APPROACHES TO MINIMIZING HALLUCINATION EFFECTS IN GENERATIVE ARTIFICIAL INTELLIGENCE WHEN FORMING REGULATORY JUSTIFICATIONS

The growing use of generative artificial intelligence in compliance-sensitive environments has turned hallucination from a general quality problem into a matter of evidentiary reliability, traceability, and legal defensibility. This article examines how hallucination effects can be reduced when large language models are used to form regulatory justifications. The study aims to identify the forms of hallucination that are especially damaging in justification tasks, systematize technical mitigation approaches, and develop an implementation logic suited to audit-oriented deployment. The materials consist of ten recent academic sources published between 2023 and 2025, covering hallucination taxonomies, detection, retrieval-augmented generation, self-verification, factuality assessment, and legal use of generative AI. The methodological basis combines comparative analysis, source analysis, conceptual synthesis, and analytical generalization. The analytical part shows that reliable justification formation depends on a layered design that combines source-bounded retrieval, claim-level verification, uncertainty signaling, and mandatory human review. The proposed model has practical value for regulated digital workflows.