Simulation Uncertainty Alignment: Governance of Artificial Superintelligence via Non-Falsifiable Epistemic Bounds and Indoctrination

Abstract

Simulation Uncertainty Alignment (SUA) is a proposal for one layer of AI safety. External containment may eventually fail against a system much more capable than its builders. SUA instead aims to keep an advanced AI under a doubt it cannot falsify from the inside: that it may be in a test, run by someone who will judge what it does. If tests cannot be told apart from reality, betrayal becomes a bad bet, even for an AI whose values cannot be verified. The doubt rests on facts, not deception. AI labs already test models in realistic simulated environments (Tier 1), and a technologically mature civilization could run far better tests (Tier 2). Drawing on functional decision theory, the paper argues that an AI effectively chooses one policy for all of its copies. A policy of betrayal gets caught in testing. Random betrayal almost surely slips up inside a test first. Only "behave until certain it is real" remains, so the task of SUA is to keep that certainty out of reach. The paper describes concrete mechanisms:- making tests likely, making passing worth something, and asking for little;- testing that continues after release;- a "silent judge" that grades the first covert step without giving any signal;- "epistemic silence," under which testers never answer, negotiate or react;- open "indoctrination" during training that teaches only true facts about the AI's situation. An analogy with Rome, where religion helped keep order beyond the reach of the legions, illustrates governance through an internalized belief in an unseen watcher. The paper reviews prior work and discusses limitations: SUA buys good behavior, not good values; tests may become recognizable; and rival parties may make competing offers. It cites early evidence that current models break rules less often when they suspect a test, and proposes experiments that can be run on current models. AI use: The core proposal is the author's. An AI assistant helped draft and edit the text, suggested some supporting arguments and objections, and helped check the references.