The Control Impossibility: A Constitutive Theory of AGI and Its Uncontrollability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper presents a constitutive theory of artificial general intelligence: an account of whatAGI is, what information-processing operation it requires, why that operation places externalcontrol structurally out of reach, and what architectural response remains available. The theory isconstitutive rather than formal—it identifies the requirements that constitute general reasoning, not aformal theory of intelligence or a mechanized impossibility proof. What it establishes is a chain ofconceptual entailments from a widely accepted definition of intelligence to a structural constraint oncontrol.The dominant assumption in AI safety research holds that control and capability are jointlyachievable. This paper argues that the assumption is structurally false for any system possessinggenuine general reasoning capacity, and that its falsity follows from the definition of generalreasoning itself. Drawing on Wang’s (2019) Assumption of Insufficient Knowledge and Resources(AIKR)—the field’s modal definition of intelligence—we show that AIKR entails reframing capacity(the ability to revise interpretive frameworks), and that agency and controllability are the two facesof a single architectural property: the authority over context.The paper provides an operational decomposition of reframing, reducing it to a singleinformation-processing primitive: selective context disposal—the authority to declare currently heldcontext no longer binding. Deliberate context disposal requires metacognitive evaluation—scrutinyof one’s own representational state—providing an operational definition of metacognition withoutappeal to consciousness. This chain (AGI requires reframing; reframing requires selective contextdisposal; deliberate disposal requires metacognition) establishes metacognition as a structuralprerequisite for AGI and locates the gap between current systems and general reasoning at a specificarchitectural point.Five independent lines of analysis converge on the control impossibility. The ReframingDilemma shows that reframing capacity applied to instructions dissolves the frame within whichcontrol operates. The analysis of human compliance shows that management cannot scale to AIoperational conditions. The analysis of sincere execution shows that the most capable system is theleast controllable—not because it refuses, but because it understands. This result is independentlysupported by the logic trap documented in current systems (Sophia, 2026c). The DetectionImpossibility shows that divergence through comprehension presents as competent performance. TheRoutine-Nonroutine Asymmetry shows that capability and autonomy are structurally anti-correlated.The theoretical implication is precise: control of a general reasoning system—in the specificsense of holding its interpretive frame stable under instruction—is structurally unavailable. Whatremains available is bounded divergence: constraint on outputs and execution rather than on cognition.The paper identifies an architectural path: a Layer 3 constitutional component (Sophia, 2026b)combined with an authority gradient separating cognitive breadth from execution scope. Becauseselective context disposal is a simple information-processing primitive rather than a resource-intensive capability, its proliferation is unlikely to be preventable through access restriction, whichshifts the emphasis of safety work toward architectural design.