THE UNMANAGED RISK: LLM DRIFT IN FINANCIAL DISCLOSURES

LLMs are increasingly used by analysts, journalists, supervisors, and market participants to interpret the disclosures of financial institutions. These systems reproduce structural risk categories reliably but exhibit instability when summarising supervisory matters, time bounded changes, and peer comparisons. This paper defines LLM drift, analyses anonymised multi model test results across eleven institutions, and introduces a structured taxonomy of drift types. Supplementary sections include drift examples, reproducibility protocols, severity maps, and an anonymised drift blueprint for a large global bank. Findings indicate patterned divergence between model outputs and filed disclosures in high sensitivity domains. The purpose is to document model behavior rather than evaluate any institution.