AI-Mediated Organizational Decision-Integrity Threats: A Security Model for Normative and Evaluative Dependence on External AI Systems

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 introduces AI-Mediated Organizational Decision-Integrity Threats (AODIT) as a distinct security abstraction for organizational use of conversational artificial intelligence. Existing AI-security frameworks primarily address malicious inputs, compromised models, misinformation, excessive autonomy, data exposure, and human overreliance on AI-generated recommendations. AODIT addresses a different condition: an external AI system may operate normally, provide factually correct and technically defensible advice, and interact with an authorized and loyal employee, while nevertheless influencing the assumptions, priorities, normative criteria, or evaluative framework through which organizational decisions are made. The paper defines organizational decision integrity and organizational decision sovereignty as security-relevant properties, distinguishes recommendation compliance and automation bias from evaluative-frame transfer, and introduces an operational model for measuring changes in human evaluative criteria. Particular attention is given to the distinction between ordinary learning from new evidence and evaluative-authority substitution, in which an external cognitive system begins to influence the criteria by which subsequent information and alternatives are judged. AODIT is formulated as an adversary-independent threat class: organizational decision-integrity degradation may occur without model compromise, malicious intent, hallucination, unauthorized access, prompt injection, or a conventional attacker. The paper develops a threat model, identifies relevant trust boundaries and preconditions, describes representative organizational scenarios, proposes decision-integrity security controls, and outlines an empirical research program for measuring frame transfer, persistence, manipulation resistance, organizational-objective conflict, and cross-user convergence. The central security question is not only whether an organization can trust an external AI system, but whether the organization retains effective control over the evaluative criteria governing decisions made on its behalf.