MDCalc puts privacy filtering at the center of clinical AI use
MDCalc has launched CrossCover, a tool designed to help clinicians use generative AI without exposing patient or provider identifiers to the underlying chatbot. The pitch is simple and unusually direct: preserve the usefulness of large language models while removing the data elements that make many clinicians uneasy about putting real clinical questions into them.
The new product arrives at a moment when medical AI tools are being pushed in two different directions. One camp is trying to make chatbots more powerful by folding in broader streams of patient, clinician, and payer information. Another, represented by CrossCover, is trying to reduce the amount of sensitive information that ever reaches the model in the first place. In clinical practice, that difference is not cosmetic. It changes the privacy risk, the regulatory burden, and the level of trust a physician may feel when asking an AI system for help with a decision at the bedside.
Graham Walker, MDCalc’s head of AI, framed the launch as a way to make generative AI safer and more usable for clinicians. “We can take away some of the challenges and risks and problems with the technology so that physicians can feel more confident in using it,” he said.
A privacy-first design in a field chasing more data
CrossCover’s design runs against the prevailing instinct in parts of the health tech market, where companies have aimed to improve AI output by granting systems access to ever more clinical context. OpenEvidence, Abridge, and Epic have each moved in that direction in different ways, building products that can work with broader clinical data or integrate more deeply into existing health system workflows. MDCalc is taking the opposite route: rather than expanding the data pipeline, it is trying to narrow it.
The clinical logic behind that choice is straightforward. When clinicians input a question into a chatbot, even a short prompt can contain protected health information, identifiers, or enough contextual detail to make a case recognizable. A privacy filter that strips those elements before the prompt reaches a generative model reduces the chance that sensitive information is exposed, retained, or reused in ways clinicians cannot fully control. In a health care environment where documentation systems, billing systems, and AI tools increasingly intersect, a “rumble strip” approach may appeal to physicians who want an audible warning before the car leaves the lane, not a full autonomy system that invites them to trust the machine more than the safeguards.
The strategy also reflects a deeper cultural problem in medical AI adoption. Many clinicians are not asking whether a chatbot can produce a polished answer; they are asking whether they can safely use it at all. If the model requires registration with a national identification number or a formal privacy arrangement before protected health information can enter the system, adoption may slow even when the output is technically strong. CrossCover attempts to remove that friction by making the question itself safer before it is submitted.
Privacy is not a side issue in clinical AI. It is the threshold question.
Trust in the answer begins with trust in the input
Generative AI in medicine has been sold largely on speed: faster summaries, quicker differential diagnoses, cleaner note drafting, easier access to reference material. Yet speed does not solve the central clinical problem, which is that physicians are ultimately responsible for what happens to the patient after the model responds. A tool that improves the quality of the prompt by removing identifiers, extraneous personal details, and other sensitive information can help clinicians use AI without handing over more than is necessary for the task at hand.
That can matter in routine practice. A clinician looking for a dosing check, a differential, or a guideline reminder may not need the patient’s name, date of birth, medical record number, or a free-text narrative loaded with identifying details. If those data points can be removed automatically, the model still receives the clinical essence of the question while the exposure footprint shrinks. The result is not perfect anonymity, but it is a practical attempt to reduce risk where the habit of convenience often overwhelms caution.
There is also a workflow angle. Systems that ask clinicians to navigate complex registration steps or privacy permissions before using AI can create another barrier in already crowded digital environments. CrossCover appears aimed at the opposite behavior: make the safer option the easier one. In health care, that can be more influential than a polished promise of intelligence.
MDCalc is betting clinicians want guardrails more than breadth
MDCalc’s decision to frame CrossCover as an accelerator for clinical generative AI is notable because it recognizes a basic truth about adoption. Physicians are unlikely to embrace AI at scale if every use case feels like a privacy compromise. A model that can be consulted without feeding it unnecessary patient identifiers may lower that barrier, especially for clinicians who already rely on MDCalc for decision support tools and are accustomed to practical bedside utilities rather than abstract AI demos.
Still, the launch raises the questions that always follow privacy tools in medicine. How much information can be removed before the query loses meaning? Which prompts remain safe enough to submit, and which are too specific to sanitize reliably? How consistently will the tool perform when clinical language is messy, abbreviated, or embedded in free text? Those are operational questions, but they are also patient-safety questions, because the quality of AI output is only as good as the system that prepares the input.
CrossCover’s launch suggests that the next phase of clinical AI may not be defined solely by more powerful models. It may also be shaped by the infrastructure wrapped around them: systems that redact, filter, and constrain what reaches the model in the first place. For many clinicians, that may be the more persuasive form of innovation. Not bigger answers, but safer ones.
The unanswered issue is whether privacy stripping alone will be enough to make physicians comfortable using generative AI for real patient care, especially when workflow pressure, model uncertainty, and liability concerns remain unresolved.