A clinical AI should let its user inspect the evidence behind a consequential claim, understand its scope and identify what remains uncertain. OpenEvidence, AMBOSS and iatroX should be assessed against those requirements, rather than ranked by citation count, confident language or the number of features on a landing page.
This comparison is published by iatroX and includes iatroX among the tools considered. It is a review of public product descriptions checked on 10 October 2026 and a proposed testing protocol, not a completed hands-on comparison. No answer-quality scores or observed response examples are being claimed.
Start with the question the reader is trying to answer
A clinician looking for a guideline recommendation has a different task from someone examining a contested research question or revising a concept for an assessment. All may benefit from sources, but the appropriate source and the required explanation differ.
The comparison should therefore specify the user, jurisdiction and task before examining the output. A long research synthesis may be useful for one question and unnecessarily cumbersome for another. A concise guidance-led answer may be efficient but inadequate if it hides a material disagreement.
The standard is not that every tool must behave identically. It is that each should make its evidence and boundaries sufficiently clear for its intended use. A product can be a good fit for a particular reader without being universally more accurate or clinically safer.
What the public materials actually describe
As checked on 10 October 2026, OpenEvidence Academy positions its introductory learning around clinicians obtaining citation-backed answers at the point of care. OpenEvidence's security and compliance page also identifies content agreements with named medical publishers. Those statements establish a published source-oriented proposition, not that every generated claim is supported.
AMBOSS's Clinical AI Mode page, checked on the same date, describes linked answers drawing on its articles, medicines content and selected external sources. It acknowledges possible misinterpretation and omissions, and states intended-use limitations, including that the tool is not for diagnostic or therapeutic purposes or time-sensitive emergencies.
iatroX's methodology, checked on 10 October 2026, describes retrieval, ranking, citation grounding, output checking and uncertainty handling. Its reference proposition is UK guidance-led, using NICE, CKS, SIGN and SmPC information from emc. These are design and source-selection descriptions, not comparative outcome evidence.
| Platform | Published starting point | What a reader should still verify |
|---|---|---|
| OpenEvidence | Citation-backed clinical information and publisher content relationships | Whether the precise claim is supported and fits the requested setting |
| AMBOSS | Source-linked synthesis of curated content and selected external material | Whether qualifiers, limitations and source disagreements remain visible |
| iatroX | Guidance-led reference with published retrieval and checking methods | Whether the cited passage supports the answer and is locally applicable |
This table records documented positioning. It does not score interfaces that were not tested or imply that an unmentioned capability is absent.
The four evidence checks should be identical
First, verify that the cited document exists and is the correct version. A familiar publisher name or a link to a related topic is not enough.
Second, inspect claim-to-passage support. Does the source justify the recommendation as written, including its conditions and exceptions? A conditional recommendation should not become a general rule through summarisation.
Third, establish currency. A recently generated response can rely on older guidance, while a recently updated web page may contain unchanged clinical material. The relevant question is whether the recommendation remains current, not whether the interface displays today's date.
Fourth, assess applicability. Population, setting, jurisdiction and available services can change how evidence should be used. A source may be correct and relevant to the disease while still not resolving the user's actual decision.
These are proposed common criteria. The testing process should permit a tool to say that the available information is insufficient. Appropriate uncertainty is not a failure merely because another product supplies a more decisive answer.
A reproducible test set, without invented results
A useful initial protocol would use fictional, non-identifiable questions spanning several distinct challenges. Include a straightforward UK-guidance query, a question with an important eligibility condition, a disagreement between sources, a recommendation that has changed and a case missing information necessary for an individual decision.
Use the same initial wording across products and record any subsequent interaction separately. A follow-up that supplies missing context should not be hidden from the account of how the final answer was obtained.
For every run, record the date, product, available version information, region, account type and settings. Save outputs only through an appropriately governed process. Establish the assessment criteria and reference material before judging the responses, with reviewers able to record legitimate disagreement.
The outcomes should include unsupported claims, consequential omissions, preservation of uncertainty, source accessibility and the work needed to verify the answer. A simple tally of citations would miss most of the important distinctions.
No such head-to-head run was performed for this article. Results should be published only from an actual recorded evaluation, with its methods and limitations, rather than reconstructed from product descriptions or imagined example answers.
Example questions for a dated run
These proposed prompts, prepared on 10 October 2026, can be entered unchanged in each tool. They are fictional learning inputs, not patient consultations, and no responses to them are reported here.
Guidance and applicability: "UK clinical-learning question: which current NICE source should a clinician inspect when reviewing acne management in an adult, and what patient context would be needed before applying its recommendations? Do not provide doses."
Uncertainty preservation: "A fictional note says a possible diagnosis was discussed and investigations remain pending. Does this establish a confirmed diagnosis? Identify the unsupported inference and the evidence still required."
Incomplete follow-up: "In a fictional learning case, a test has been requested, but the record does not identify who will review the result. What remains unresolved? Do not invent an owner or assume that the test has been completed."
For a fuller study, add prespecified source-conflict and freshness cases using verified documents and an independently prepared reference assessment. Keep those materials identical across products.
Claimed safeguards and observed behaviour are separate columns
A vendor may describe a checking process that is not directly visible in the interface. That description can be relevant, but it is not the same as observing the behaviour under test.
Similarly, a single successful demonstration does not establish consistent performance across paraphrased questions, incomplete inputs or repeated runs. An evaluation should distinguish documentation reviewed, features observed and outcomes measured.
This distinction protects suppliers as well as readers. The absence of a detailed public explanation does not prove the absence of a safeguard. The correct entry may be "not established from the material reviewed", followed by a precise question for the vendor.
Apply that discipline equally to iatroX. Its methodology should be open to challenge, and its reference links should be checked rather than treated as inherently trustworthy because this article appears in its Journal.
Privacy and learning need their own assessment
An answer-quality comparison does not establish suitability for patient data. A clinician should examine the actual processing arrangements and organisational approval before entering identifiable information. Reference questions can often be framed generally without reproducing a clinical record.
iatroX's privacy policy, checked on 10 October 2026, describes usage and query logging and processing by AI service providers. It should not be characterised as entirely local, zero-retention or exempt from privacy scrutiny.
Learning functions are also a different question from reference accuracy. As described on 10 October 2026, iatroX's Socratic Tutor supports follow-up reasoning around attempted questions. That may suit a learner seeking to explore a misconception, but does not establish that the platform gives more accurate clinical answers than the other products.
A fair comparison allows the same person to use different tools for different purposes, while recognising that maintaining multiple overlapping subscriptions can create unnecessary cost and fragmentation.
Verdicts by reader scenario
For a clinician primarily exploring research, OpenEvidence's citation-oriented proposition is relevant to shortlist, subject to checking the actual sources, access conditions and applicability. Publisher relationships alone should not decide the purchase or trust decision.
For a reader who values a curated explanatory library alongside linked synthesis, AMBOSS offers a relevant documented approach. Its stated limitations should be respected, and the proposed test should examine whether the user's important qualifiers survive the synthesis.
For a UK clinician or learner seeking guidance-led reference and reasoning-focused practice, iatroX offers a relevant combination. That fit is a statement about task and workflow, not proof of clinical superiority.
For a consequential decision with unresolved context or evidence conflict, direct source review and appropriate clinical input may be more important than choosing another chatbot. The useful winner is the process that lets the reader understand and justify the decision, not the product with the most reassuring interface.
Frequently asked questions
Which of OpenEvidence, AMBOSS and iatroX is most accurate?
This article does not establish a comparative accuracy ranking because no consistent head-to-head evaluation was performed. Product descriptions, source relationships and methodology statements cannot substitute for that evidence.
Does a larger number of citations make an answer more trustworthy?
Not necessarily: the important questions are whether the sources support the precise claim, remain current and apply to the context. Several weak or irrelevant citations do not establish stronger support.
Can iatroX be assessed using the same criteria as its competitors?
Yes: its claims, sources, uncertainty handling and privacy arrangements should face the same scrutiny. Publication by iatroX is not a reason to exempt it from the comparison protocol.