Running summarizing tools on a new article

Abstract: The rapid adoption of large language models (LLMs) and generative artificial intelligence (AI) systems has brought AI hallucinations to the forefront of contemporary natural language processing research. AI hallucinations refer to statements made by the system which are either factually incorrect, illogical, or not substantiated by any credible sources but sound coherent and fluent on the face of it. These behaviors raise serious problems when using AI in high-stakes applications such as medicine, teaching, legal matters, finance, or science, where accurate facts are required. The present survey aims to provide an extensive overview of current techniques employed to detect, verify, mitigate, or correct hallucinations in generative AI models. Initially, we provide a taxonomy of different types of AI hallucinations. Subsequently, we examine major detection and verification paradigms, including retrieval-augmented generation (RAG), factuality evaluation metrics, automated fact-checking, natural language inference (NLI), knowledge-grounded reasoning, multi-model verification, and self-reflection-based approaches. The survey further reviews widely used benchmark datasets, evaluation frameworks, and emerging multi-layer verification architectures for improving AI trustworthiness. In addition, enabling technologies such as transformer-based language models, dense retrieval systems, knowledge graphs, credibility assessment mechanisms, and truth-scoring frameworks are analyzed from the perspective of reliable content verification. Lastly, we highlight important limitations of existing approaches along with potential future research directions, such as scaling verification, multilingual reasoning, real-time verification, and privacy-enhanced trusted artificial intelligence systems. In summary, this survey offers a comprehensive guide to researchers and developers aiming to design reliable, interpretable, and trustworthy generative AI models.