Hallucination In Large Language Models: A Review Of Causes, Taxonomy, Detection Methods And Mitigation Strategies

Large Language Models (LLMs) generate human-like text, answer questions and summarize information, but they sometimes produce incorrect, misleading or fabricated content while presenting it as fact. This phenomenon, known as hallucination, reduces the accuracy, trustworthiness and safety of generative AI systems, particularly in education, healthcare, law and finance. This paper reviews LLM hallucination in a structured way. It explains causes arising from training data, model architecture and training objectives, decoding and missing context, and presents a taxonomy that distinguishes factuality from faithfulness errors. Benchmarks that measure truthfulness and hallucination rates are discussed. Detection methods are reviewed, including self-consistency checks across sampled answers, uncertainty measures such as semantic entropy and fine-grained atomic fact verification. Mitigation strategies are examined across the model life cycle: data curation, alignment with human feedback, retrieval-augmented generation, prompting for verification and abstention when uncertain. A layered hallucination-aware LLM application architecture is proposed, along with research gaps, objectives and a comparison of methods. The review concludes that hallucination cannot be eliminated entirely but can be detected, reduced and communicated so that users can rely on LLM outputs appropriately.