Answers are grounded in this paper and cite their sources.
Position: Before Scaling Oversight, Measure the Overseer — The Reliability Axis of AI Governing AI AI systems increasingly supervise AI systems. The 2026 Singapore Consensus names the gap: "the alignment of overseer AIs is itself unverified." This position paper argues the gap is primarily a reliability problem and proposes the reliability axis of AI-governs-AI as a research direction: (1) four dominant families (constitutional AI, LLM-as-judge, self-critique, scalable oversight) each treat the overseer as an instrument, never as a failure source; (2) the precise fallacy is same-origin governance — shared weights/corpora/alignment mean joint failure cannot fall below the common cause (measured φ = 0.916 for same-model pairs, Zenodo 23696391); (3) the governing discount 1−ρ_cross prices the independence an overseer actually delivers; (4) legitimacy is layered: deterministic green core, heterogeneous-model yellow tier, human red tier; (5) three reporting standards: ρ for every oversight layer, the discount on the invoice, provenance triplicity for heterogeneity claims. A preregistered six-family measurement protocol is registered (Zenodo 23719903). Alternative views are stated and answered.
No takes yet. Share an insight, caveat, or question.
Cunxun Zang (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: