Dive Brief:
- When nine widely used AI models were prompted with questions about medication dosing, about 19.5 per 1,000 answers contained an error that was serious or potentially lethal, according to new research from health tech company Doximity.
- Doximity’s research, which was launched in tandem with the company’s new dose verification tool, claims other popular AI models got doses wrong in at least 2.3% of answers.
- It’s the company’s latest attempt to differentiate itself from major AI companies like OpenAI and Anthropic that are expanding their presence in healthcare.
Dive Insight:
As the appetite for clinical AI grows, companies offering specialized medical AI tools are trying to prove that they have accuracy and medical context on their side in a way that general-use AI tools do not.
The research comes at a time when major AI companies like Anthropic and OpenAI are entering the healthcare space, presenting steep competition for industry-specific tools.
Doximity, which claims its clinical AI platform is used by more than 165 health systems, says it tested a handful of large language models released in 2026 by asking them more than 8,500 clinical dosing questions covering more than 2,200 FDA-approved drugs. They checked how often the answers contained errors and how severe those errors were.
Drugs that were freshly approved by the FDA had five times the odds of an error, largely because they are newer to the market and have less real-world safety data compared to established drugs, according to Doximity.
Of the errors detected, 13.2% were minor, 43% were significant and 3.4% were potentially lethal. The most common mistake was failing to adjust doses for age, kidney function or drug interactions, even when that information was given in the question, the report states.
Researchers then tested DoseCheck, Doximity’s new verification tool, and claimed that its dosing findings were correct 99.9% of the time.
The company frames the tool as an additional safety net within Doximity Ask, its AI clinical assistant and medical search engine for clinicians.
“The failure modes we saw aren’t unique to dosing,” Dr. Louis-Antoine Mullie, chief AI researcher at Doximity, said. “Those same weaknesses, alongside outdated knowledge and missing patient context, can show up in other aspects of clinical care.”
Physician adoption of AI is growing rapidly — more than 81% of physicians report using AI in their practices, according to the American Medical Association, which is more than double the rate in 2023.
But accuracy remains the key point of contention in debates about implementing AI tools in healthcare, especially when they involve life-or-death decisions like dosing and drug prescribing. AI companies have the tall task of winning over skeptical clinicians who want to ensure that any tool used in their practice will not harm patients.
Companies that specialize in clinical-grade AI tools say they are best equipped to handle the industry’s concerns, and research has been released on either side of the debate. For example, physician researchers from Stanford University School of Medicine and Harvard Medical School found that clinically specialized AI tools consistently outperformed general-purpose AI models. Doximity celebrated the results of that study earlier this month.
But another research team published findings this summer showing that general-purpose AI can outperform specialized clinical AI tools.
