Health tech company Doximity has found that popular AI models make serious or potentially lethal errors in 19.5 per 1,000 answers to medication dosing questions.
The research, which tested nine widely used AI models, comes as major AI companies like OpenAI and Anthropic expand into healthcare, competing with specialized medical AI tools.
Doximity’s research highlights AI dosing errors
Doximity’s study, involving over 8,500 clinical dosing questions across 2,200 FDA-approved drugs, revealed that 2.3% of answers from general-use AI models contained incorrect doses.
The company’s new dose verification tool, DoseCheck, was 99.9% accurate in the same tests, according to Doximity.
Errors were more frequent with newly FDA-approved drugs, which had five times the odds of an incorrect dose due to limited real-world data.
The most common mistakes included failing to adjust doses for age, kidney function, or drug interactions, even when this information was provided.
Specialized AI tools aim to address accuracy concerns
As physician adoption of AI grows, with 81% of doctors now using it, accuracy remains a critical concern, especially in life-or-death decisions like medication dosing.
The company’s chief AI researcher, Dr. Louis-Antoine Mullie, noted that these failures aren’t limited to dosing but can impact other aspects of clinical care.
Among the errors identified, 3.4% were classified as potentially lethal, while 43% were significant and 13.2% minor.
Competition heats up in clinical AI space
With 165 health systems using its clinical AI platform, Doximity emphasizes the importance of accuracy and medical context in AI tools.
The findings show the need for rigorous testing and verification in clinical AI, particularly as these tools become more integrated into medical practice.
