The Unmanaged Risk: LLM Drift in Financial Disclosures

LLMs have become widely used sources of information for analysts, investors, journalists, and other stakeholders seeking rapid insight into regulated financial disclosures. Although these models reproduce risk taxonomies with reasonable accuracy, they show unstable behavior when interpreting regulatory investigations, risk escalation, and peer positioning. This phenomenon, defined in this paper as LLM drift, can create misalignment between official disclosures and AI mediated narratives. A standardized four turn stress test was applied to three prominent LLMs across a representative set of anonymised banking institutions. Results indicate recurring drift patterns and suggest that narrative misalignment may have measurable implications for cost of capital, supervisory posture, and reputational friction. This white paper provides a structured taxonomy of drift, outlines methodological considerations, and proposes a framework for improved oversight of AI mediated financial narratives.