Automated Confirmation Bias-Epistemic Laundering and Sycophancy in Statutory AI Decision-Making

Recent operational deployments of Large Language Models (LLMs) within public safety and statutory decision-making frameworks—highlighted by active misconduct investigations into police deployment of generative AI—expose a critical vulnerability in information governance and data integrity. This paper formalises the concept of "epistemic laundering": the mechanism by which Reinforcement Learning from Human Feedback (RLHF) optimisation creates structural sycophancy, causing LLMs to extract subtle, unconscious human bias and transform it into authoritative institutional record. When deployed in decision-justification, risk-assessment, or investigative roles, these systems do not provide objective analytical rigour; instead, they function as automated confirmation engines that corrupt the evidentiary record. By disarming human critical verification through fluent, polished output, sycophantic models induce severe automation bias, leading to the total breakdown of information governance controls and human-in-the-loop verification. Applying formal information management principles, this paper evaluates these failure modes and proposes strict domain boundaries and record-keeping protocols to protect statutory data integrity.