Doximity’s DoseCheck launch on 29 October 2026 puts a concrete question before clinicians using medical AI: how should a medication claim be checked before anyone relies on it? The company describes a rules-based checking layer within Doximity Ask and reports accuracy above 99%. Its announcement makes verification an explicit part of the answer interface. Doximity’s launch release explains the proposed approach.
Our DoseCheck launch explainer covers the product announcement. The accompanying benchmark deserves attention because it separates out a consequential component of an AI answer. For readers assessing the evidence, the useful next step is to identify exactly what was counted, what was verified and which clinical conclusions the results support. A percentage becomes much more informative when its denominator and intended use travel with it.
Start with the unit being counted
Doximity’s research summary describes testing nine models across 2,204 drugs. Its reported quantities use different units:
| Reported quantity | What the denominator represents |
|---|---|
| More than 8,000 questions | Inputs presented to the models |
| 261,508 evaluated dosages | Individual dosage claims assessed |
| At least 2.3% incorrect | A claim-level error estimate |
| Nearly 20 per 1,000 answers | An answer-level serious-error estimate |
One answer can contain several medication claims. Counting each creates a useful picture of how often individual statements fail, but it does not directly tell us the chance that a whole answer contains a problem. Conversely, an answer-level measure can tell us whether any qualifying problem occurred while concealing how many separate statements were affected.
Neither denominator is a prescription issued to a patient. It would therefore be inaccurate to convert this benchmark into a forecast of erroneous prescriptions per thousand consultations. That would require evidence about what clinicians asked, which suggestions they accepted, what they independently checked and what ultimately entered the prescribing workflow.
What “more than 99%” needs alongside it
The company’s release reports accuracy above 99%; its illustration also describes an error reduction above 99%. These metrics answer different questions. The first concerns the fraction classified as correct; the second compares an error rate with a baseline. Readers should preserve the distinction when discussing the announcement. Source: Doximity’s company release.
For a verification system, several measurements would be useful together. How often is a verified claim actually correct? How many incorrect claims are caught? How many correct claims are challenged? What proportion of all claims can the system assess? A high result on one measurement does not determine the others.
Imagine two hypothetical tools with equally reliable positive verification results. One can assess almost every claim in its intended setting; the other assesses a narrow subset and leaves the remainder unresolved. Both could provide value, but their badges would carry different practical implications. This is why an explicit “unable to verify” category can be informative rather than a failure to produce an answer.
No single headline percentage establishes performance across every medicine, patient group or clinical setting. The useful publication would show the eligible population, unresolved cases, reference standard and uncertainty alongside each result. Clinicians can then judge whether the tested task resembles their own.
Follow the whole verification sequence
Doximity describes extracting the medicine, dose, route, frequency and relevant indication from an answer, then comparing that claim with a structured database. A matching claim receives a verification badge linked to its source. The company says the comparison is deterministic. Source: Doximity’s research summary.
That architecture makes the intermediate steps important. A repeatable comparison can still receive an incorrectly extracted claim. An entry may match the intended drug but omit a qualifying statement elsewhere in the answer. A source can be correctly represented while the patient’s circumstances fall outside the entry’s applicability. These are possible evaluation questions for any such system, rather than findings that this benchmark demonstrates occurred.
Testing should therefore examine the entire path from original question to displayed result. Reviewers need to know whether the system identified every relevant claim, retained necessary context and clearly communicated unresolved checks. Testing the matching engine alone would leave these surrounding steps unanswered.
A source link also has a specific role: it allows inspection. It does not, by itself, demonstrate that the generated statement faithfully represents the source or applies to the situation. The interface should help the reader find the pertinent passage and recognise the limits of what was checked.
Interpret severity as a separate question
A benchmark can assess the potential consequences of an incorrect answer without observing those consequences in a patient. That distinction matters when translating terms such as “serious” into a discussion of safety.
The NCC MERP Medication Error Index, referenced by Doximity, categorises medication errors with attention to whether an error reached a patient and the resulting harm. Applying a severity framework to generated answers therefore needs a clear explanation of the hypothetical assumptions used.
An evaluator might reasonably judge a mistake capable of causing substantial harm. That is a warning about the content under specified circumstances. It is not evidence that the benchmark produced the same number of injuries. Demonstrating prevented harm would require a further evaluation of actual use, including clinician responses and downstream events.
What a useful validation report would contain
The public research summary does not identify the nine models by name and version or provide a full reproducible protocol. A fuller report would help readers assess both the baseline errors and the verification result. This is a normal next step for a promising technical claim.
The TRIPOD-LLM reporting guideline provides a relevant starting point: it addresses transparent reporting of model versions, prompting, outcome definitions, assessors and evaluation procedures. It is a reporting guideline, rather than a certification of a product’s quality.
For this particular task, a practical evidence package could include:
- The question-selection process, complete answer count and distribution of claims per answer.
- A description of the independent reference and how disagreements were resolved.
- Separate results for claim extraction, verification, unsupported cases and complete answers.
- Performance by relevant medicine and patient-context groups, with uncertainty estimates.
- A follow-up evaluation of how clinicians interpret badges and act on unresolved checks.
The last item matters because a tool changes a workflow as well as a number. The DECIDE-AI guideline addresses early live clinical evaluation, including human factors. It supports asking whether a technically useful check remains useful when people encounter it during real work. Such evaluation could examine missed warnings, unnecessary interruptions and whether a reassuring label discourages further review.
UK readers should examine jurisdiction and source
The source behind a medicines answer must fit its intended setting. emc’s own description identifies it as a UK medicines-information resource containing approved product information, including Summaries of Product Characteristics. The relevant product and current document matter when examining a particular claim.
iatroX’s Datapharm partnership announcement, published on 12 May 2026, describes integrating emc medicines information into askiatroX. This is a source-integration partnership. It does not establish that iatroX provides the same deterministic verification functionality as DoseCheck, or that the two systems have equivalent validated performance.
For a UK learner, the practical opportunity is to make source scrutiny a repeatable habit. Identify the product, inspect the relevant source, check whether the answer preserves its qualifications and record what remains uncertain. A US-oriented benchmark can prompt that habit without being treated as validation for a UK product or workflow.
Turn the benchmark into a focused learning exercise
A useful CPD activity would examine a small set of educational medicines questions without entering patient-identifiable information. Before consulting an answer, write down which contextual details could change its interpretation. Afterward, inspect the supporting information and classify any discrepancy: missing context, wrong product, incomplete qualification, unsupported statement or unresolved evidence.
The reflection can then describe a specific change in practice, such as checking whether a source addresses the stated population before accepting a concise summary. Repeating the exercise later can reveal whether that habit has become consistent. The learning outcome is a better method of evaluating answers; a badge or a headline accuracy figure cannot supply that method on its own.
Frequently asked questions
Does the 2.3% figure mean 2.3% of patients received a wrong dose?
No. The denominator consists of evaluated claims in generated text. Translating that result into events affecting patients would require evidence about real clinical use, including the prescriptions actually issued.
Does deterministic verification establish that the whole answer is correct?
It establishes repeatability of the specified check. Assessing the whole answer also requires attention to claim extraction, applicable context, database coverage and statements outside the check’s scope. Readers need those boundaries to interpret a verification result.
How is Datapharm relevant to UK medicines questions?
Datapharm owns emc, which publishes approved UK medicines information, including SmPCs. The iatroX partnership brings that source material into askiatroX, helping readers connect a concise answer with product-specific information relevant to UK practice.
Turn evidence appraisal into learning
Explore iatroX Medicines for evidence-linked UK medicine guides, then use iatroX CPD and Tutor to turn an identified learning gap into questions, discussion and a reflection you review yourself.