Grok AI Misdiagnoses Medical Scans After Musk Urged Users to Upload Data
Elon Musk encouraged X users to upload medical images to Grok, xAI's AI chatbot, for diagnostic analysis. Physicians subsequently reported that Grok made significant diagnostic errors, including misidentifying tuberculosis and breast cysts, while experts raised concerns about both accuracy and privacy risks of using unprotected user data for AI training.
What happened
Elon Musk encouraged X users to upload medical images such as X-rays, MRIs, CT scans, and bone scans to Grok for analysis and diagnosis, stating "This is still early stage, but it is already quite accurate and will become extremely good." Medical images uploaded to Grok were to be used to train the chatbot.
Physicians who tested Grok reported significant diagnostic errors. Grok mistook a textbook case of tuberculosis for a herniated disk or spinal stenosis, and mistook a mammogram of a benign breast cyst for an image of testicles. The bot did successfully analyze some blood test results and identified breast cancer in some cases.
Reported impact
Experts raised concerns about both the accuracy of Grok's medical diagnoses and user privacy risks. Dr. Laura Heacock, associate professor at New York University Langone Health Department of Radiology, noted that "non-generative AI methods continue to outperform in medical imaging." Ryan Tarzy, CEO of Avandra Imaging, stated that using data directly input by users rather than from secure databases with de-identified patient data raises privacy concerns and appears designed to accelerate development at the expense of data security.
Index analysis
Musk claimed Grok had saved a man in Norway by diagnosing a problem his doctors failed to notice, though no medical evidence was presented. Musk stated he submitted his own MRI to Grok, saying "None of the doctors nor Grok found anything," but did not disclose why he had received the MRI.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.