Particle.news

Download on the App Store

Google’s Med-Gemini AI Still Lists Fictional Brain Part Months After Error Flagged

Experts warn that leaving the fabricated term in the research paper poses patient safety risks.

Overview

  • The 2024 research paper by Google’s Med-Gemini model reported an “old left basilar ganglia infarct,” inventing a non-existent brain structure.
  • Neurologist Dr. Bryan Moore exposed the mistake on LinkedIn in August, leading Google to silently update its blog post but not the original paper.
  • Google has characterized the error as a mere misspelling of “basal ganglia,” yet the published study continues to reference the fabricated “basilar ganglia.”
  • Google’s companion model, MedGemma, has also shown inconsistent diagnostic outputs, varying results for the same X-ray based on question phrasing.
  • Medical professionals caution that unchecked AI hallucinations risk causing misdiagnoses and underscore the need for stricter oversight in clinical AI deployments.