Doximity adds DoseCheck after finding dosing errors in AI answers
Doximity has introduced DoseCheck, a tool that compares AI-generated medication doses with structured drug information. Its launch accompanies company research reporting errors in at least 2.3 percent of evaluated dosages across nine language models. The tool extracts dosing details from answers and links checks to source information; the developer’s evaluation does not establish independently measured patient outcomes.
Artificial Intelligence··Evening
DoseCheck compares generated doses with drug information
Doximity, the company behind the clinical assistant and medical search service Doximity Ask, has introduced DoseCheck to verify AI-generated medication doses. Announced on October 29, the free additional checking layer extracts the drug, dose, route of administration, frequency and relevant indication from an answer. It then compares those details with structured drug information, including labeling from the US Food and Drug Administration. A badge beside a matching dose opens the underlying information for the clinician to inspect.[1], [2]
The comparison is deterministic: the same dosage claim checked against the same database gives the same result. Doximity describes using structured information rather than asking another language model to judge the first answer. Its RxIndex resource contains approximately 1,900 drug monographs and 36,000 dosing regimens, organized by indication, patient population and route. The company says the checking layer preserves the need for clinical judgment.[1]
Company evaluation found errors across nine models
The launch accompanies Doximity’s own evaluation of nine widely used language models released in 2026. More than 8,000 questions modeled on physician queries covered 2,204 medicines. Across 261,508 evaluated dosages, the company reports that at least 2.3 percent were wrong, calling that a conservative lower bound. Its separate severity assessment identified nearly 20 serious dosing errors per 1,000 answers. The dosage count and answer count describe different evaluation units; neither figure is a measurement of injury rates among patients receiving treatment.[1]
Patient details were often missed despite being supplied
The most common problem was failure to adjust a dose for patient characteristics already supplied in the question, such as age, kidney function or interactions between medicines. Doximity also describes errors involving units, decimal placement and medication matching. Recently approved medicines were especially vulnerable in the company’s analysis. Healthcare Dive’s coverage places the launch alongside growing competition from general-purpose AI companies entering healthcare. The evaluation comes from DoseCheck’s developer; independent reporting of those findings supplies no independent replication of the product or patient-outcome assessment.[1], [2]
A sensitivity analysis using a separate prescribing-error severity scale identified approximately 30.5 errors per 1,000 answers in categories encompassing harm or the need for intervention. The company’s evaluation describes different types of mistakes, including incorrect units, decimal placement and drug matching. These assessments classify the severity of generated answers, rather than following a group of treated patients.[1]
Related columns
For more information on this topic, you can read the related columns.
