Google Med-Gemini Fabricates Nonexistent Brain Structure

Google's Med-Gemini hallucinated an 'old left basilar ganglia infarct' in a radiology case study -- a diagnosis referencing a brain structure that simply does not exist in human anatomy. The error surfaced publicly after a board-certified neurologist reviewed the published research and flagged it to The Verge. Google's response was to quietly edit the blog post describing the research, with no public acknowledgment of the correction. The original peer-reviewed paper, which circulates in the scientific record, remains unchanged. Essentially: Google published AI medical research containing a clinical hallucination, corrected the marketing layer, and left the scientific layer wrong. - Med-Gemini was trained on radiology transcripts that may themselves contain transcription errors, compounding the hallucination risk at the source data level. - The neurologist who caught the error says the case illustrates the clinical danger of deploying AI trained on mis-transcribed medical text without rigorous verification. - The uncorrected paper continues to be indexed and cited, meaning the fabricated anatomy persists in the literature. The incident draws a sharp line between AI performance on benchmarks and AI safety in clinical deployment, where a confident wrong answer about neuroanatomy is not a benchmark failure but a patient risk.