Hallucinations in generative LLMs for radiology: a comprehensive review

Hallucinations generated by large language models (LLMs) represent an emerging safety top event in radiology, where subtle interpretive errors can rapidly propagate into clinical decisions. Although these systems demonstrate remarkable linguistic fluency, their underlying probabilistic architecture is, in some cases, not aligned with the demands of diagnostic intelligence, factual reliability, or safety standards. This work presents a comprehensive review describing the pathways that lead to hallucinations in radiology domain generative LLMs, mapping the risks derived from model architecture, data quality, black-box opacity, and automation bias. It also identifies the preventive strategies required to control these risks, including legal and governance safeguards, rigorous data curation protocols, explainability and simulation frameworks, and structured human oversight. This review shows how hallucinations may distort diagnostic reasoning, generate fictitious clinical data, and precipitate erroneous management decisions with potential patient harm, underscoring the need for controlled deployment strategies before integrating generative AI into radiological workflows.