AI-Generated Radiology Reports: Opportunities For Efficiency, Diagnostic Support, And Risks To Patient Safety
Description
Generative artificial intelligence (AI) and large language models (LLMs) are increasingly being embedded into radiology reporting workflows, promising to ease a mismatch between imaging volumes that grow 8-12% annually and a radiology workforce that expands by only 2-3% per year. This paper reviews the current evidence on AI-generated radiology reports across three dimensions: efficiency, diagnostic support, and patient safety. Across nine recent clinical studies, AI-assisted drafting reduced radiologist reporting time by 7.8% to 43.8%, with the largest gains observed for structured, high-volume examinations such as outpatient CT. As of March 2026, radiology accounted for 1,163 of 1,524 (76.3%) FDA-cleared artificial-intelligence medical devices, with representative diagnostic-support tools reporting sensitivity and specificity above 90% for intracranial haemorrhage, pulmonary embolism, large-vessel-occlusion stroke, and fracture detection. However, systematic evaluations also document clinically meaningful hallucination and omission rates — up to 5.5% for major omissions and 1.0% for malignant hallucinations in some report categories — alongside documented automation bias and unresolved medico-legal liability when radiologists over-rely on AI output. We synthesise this evidence into a risk-mitigation framework spanning mandatory human-in-the-loop review, hallucination-detection tooling, transparent error-rate disclosure, and regulatory post-market surveillance. We conclude that AI-generated radiology reports offer substantial, reproducible efficiency and diagnostic gains, but that these gains are only safely realisable within governance structures that actively guard against automation bias and hold hallucination rates to clinically negligible levels.