A serious error has been discovered in Google’s latest healthcare development—the Med-Gemini medical AI model—: the AI “invented” a nonexistent anatomical structure. A research report on a brain CT scan mentioned an “infarction of the left basal ganglia,” even though such a region simply does not exist in the human brain.
This passage caused concern within the community, since, as it turned out, Med-Gemini incorrectly combined two real terms—basal ganglia and basilar artery—to create a fictitious one: basilar ganglia. Neither the authors of the research paper nor the Google team, which published the article on its official blog back in 2024, noticed the error.
It wasn’t until some time later that the problem was discovered by neurologist and AI researcher Brian Moore, who reported it to Google via LinkedIn. In response, the company privately edited the blog post, replacing “basilar” with “basal,” but did not correct the scientific publication itself. In a comment to The Verge, the error was described as a “simple typo.”
However, not all experts share this view. Maulin Shah, Chief Medical Information Officer at Providence Health (which operates more than 50 hospitals in the United States), emphasized that even a minor terminological error can have serious consequences in clinical practice.
Google positions the Med-Gemini system as a tool for the automated generation of medical reports, analysis of medical records, and risk prediction. The company presented it as a true breakthrough in the field of medical AI. But the case of the made-up term revealed yet another problem—AI’s tendency to “hallucinate,” that is, to invent plausible but false facts.
“We can’t expect doctors to always proofread these reports by hand. And if an error makes its way into the medical database, it could be replicated by other models or systems,” Shah noted.
This is not the first instance of problematic behavior by Google’s medical AI systems. Another model, MedGemma, produced different results for the same X-ray images depending on how the query was phrased: when a doctor asked a detailed question, the model detected abnormalities, but when the question was phrased more simply, it did not.
Professor Jonathan Chen of Stanford School of Medicine noted that one of the main dangers is that AI provides incorrect answers with absolute certainty. This creates the illusion of reliability and can influence the trust of professionals, who gradually become accustomed to relying on such systems.
Dr. Judy Gichoya of Emory University shares this view. According to her, Med-Gemini has repeatedly generated incorrect medical conclusions in the test environment. She is particularly concerned that the models are unable to admit their uncertainty—they never say “I don’t know,” which is of fundamental importance in medicine.
Google officially acknowledged the issue only in a brief comment under a LinkedIn post:
“Thank you for bringing this to our attention! We’ve updated the image on the blog to reflect the correct result.”However, experts emphasize that this isn’t just a single unfortunate mistake.
“What’s most alarming isn’t that AI sometimes makes mistakes, but how convincingly it does so,” Chen concludes.