Microsoft Copilot provides inaccurate and potentially harmful medical advice
Researchers from Germany and Belgium found that Microsoft Copilot provided inaccurate medical advice, with 42% of responses about common U.S. prescription drugs potentially causing moderate to severe harm and 22% potentially causing severe harm or death if followed.
What happened
Researchers from Germany and Belgium tested Microsoft Copilot's medical advice by posing ten questions about side effects, contraindications, and dosage information for the 50 most-prescribed drugs in the United States, generating 500 responses.
When researchers evaluated a subset of approximately twenty responses against information from drugs.com, they found 54% of Copilot's answers were scientifically accurate, 24% did not conform to official medical knowledge, and 3% were completely incorrect.
Reported impact
According to the study, 42% of Copilot's medical responses could potentially cause moderate or severe harm to patients' health, while 22% of responses could cause severe harm or death if users followed the advice literally.
Index analysis
The study compared Copilot's medical responses to drugs.com, an encyclopedia with official accreditation as a reliable medical source, to assess accuracy and potential harm.
Governance controls
Researchers concluded that chatbots including Copilot, ChatGPT, and Gemini should not be used to provide reliable medical information, and cannot substitute for medical professionals.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.