MDCalc has launched CrossCover, a service designed to add scrutiny around clinical AI conversations. The announcement on 11 November names AMBOSS and Vera Health as founding AI partners, with ACP and ECRI supplying reference content. Participating partners generate the underlying answers; CrossCover applies checks around that interaction. MDCalc's announcement
The proposition is significant because it changes the question from “Which assistant should I ask?” to “What should happen around the answer before I use it?” A persuasive response can contain a subtle error, miss an exception or apply evidence to the wrong setting. A separate review layer could help a clinician notice those problems, provided its findings are understandable and its limits remain visible.
For a working clinician, CrossCover therefore deserves a precise reading. Its name describes a supplementary role, while the meaningful product questions concern the information it receives, the checks it performs and the decisions that remain with the user. For iatroX, the launch also reinforces a substantial opportunity: helping professionals develop the understanding needed to question an answer and use the evidence intelligently.
How the proposed interaction works
The launch path is a conversation with a participating AI through CrossCover, followed by review of the answer and supporting material. The product's FAQ says CrossCover is not itself another large language model. Its role is to add checks outside the system producing the original response. CrossCover features and FAQ
That distinction matters when assessing a demonstration. There is an answer-producing component and a reviewing component, potentially with different sources and failure modes. A useful interface should help the clinician understand which statement came from which component and what action, if any, follows from a discrepancy.
Consider the amount of work hidden in a simple clinical question. The user has to describe the uncertainty, decide what information can appropriately be submitted, interpret the returned answer and inspect consequential claims. CrossCover's opportunity is to make parts of that work more visible. Its success would depend on whether the additional information improves review without creating so much noise that clinicians stop paying attention.
This is also why access through participating partners should not be mistaken for a universal plug-in covering every medical chatbot. The reviewed launch materials describe partner access within CrossCover. They do not establish that a browser extension already checks any response displayed anywhere, or that a hospital's existing assistant automatically acquires the same safeguards.
Four features addressing different problems
The four named features deserve separate evaluation because they act at different points in the interaction.
| Feature | Described function | Practical question for a clinician |
|---|---|---|
| PHI Scrubbing | Identifies potential patient identifiers and recommends redaction before transmission | What remains in the text I am approving? |
| CrossCheck, in beta | Compares clinical output with medical references and surfaces discrepancies or uncertainty | Does the finding explain what needs checking? |
| TrackerBlocking | Limits unnecessary information about the clinician reaching partner models | Which information is withheld, and from whom? |
| AI-Assisted MDCalc | Surfaces a relevant clinical tool within the workflow | Is the selected tool appropriate for the question? |
These are the vendor's described functions, rather than measured estimates of detection performance. CrossCover feature descriptions
A product can perform one of these functions well and still require improvement in another. Removing an identifier does not verify a recommendation. Finding a disagreement does not establish which source should govern the case. Opening a calculator does not confirm that the inputs or the chosen population are correct.
PHI scrubbing changes the review step
The important verb in the feature description is “recommends”. The clinician still needs to inspect the proposed submission. A redaction suggestion is useful only if the remaining text is appropriate to send and retains enough relevant information for the intended question.
CrossCover's current standard terms prohibit submitting protected health information. They also explain that automated safeguards do not guarantee detection of every identifier or create a business-associate relationship. Separate institutional agreements may govern a different arrangement. CrossCover terms, effective 28 October 2026
This makes screening and permission distinct questions. A clinician should not add identifiable material merely to see whether a filter will catch it. Nor should a clean-looking prompt be treated as proof that every privacy requirement has been met. The practical workflow begins with a question framed for the service's permitted use, followed by inspection of any suggested changes.
There is a second balance to consider. Removing information can reduce exposure, but indiscriminate removal could also change the meaning of a clinical question. An evaluation should examine both outcomes: what potentially identifying detail remains, and whether the submitted question still communicates the relevant clinical uncertainty. That is a proposed test of the workflow, not a claim that CrossCover has been shown to fail it.
A separate check needs an understandable explanation
CrossCheck's attraction is the possibility of examining a response from outside the system that wrote it. A model's own citations help locate its sources, but a reviewer may also want to know whether a cited passage supports the particular claim and whether another reference changes the interpretation.
Imagine a fictional teaching exercise in which an AI explanation describes a study's follow-up as six weeks, while the supplied abstract says six months. A useful review would identify the exact statement and the relevant passage, making the difference easy to inspect. A vague warning that “the evidence may differ” would leave considerably more work for the clinician.
Other disagreements will be less tidy. Two recommendations may differ because of their publication date, patient population or practice setting. A checking system should help the user investigate that distinction. Counting every disagreement as an error could be misleading, just as treating agreement as proof could be misleading.
The current terms make a crucial boundary explicit: the absence of a CrossCover flag does not establish that an answer is accurate, complete or safe. CrossCover Schedule C
The product's performance therefore cannot be inferred from the number of warnings on a screen. The clinically useful question is whether the review draws attention to the right passage, explains the significance and makes the next verification step clearer.
Independent checking is a design claim requiring evidence
Several forms of independence are possible. A checking process can sit outside the original answer model, use separately selected references or be operated by another organisation. Those arrangements can matter without establishing that the two processes make unrelated mistakes.
For example, agreement may occur because two systems rely on the same incomplete source. A discrepancy may arise because they retrieve different editions of a reference. Neither situation can be resolved simply by preferring whichever component carries the label “checker”.
The reviewed launch pages do not provide a prospective clinical-outcome study or public estimates of CrossCheck's false-positive and false-negative rates. That leaves important evaluation questions open. Which kinds of error can it detect? How often does it flag a clinically reasonable qualification? Does performance change when the upstream model changes?
These questions support constructive scrutiny. A review layer is a promising product idea precisely because checking clinical AI is difficult. The evidence needed to assess it should describe that difficulty and show how the finished workflow handles it.
TrackerBlocking has a defined destination
Information withheld from a partner model is different from information collected by the platform operating the account. CrossCover's privacy policy says account identity is not transmitted with a query and describes user-approved screening before transmission to AI providers. It also permits use of de-identified user content to train, evaluate or improve MDCalc or third-party models. CrossCover privacy policy, updated 28 October 2026
Consequently, phrases such as “anonymous”, “no data collection” or “never used for training” would give an inaccurate impression of the documented offer. The practical question is which data moves to which recipient for which purpose.
This is particularly relevant when comparing CrossCover with an enterprise workspace that deliberately brings authorised patient context into an organisational environment. The two approaches address information differently. A fair assessment examines the complete data path and applicable terms, rather than assuming that more context is always preferable or that less context removes every risk.
Access starts with a defined group
CrossCover's FAQ describes current free access for eligible healthcare professionals and supervised medical trainees who reside and work in the United States. Its Charter Physician programme uses an MDCalc account and says users will be emailed when access is ready. The MDCalc landing page separately states that international clinicians can register interest but do not currently have international product access. CrossCover access FAQ, MDCalc launch landing page
For UK readers, an interest form is therefore different from an available clinical service. For an eligible US reader, the launch still requires checking the actual access offered to the account. The geographical reach of MDCalc's existing calculator platform should not be transferred automatically to this newer service.
The FAQ also says advertising is expected to help support the platform over time, alongside a commitment that commercial relationships will not determine clinical answers. CrossCover business-model FAQ That commitment provides a question for future transparency reporting; it is not itself an independent audit of commercial influence.
What a useful early evaluation would record
An organisation or research group assessing a checking workflow should define success before collecting attractive examples. Useful measures could include whether a material issue was identified, how much verification work remained and whether a reviewer understood the information boundary.
| Evaluation dimension | Observation worth recording |
|---|---|
| Problem detection | Did the check identify the issue the case was designed to contain? |
| Explanation | Could the clinician locate and understand the supporting passage? |
| Unnecessary warnings | Did a reasonable statement receive a distracting flag? |
| Missing information | Was uncertainty made visible rather than silently resolved? |
| Review effort | How long until the reviewer reached an acceptable interpretation? |
| Subsequent understanding | Could the clinician explain what changed and why? |
These are proposed assessment questions, not results from a CrossCover trial. Any actual evaluation should use authorised materials and respect the applicable terms governing access and outputs. The purpose is to examine useful behaviour, rather than manufacture a single headline score from unlike tasks.
Where iatroX makes a positive case
A checking layer can identify something that deserves attention. A learning workflow can help the clinician understand why it matters and recognise the issue next time. That is a concrete place for iatroX to offer value alongside developments such as CrossCover.
A useful learning conversation examines a specific uncertainty exposed by an attempted answer. The learner has committed to a choice, so the discussion can investigate that reasoning rather than begin with another general summary. Our comparison of reference and learning workflows explains how to assess that experience in iatroX.
For a UK clinician, iatroX is a strong starting point when the immediate goal is source-linked reference followed by deliberate learning. The useful sequence is to inspect the evidence, identify what remains unclear and practise the relevant distinction. A product that helps with that sequence has a substantial role even as other software adds checks around clinical answers.
CrossCover's launch makes the surrounding work more visible. Clinicians still need to frame the question, interpret the evidence and develop the judgement they carry into the next case. Explore Ask iatroX for a reference question, then choose relevant practice and Tutor discussion for the learning need it reveals.