Reframing Hallucination in Large Language Models: A Lifecycle-Based, Mechanism-Aligned, and Phenomenon-Consistent Definition

While recent advances in Large Language Models’ reasoning capabilities have improved their performance across many tasks, the fundamental challenge of hallucination, a form of unexpected generation, persists. Researchers are already attempting to develop methods to detect and mitigate hallucination, even though the term itself lacks a unified definition and is applied inconsistently across the literature. This incongruity underscores the need for a clear and conceptually grounded definition, which is the prerequisite for pursuing effective solutions. We reviewed 76 definitions of hallucination from AI-related literature published between 2023 and 2025. Our analysis revealed substantial overlaps and ambiguities in the definitional boundaries between hallucination and other related concepts, leading to inconsistent usage across the literature. The current definitional vocabulary also remains largely focused on observable properties, such as factuality and faithfulness, rather than the underlying mechanisms that give rise to these outputs. To address these limitations, we propose a Lifecycle-Based, Mechanism-Aligned, and Phenomenon-Consistent definition of hallucination that integrates both the generative mechanisms and their observable manifestations across the stages of data collection, processing, and quality assurance, through model development, to deployment. This reframed definition provides a conceptually coherent foundation for hallucination evaluation and mitigation, offering future guidance for developing benchmarks, designing diagnostic tools, and implementing mitigation strategies.