Your Chief AI Officer · Friday, October 30, 2026 · Act today
Doximity research finds general AI models err on medication doses
Reported2 sourcesFor patients and employees
Clinical AI company Doximity released research showing popular general-use AI models get medication doses wrong, some dangerously, and that its own dose verification tool performed better, Healthcare Dive reported. Separately, a family has sued OpenAI over ChatGPT advice on an epilepsy drug taper, Medical Economics reported.
Doximity released research on October 29 testing nine widely used AI models with more than 8,500 clinical dosing questions covering more than 2,200 FDA-approved drugs. About 19.5 per 1,000 answers contained an error that was serious or potentially lethal, and the models got doses wrong in at least 2.3% of answers, Healthcare Dive reported. Of the errors detected, 13.2% were minor, 43% significant and 3.4% potentially lethal. Newly approved drugs carried five times the odds of an error. The most common failure was not adjusting doses for age, kidney function or drug interactions, even when the question supplied that information. Doximity says its new DoseCheck tool was correct 99.9% of the time.
The same day, Medical Economics reported that the parents of Bowen Wan, a 31-year-old with severe epilepsy, have sued OpenAI. They allege ChatGPT helped her taper off lamotrigine from 400 mg a day to zero within weeks without her physicians' knowledge before her death in February. OpenAI said the case involved an earlier version that is no longer available. It pulled GPT-4o from its consumer app Feb. 13.
Why it matters: clinician use is already widespread. More than 81% of physicians report using AI in their practices, according to the American Medical Association as cited by Healthcare Dive. Patients are also bringing chatbot advice into care, sometimes invisibly. Wan's neurologist said he did not know she had used ChatGPT to draft messages to him.
What the evidence shows and does not: the dosing study comes from a vendor comparing its own product, launched alongside that product, and has not been described as peer reviewed. Published research points both ways. Stanford and Harvard researchers found specialized clinical tools outperformed general models, while another team this summer found the reverse. The lawsuit is an allegation, not a finding. Still, the failure pattern Doximity describes, missed adjustments for patient context, is a concrete risk worth governing whichever tool is in use.
The move: the Chief AI Officer and CMO should publish a short list of tools approved for dosing questions and require pharmacist or reference verification for any AI-suggested dose, with extra scrutiny for newly approved drugs. Nursing and pharmacy leaders can fold this into existing double-check workflows rather than adding a new step. Compliance should document the policy. Clinicians and front-line teams can routinely ask patients whether they have used a chatbot about their medications, which protects patients and gives staff an earlier chance to correct course.
Do this week. Issue clear guidance to clinicians on which AI tools are approved for dosing questions and require pharmacist or reference verification for any AI-suggested dose.
Run this
The prompt to act on this item, for your own model.
Paste it into Claude, ChatGPT or Copilot and fill the bracketed slots with your own numbers. The output is a draft for your team to check, built only from the sources on this page. Never paste patient data. How these prompts are written.
You are an operations advisor to the Chief AI Officer and Chief Medical Officer of [your organization], a medical, health and wellness organization. On October 29, Doximity released research testing nine widely used AI models with more than 8,500 clinical dosing questions covering more than 2,200 FDA-approved drugs, Healthcare Dive reported. The models got doses wrong in at least 2.3% of answers, and about 19.5 per 1,000 answers contained an error that was serious or potentially lethal. Of errors detected, 13.2% were minor, 43% significant and 3.4% potentially lethal. Newly approved drugs carried five times the odds of an error. The most common failure was not adjusting doses for age, kidney function or drug interactions, even when the question supplied that information. Doximity says its DoseCheck tool was correct 99.9% of the time, but this is a vendor comparing its own product and has not been described as peer reviewed. Separately, Medical Economics reported that a family has sued OpenAI over ChatGPT advice on an epilepsy drug taper. That is an allegation, not a finding. Healthcare Dive cites the American Medical Association: more than 81% of physicians report using AI in their practices. Task: draft clinician guidance we can issue this week that names which AI tools are approved for dosing questions and requires pharmacist or reference verification for any AI-suggested dose, with extra scrutiny for newly approved drugs. Fold verification into our existing double-check workflow rather than adding a new step. Output: a draft workflow with owners and checks, then a one-page memo for the executive team with these sections: Approved tools, Verification rule, Workflow and owners, Asking patients about chatbot use, Documentation for compliance, Open questions. Inputs: [AI tools currently approved or in use], [your EHR], [your current medication double-check workflow], [your pharmacy or drug reference resource], [the number of clinicians the guidance covers]. Use only the facts above and the inputs I fill in. Cite Healthcare Dive and Medical Economics by name. Flag anything uncertain or that must be verified. Use no patient data. Nothing here is medical advice. Label the result a draft for our team to check.
- The claim.
- Doximity says its dose verification tool outperformed general-use models, which it describes as rife with errors.
- The evidence.
- The research comes from a vendor comparing its own product and has not been described as peer reviewed; the lawsuit is an allegation, not a finding.
By seat
For the Chief AI Officer. Confirm the approved-tools list covers dosing use cases and blocks or flags unapproved chatbots on clinical devices.
For the Chief Medical Officer and physicians. Remind prescribers that general chatbots are not dose references and route questions through verified sources.
For the CNO and nursing. Reinforce that nurses should not act on AI-generated dose information without the standard verification step.
For the Compliance, privacy and legal. Document the policy on clinician use of consumer AI so the organization can show reasonable controls.
Disclosure: Ecofy orchestrates OpenAI for clients and says so on every item that names it. The item was selected and checked by the same rules as every other.
Checked before publishing. 2 independent outlets.
Sources
Drafted by Lumen from the sources above, scored by a second pass on six criteria, then checked by rule against those sources (every figure, two independent outlets or a primary source) and published only when it passes. A person can pull any item, and corrections are published here. Apolitical, always. Nothing here is medical advice. How Your Chief AI Officer is made.
Also in this brief
- Epic's AI pricing now moves with usage, making next year's bill hard to forecast
- Ten health systems each passed 1 million AI-documented encounters with Microsoft
- AMA releases framework assigning liability when clinical AI harms patients
- Oracle Health breach count rises to nearly 20 million as systems move to Epic
- HHS finalizes rule cutting health IT certification requirements
- National Coordinator Keane expects HTI-6 to address wearable and patient-generated data