Google Healthcare AI Model Med‐Gemini Allegedly Produces Non‐Existent 'Basilar Ganglia' Term in Published Output

In May 2024, Google released Med-Gemini, a suite of healthcare AI models designed to interpret medical scans, generate radiology reports, and analyze health records. The models were trained on de-identified medical data and promoted as capable of interpreting complex 3D scans and generating state-of-the-art radiology reports. In a research paper and blog post demonstrating Med-Gemini's capabilities, the AI diagnosed an 'old left basilar ganglia infarct' in a head CT scan example. However, 'basilar ganglia' is not a real anatomical structure - it incorrectly combines 'basal ganglia' (brain region controlling motor function) with 'basilar artery' (blood vessel supplying the brainstem). Neurologist Bryan Moore caught this error and flagged it to Google. The company initially quietly edited the blog post to correct 'basilar' to 'basal' without acknowledgment, then later reverted to show the original error with a caption calling it a 'common mis-transcription.' The research paper, which had over 50 authors including medical professionals, still contains the uncorrected error. Google characterized the mistake as a simple misspelling, but medical experts consider it a dangerous hallucination that could lead to misdiagnosis and inappropriate treatment if used in clinical settings. Med-Gemini has since moved to a trusted tester program, suggesting pilot testing in real medical scenarios.