Doximity benchmark found nine unnamed popular AI models made potentially lethal medication dosing errors
Doximity tested nine widely used AI models against more than 8,500 clinical dosing questions covering over 2,200 FDA-approved drugs. It found about 19.5 dangerous or potentially lethal dosing errors per 1,000 answers, with 3.4% of errors classed as potentially lethal. The models often failed on patient-specific dose adjustments, such as age, kidney function and drug-drug interactions, even when that information was in the prompt, and newly approved drugs carried five times the error odds. The nine models were not named, and the results were released alongside Doximity's own DoseCheck product.
What the AI did
In a test run by Doximity, nine widely used AI models, which have not been named, answered more than 8,500 clinical dosing questions and gave about 19.5 dangerous or potentially lethal dosing errors per 1,000 answers. The models often got patient-specific dose adjustments wrong, such as those for age, kidney function and drug-drug interactions, even when that information was in the question, and errors were five times more likely for newly approved drugs.
First reported October 29, 2026 · Added to the register November 8, 2026 · 1 source
- When it happened
- Not stated in the sources
- First reported
- October 29, 2026
What this means for you
Could this affect you?
If you use general-purpose AI models for medication dosing questions. Hospitals, clinicians and patients who rely on general-purpose AI models for dosing questions are exposed, especially for newly approved drugs and doses that must be adjusted for the individual patient.
What to check
- Require pharmacist or clinician review before any AI dosing recommendation reaches a patient.
- Check whether any AI dosing tool in use has been independently validated.
- Check whether any AI dosing tool in use needs FDA clearance.
Areas of your AI programme this touches
Every fact and its source (8)
- Results releasedOct. 29“The results, released Thursday, Oct. 29 alongside the company's new clinical dose-verification product DoseCheck”[1]
- Test scopeNine models, 8,500+ dosing questions, 2,200+ drugs“tested nine widely used models against more than 8,500 clinical dosing questions spanning over 2,200 FDA-approved drugs”[1]
- Dangerous error rateAbout 19.5 per 1,000 answers“approximately 19.5 dangerous or potentially lethal dosing errors per 1,000 answers”[1]
- Share of errors potentially lethal3.4%“3.4% were classified as potentially lethal”[1]
- Benchmark run byDoximity“Doximity, the clinical networking platform used by more than 165 health systems”[1]
- Newly approved drugsFive times the dosing error odds“Newly approved medications carried five times the dosing error odds”[1]
- Models testedNot named“The nine models tested are not named in the published summary.”[1]
- Main failure modePatient-specific dose adjustments even when info provided“AI repeatedly failed on patient-specific dose adjustments — age, kidney function, and drug-drug interactions — even when that information was explicitly provided in the prompt”[1]
Sources
- Popular AI Models Make Potentially Lethal Dosing Errors: Doximity Benchmarktechtimes.com · October 30, 2026
How this record is classified. Severity N: a legal action or test finding, not a harm. OECD level: hazard, an event that could plausibly have led to harm. Evidence: Confirmed, meaning a company, official or court statement.
The facts in At a glance and in Every fact and its source are quoted from the sources listed. The summary, What the AI did and What this means for you are written from those sources and checked against them automatically before publication. Records reflect what has been disclosed, not everything that has happened. Incidents are found through news feeds and the GDELT Project.
