Chatbots still get drug doses wrong often enough that one of the biggest physician platforms now checks every AI dose against plain rules before a doctor sees it.
Doximity added a deterministic check to its clinical AI tool. Each dosing claim the model makes is compared with FDA labeling and professional guidelines, instead of trusting the model's best guess.
In the company's own testing of nine widely used models on thousands of real physician-style dosing questions, at least 2.3% of dosing claims were wrong, and nearly half of those errors were serious or potentially lethal. The misses were the ones we'd expect: newer drugs, adjustments for kidney function and weight, units and decimal points, and look-alike drug names.
At the bedside, that's the ICU med list: renally cleared antibiotics, weight-based drips, a drug approved last year. A rule-based check is the right design. It still doesn't know your patient's creatinine trend. Verify the dose. The physician still decides.
The limit: this is a press release, not a peer-reviewed study. The accuracy figures are the company's own and haven't been independently tested.
Source: Doximity press release via Business Wire, Oct 29, 2026. Vendor-reported internal testing, not peer reviewed. www.businesswire.com/news/home/20261029499969/en/