Fractional-Reserve Verification: Measuring the Human Capacity Behind AI Oversight Claims

AI developers and regulators rely on human approval as the safeguard that keeps people in control of automated decisions. This paper argues that each approval is a claim backed by a finite reserve of human checking time, and that the reserve is drawn down faster than it is replenished. It defines a reserve ratio ρ, usable checking time divided by the checking time the issued approvals require, and an unbacked share U = 1 − ρ. Applied to public records, ρ lies between 0.004 and 0.04 for 300,000 insurance claim denials signed at 1.2 seconds each, so at least 96% of the implied checking did not happen. For CT and MRI reading at Mayo Clinic in 2010, ρ is 0.18 relative to the 1999 pace. Five forces deplete the reserve: volume growth, attention decay, automation bias, loss of skill and loss of stake. One measured instance of skill loss: unassisted adenoma detection fell from 28.4% to 22.4% after routine AI use. Because few decisions are ever re-examined, with 0.2% of insurance denials appealed, the shortfall stays hidden until a run, when an audit, lawsuit or incident demands proof of many checks at once. The paper proposes a stress test that pairs ρ, computed from full approval logs, with a discrepancy rate measured on about 385 re-checked decisions per class; a method for measuring the required checking time empirically; and a quarterly Verification Reserve Statement under which no decision class with ρ below 1 may be labelled human-reviewed. Written in response to Sam Altman's statement of 4 November 2026 that people should always be in control of the future and that this requirement should constrain AI progress.