Doximity Puts Medication Dosing Errors at the Center of the AI Debate
When clinicians ask a language model about a medication dose, a wrong answer can become a bedside hazard in seconds. Doximity’s latest research places that risk in stark relief: among nine widely used AI models, about 19.5 per 1,000 answers contained an error judged serious or potentially lethal.
The company released the analysis alongside a new dose verification tool, part of a broader effort to separate clinical-grade software from general-purpose AI systems moving into healthcare. In practical terms, the findings speak to one of the most sensitive use cases in medicine. Dosing is not an abstract knowledge task; it sits at the intersection of renal function, age, comorbidities, drug interactions, and narrow therapeutic margins.
Freshly Approved Drugs Carried More Risk
Doximity tested models with more than 8,500 clinical dosing questions spanning more than 2,200 FDA-approved drugs. The questions were designed to examine how often the models gave incorrect answers and how severe those mistakes were. Among the errors identified, 13.2% were minor, 43% were significant, and 3.4% were potentially lethal.
A particularly concerning pattern emerged around newly approved medications. Drugs recently cleared by the FDA had five times the odds of an error, a gap Doximity linked to the simple reality that newer agents have less real-world safety data and less embedded clinical memory than older therapies. In daily practice, that matters because clinicians often rely on accumulated experience as much as monographs and labels when choosing a dose for a patient whose physiology does not fit a textbook profile.
One sentence captures the problem plainly: the models were often wrong even when the relevant clinical context was included in the prompt.
The most common failure involved missing dose adjustments for age, kidney function, or drug interactions. Those are not peripheral details. They are the core variables that determine whether a dose is therapeutic, excessive, or toxic. A model that can repeat a standard dose but fails to adjust for reduced renal clearance or a competing medication may sound confident while steering a clinician toward an unsafe decision.
DoseCheck Is Positioned as a Safety Layer, Not a Replacement
Doximity says its new verification tool, DoseCheck, answered dosing questions correctly 99.9% of the time in testing. The company frames it as an added safety net inside Doximity Ask, its AI clinical assistant and medical search engine for clinicians. The design logic is straightforward: if a general model can draft an answer and a verification layer can check it against dosing rules, the combination may reduce the chance of a dangerous miss.
That architecture reflects a growing shift in healthcare AI. General-use models may be broad and fast, but dosing requires precision, current knowledge, and patient-specific context. Errors become more consequential as the workflow moves from documentation support to prescribing support. A missed renal adjustment can turn a routine prescription into an adverse drug event, and a failure to recognize an interaction can compound toxicity across an entire treatment course.
Dr. Louis-Antoine Mullie, Doximity’s chief AI researcher, said the failure modes seen in dosing are not unique to that task. Outdated knowledge and missing patient context can affect other aspects of clinical care as well. His point goes beyond medication selection. Any clinical AI system that lacks a reliable way to integrate context risks drifting from helpful assistant to unsafe shortcut.
Competition Is Tightening as Big AI Firms Enter Healthcare
The research arrives as major AI developers expand into medicine and compete directly with specialized vendors. Anthropic and OpenAI have both moved into healthcare-facing products, increasing pressure on companies like Doximity to prove that clinical specialization delivers measurable safety advantages. In this market, accuracy is no longer an abstract differentiator. It is the product.
Physician adoption of AI is rising quickly. More than 81% of physicians report using AI in practice, more than double the rate in 2023, according to the American Medical Association. Even so, use does not equal trust, and the stakes in dosing are high enough to keep skepticism alive. Clinicians need a tool that can be checked, audited, and verified, not just one that sounds knowledgeable in natural language.
Research on both sides of the debate is already shaping how buyers interpret these tools. Physician researchers at Stanford University School of Medicine and Harvard Medical School found that clinically specialized AI tools outperformed general-purpose models in some assessments, while another research team later reported that general-purpose AI could outperform specialized clinical AI tools in other settings. The result is a fragmented evidence base, and that makes head-to-head testing in real clinical tasks especially important.
Why Dosing Errors Cut Deepest in Clinical Practice
Medication dosing sits in a uniquely unforgiving corner of medicine because the right answer changes with the patient. The same drug can require a lower dose in kidney impairment, a different regimen in older adults, or avoidance altogether when interacting therapies are present. AI systems trained on broad text patterns can reproduce familiar dosing language without truly modeling the physiologic reasons behind it.
That gap explains why a seemingly small error rate can still produce meaningful risk. Even a few dangerous answers per thousand can matter if the tool is embedded in workflows used repeatedly across large clinical networks. Doximity says its platform is used by more than 165 health systems, which gives any improvement in verification a potentially wide clinical footprint.
The central question now is not whether AI belongs in healthcare. It already does. The harder issue is which layer of the workflow should be trusted, when, and with what guardrails. Doximity’s study argues that the answer cannot be a free pass for general-purpose models, especially when the output concerns dosing for drugs with narrow safety margins.
The company’s own data also leave open a basic operational challenge: verification tools are only as useful as their coverage, maintenance, and integration into clinician behavior. A 99.9% correct verification result in testing is an impressive claim, but bedside safety depends on how often clinicians actually use the checker, how it handles edge cases, and whether it stays current as new drugs and new dosing rules arrive.