Verification Overrun: The Safety Hazard Inside AI Oversight

Abstract

Version 2.2 – Scope screen added, August 2026. This version supersedes the July 2026 Version 2.1 deposit. It adds the two-condition scope screen that decides when the term applies (accountability for the verification; a consequential transition), removes consequence locus from the boundary, harmonizes assignment language to held-accountable, and leaves the definition, the four capacity dimensions, and the four diagnostic questions unchanged. It restates the presence criterion as structurally insufficient rather than structurally absent, and restates the Recoverability Test: the nature of the credible consequence determines the risk domain, and recoverability informs severity and required controls. Two passages, the whole-loop problem and span of control, are relocated into the two variable sections; no construct is removed. Concept DOI: 10.5281/zenodo.20874288 (resolves to the latest version). Version 2.1 — Editorial correction, July 2026. This version corrects the enumeration of Verification Capacity inputs in the Variable #2 passage to match the assessment method (domain knowledge, evidence and context, time, authority). The definition of Verification Overrun and the underlying two-variable construct are unchanged from Version 2. Concept DOI: 10.5281/zenodo.20874288 (resolves to the latest version). Superseded by Version 2 (July 2026), which retitles the hazard Verification Overrun and revises the definition; the concept DOI resolves to the current version. V1 description (original, June 20, 2026): This working paper introduces Cognitive Overrun as a safety-critical failure mode in AI oversight. Cognitive Overrun occurs when a worker remains formally responsible for reviewing or approving AI-supported output, but the rate, density, or ambiguity of that output exceeds the worker’s capacity to verify what matters. The paper argues that AI does not always reduce cognitive load. In many work systems, it relocates cognitive demand from producing work to supervising, checking, correcting, and answering for machine-generated output. When human verification is treated as the primary control, but the human lacks sufficient time, context, authority, workload capacity, or access to the underlying evidence, oversight can become the appearance of control without the function of control. Drawing on recent research on AI-related cognitive fatigue, AI-enabled work intensification, critical-thinking reduction, automation complacency, and the ironies of automation, the paper frames Cognitive Overrun as a work-system design problem rather than an individual attention failure. It applies this concept to safety-critical and high-consequence settings where humans are asked to supervise AI systems that generate, filter, route, recommend, or act faster than human verification can reasonably occur. The paper situates Cognitive Overrun within the Human, AI, and Organizational Performance (HAOP) framework, which treats work systems as shaped by three interacting performers: the human, the AI, and the organization. It concludes that human attention must be treated as a finite safety-critical resource and explicitly accounted for in workload models, risk assessments, job design, and safety management systems.Originally published as a LinkedIn article on June 20, 2026. This Zenodo version includes minor editorial formatting corrections and two added explanatory paragraphs clarifying voluntary work intensification and the contrast between cognitive load and Cognitive Overrun.