AI Hallucinations — Analysis of Causes and Targeted Solutions

Abstract

This paper reviews the problem of ‘AI hallucinations’ that occurs in generative models for open-domain and knowledge-intensive tasks, and proposes transforming it into a system engineering process for visual evaluation and correction. This study constructs an ‘Evidence–Evidence-Generation-Verification-Abandon’ closed loop: evidence is pre-placed using Retrieval-Augmented Generation (RAG) with sentence-level evidence binding; post-generation self-verification is achieved via Chain-of-Verification (CoVe) and consistency sampling; and a risk upper bound is provided through a calibratable refusal to answer, establishing an intuitive report based on FACTSCORE. Implementations and case studies on Chinese questionanswering and long-text generation tasks demonstrate that this closed loop can reduce high-risk outputs while maintaining usability. This research facilitates the more convenient integration of handling AI hallucinations and reduces their occurrence. The newly established closed loop can significantly lower the error rate of large language models. It innovatively combines statistical knowledge to calculate the confidence interval of AI responses, providing a mathematical model for the occurrence of AI hallucinations. This offers a new approach to addressing the issue of AI hallucinations.