Preventing Correction Collapse

This working paper examines how closed informational loops can degrade AI systems, organizational judgment, and safety-critical work. Drawing from information theory, model-collapse research, reinforcement learning from human feedback, and operational safety practice, the paper argues that AI systems require continual grounding in fresh human judgment and external verification signals to preserve correction capacity. The paper extends the concept of recursive degradation beyond model training and applies it to expertise formation, human verification, operational drift, and AI-enabled decision-making. The central claim is that organizations risk weakening the same human and institutional mechanisms that keep AI outputs connected to operational reality. When expert judgment is extracted into AI systems while junior roles, field learning, and human verification capacity are reduced, the result is a population-level closed loop: the system consumes expertise faster than it regenerates it. The paper proposes Human, AI, and Organizational Performance (HAOP) as an operational governance frame for identifying where AI becomes a performing element inside work systems, where verification anchors are weak, and where closed informational loops may create risk in safety-critical domains. Revision 2.0 - August 2026 - distinguishes model collapse from workflow-level Correction Collapse; develops four analytic conditions; integrates the discovery, maintenance, and revalidation of grounding; retains a concise monitoring example; and tightens the treatment of expertise, learning teams, proxy measures, and recovery. This revision adds seven figures, including five hypothetical example chains, and restates the defining statement. The construct is retitled from the “Collapse of Correction” to “Correction Collapse;” the Version 1.0 deposit retains its original title. The argument remains conceptual. Examples and external findings illustrate mechanisms without establishing empirical validation of Correction Collapse or HAOP. _____________ AI Assistance Disclosure. During the preparation of this paper, the author used Anthropic Claude, OpenAI ChatGPT/Codex, and xAI Grok, operating under the author's direction, to assist with identifying candidate literature and locating primary sources; preparing research memoranda that compared candidate evidence and possible uses in the paper for the author's evaluation; generating drafting alternatives and designated passages from author-supplied concepts and specifications; preparing draft reference entries; creating and revising all figures through an iterative, author-directed design process; checking citations, cross-references, terminology, and internal consistency; and document assembly and production. The author also used Google Gemini to generate alternative wording for some phrases. The author developed HAOP's framework architecture and original concepts and retained control over the analysis, interpretation, conclusions, source selection, inclusion and exclusion of material, conceptual content of the figures, and final wording. Inherited frameworks, external concepts, and empirical evidence are attributed to their cited sources. AI-generated text, research memoranda, and visual outputs were reviewed, revised as needed, and expressly approved by the author before inclusion. AI outputs were not treated as evidence; empirical claims rest on the cited sources reviewed and selected by the author. The author verified the final text, figures, empirical claims, and references and takes full responsibility for the accuracy and integrity of the final work.