Medical Study Identifies Incorrect Prescription Doses Generated by AI Models

A study by medical platform Doximity revealed that leading artificial intelligence models regularly make hazardous mistakes when writing drug prescriptions. Researchers evaluated nine AI models on thousands of medical scenarios, finding that over two percent of the calculated dosages were incorrect. The software frequently ignored essential patient details provided in the prompts, such as weight, age, and kidney function.

Although a small percentage of mistakes might seem manageable in standard software applications, any error in medicine can lead to severe harm. The analysis showed that over forty percent of the flaws were clinically significant, including potentially fatal mix-ups involving swapped drug names or incorrect measurement units. Newly released drugs were especially vulnerable, triggering errors at five times the rate of older medications.

To reduce these hazards, Doximity created a verification system that automatically cross-checks AI outputs against an established database of official drug labeling. While this provides a safety net for doctors using specialized clinical software, it leaves a major question about public safety. It is still uncertain how general consumers using popular web chatbots at home will be protected from convincingly written but dangerous medical advice.