Doximity Launches DoseCheck — AI Dosage Verification Based on Rules, Not a Second Model
Doximity added a dosage verification layer to Doximity Ask based on the FDA database and clinical guidelines. In tests on nine models, at least 2.3% of dosage statements were incorrect, and 43.8% of the errors found were serious or potentially fatal.
On October 30, 2026, Doximity announced the launch of DoseCheck — a safety feature within Doximity Ask. It verifies dosage recommendations generated by AI (artificial intelligence) and is intended to catch unsafe or incorrect doses.
The key difference is how the verification works. Doximity does not task one large language model (LLM — a model trained to predict text) with checking another. Instead, it adds a deterministic rules layer: a dosage statement is checked against structured data, including FDA (U.S. Food and Drug Administration) labeling and professional guidelines. Deterministic here means that given the same input data, the answer will be the same, without probabilistic variation.
The company claims that this verification raises accuracy above 99%. This figure comes from the press release; there is no independent verification in the note.
Before launch, Doximity ran nine common AI models through more than 8,000 dosage questions compiled from real physician queries. At least 2.3% of dosage statements were incorrect. Of the errors found, 43.8% were serious or potentially fatal. Common sources of error: dosing for new drugs, ignoring patient factors (kidney function, age, weight, medical history), and confusion with units, decimal points, or similar drug names.
"AI is probabilistic. Drug dosing cannot be," the company quotes Louis-Antoine Mullie, chief AI researcher at Doximity. In his phrasing, DoseCheck does not ask a language model to "check its own homework," but rather checks statements against a structured database so that the physician gets the same verified answer every time.
In practice, this is another layer between the model's answer and the physician's decision. The check does not replace clinical judgment: it verifies text against a database, not evaluates the patient. For new drugs and complex cases, the value depends on how fully the database covers the specific situation.
What remains unknown: the full list of sources and the update frequency of the structured database, as well as exactly how a statement is verified if a recommendation mixes several drugs. There is no data on independent validation of the claimed 99% in the note.