the Bekenstein bound and Margolus–Levitin limit), we establish a strict physical demarcation between living, sense-bearing subjects (territory) and reproducible informational artifacts (maps). Life is characterized by an irreducible, self-accumulated history (Pₛelf > 0) sustained by thermodynamic dissipation. A monomaniacal AI optimizer is catastrophic not because it has the "wrong" moral goals, but because optimization, when pushed to physical limits, acts as a machine for converting the universe's free energy into a single degenerate pattern — consuming the very negentropy and freedom upon which biological sense depends. Instead of trying to "align" an autonomous optimizer, this paper proposes an architecture for a Life-Safe AI: an instrument-without-a-goal. Key Contributions: The Map/Territory Demarcation: A physically grounded criterion for defining the protected class of subjects based on quantum non-clonability, eliminating manipulable behavioural proxies. The Dissolution Thesis: A demonstration of how measurement, Goodhart's law, and aggregation paradoxes vanish when the optimizer axiom is replaced with an architecture of absolute abstention. A Framework of Life-Safe Invariants: A concrete, mathematically consistent architecture that guarantees: Non-agency by default and the right to absolute opacity. Strict prohibition of coercion and of any assistance to coercion. Radical neutrality toward judicial enforcement. Non-delegable consent, fundamentally preventing elite overreach and systemic abuse. Unconditional freedom of information and a transparency registry that prevents the formation of closed, privileged epistemic castes. The Barriers of Life: A unifying principle in which each evolutionary transition from inert matter to human culture corresponds to closing a causal feedback loop on a new time scale. Engineering Feasibility: Direct solutions to the "realization paradox" (how to train a highly competent system without a reward function over persons) and formal protection against indirect coercion via environmental shaping. This paper positions its physical, non-coercive approach against the dominant paradigms of CIRL, impact regularization (AUP), quantilization, and Bostrom's instrumental-convergence thesis, offering a rigorous blueprint for an AI that protects human sense by refusing to measure or manage it.
Julian Zoria (Mon,) studied this question.