At a glance: A virtuous cycle between AI-augmented clinical practice and clinician education is a loop where what happens at the bedside shapes what gets taught, and the performance of those who were taught becomes the evidence for what to teach next. The evidence suggests most institutions cannot yet verify that loop is closing.
Breast Cancer Detection as the Case in Point
October is Breast Cancer Awareness Month, and detection is changing on three fronts at once. Imaging is improving, including AI-supported mammography reading. Blood-based tests, including multi-cancer early detection (MCED) tests that look for signals of many cancers from one blood draw, are arriving. And AI is analyzing electronic health record data to flag patients at higher risk. The evidence for all three is real, and still contested.
A randomized trial of about 106,000 women in Sweden’s screening program found that AI-supported screening detected 29% more cancers while keeping false positives at a comparable rate (Hernström et al., 2025). In the follow-up analysis, AI-supported screening met the non-inferiority standard for interval cancers, at 1.55 versus 1.76 per 1,000 (Gommers et al., 2026), but a correspondent in The Lancet argued that interval cancer rate does not directly measure advanced-stage disease or breast cancer mortality (Silipigni, 2026). Even the strongest trial leaves a live argument about what counts as proof.
Blood tests raise the same question. Congress has created a Medicare coverage pathway for FDA-approved MCED tests, yet robust outcome data, including mortality reduction, are still needed (O’Donnell & Marinac, 2026), and the NHS-Galleri trial missed its primary endpoint of a significant reduction in stage III to IV diagnoses, though stage IV cancers fell 14%, a key secondary result (Swanton et al., 2026). On September 23, an FDA advisory panel voted in favor of the first MCED application, but split 6 to 4 on effectiveness (The ASCO Post, 2026), and recommended that marketing materials avoid characterizing the test as an “early” detection tool (NPR, 2026). The FDA has not yet decided. Breast cancer may also shed too little circulating tumor DNA for reliable detection (Xu et al., 2025).
AI applied to the medical record adds a third layer. A study in npj Precision Oncology testing EHR-based models across eight cancers found that flagging high-risk individuals with EHR models concentrated true cancer cases three to six times more than traditional risk factors did, but the breast model’s top predictors featured “carcinoma in situ of the breast” (Park et al., 2026), a reminder that clinicians must know how to question what a model is keying on.
Three questions follow. How do we use these tools responsibly in service of the patient? How do we make sure both current and future clinicians are trained to lead in this area? And how do we assess whether we are teaching what is actually important to know, learn, and scrutinize? This piece answers the third.
Much has been written about whether AI is deskilling clinicians or adding to their toolbox, and Lecturio has already covered that debate. I want to ask a more operational question. If the honest answer is “it depends on how it’s done,” how would an institution know which outcome it is producing, for the clinicians it trains now and the ones it trained years ago? That is a measurement problem, and most institutions do not yet have the infrastructure to solve it.
Why “Is AI Deskilling Students?” Misses the Actual Stakes
Whether AI use erodes or extends skill is a real question, but even a perfect answer about one moment of use would leave the bigger gap: no one could say whether it holds up in practice over time, or whether that reality changes what gets taught next.
I will not claim that AI is or is not deskilling clinicians, because that cannot be settled without measurement. But deskilling need not happen, and the reason is worth teaching. One benefit of keeping live clinicians in the mix is understanding the patient in context: aspects of a patient’s life that may not be documented in an electronic health record, such as living circumstances, who is close to them, and their literacy level. AI alone also potentially leaves out a skill clinicians learn: leveraging their relationship with patients to help those patients manage their own health.
The Loop: Practice Informs Training, Training Produces Clinicians, Clinicians Become the Next Data Point
A virtuous cycle has three steps, and each depends on the one before it.
- Clinicians in practice use AI-enabled tools in real decisions, generating signals about what competence now requires: which results get acted on well, and which get over- or under-trusted.
- Those signals should shape what gets taught, first to students and residents, and in time to practicing clinicians through continuing education. Not just how to operate a tool, but what to scrutinize. Teaching this requires people who can teach it, and in my view not all current faculty can yet, so learning resources will need to be tiered for faculty as well as for learners.
- Newly trained clinicians enter practice, and their performance becomes the next round of signal.
The failure mode is a “vicious” version, where curriculum changes on faculty intuition or vendor pressure rather than downstream performance data, and no one learns whether a decision made three years ago helped or hurt.
The training question already spans the continuum: AAMC is developing AI competencies that span medical school, residency, and continuing education (Association of American Medical Colleges [AAMC], n.d.), and AMA is funding 11 precision education teams across more than 80 institutions (American Medical Association [AMA], 2026a). What counts as competence also depends on where humans stay in the loop, and that is being argued from several directions.
Nursing’s view of AI in clinical care has been less welcoming. National Nurses United surveyed more than 2,300 members in 2024, and 60 percent disagreed that they trust their employer to put patient safety first when implementing AI (National Nurses United [NNU], 2024). Its affiliates have bargained for a voice in which AI tools their hospitals adopt and how those tools are evaluated (Nurse.org, 2026). I read that as a reasonable demand for a voice in governance, and for the skills that must still work when the technology fails or is wrong, rather than resistance to AI itself. Press coverage describes the union as critical of how AI has been deployed but open to AI that could help its members (Tech Policy Press, 2026), and Bonnie Clipper’s framework in Lecturio’s own piece on the deskilling debate makes the same case: nursing leadership on the AI governance committee, frontline nurses in tool evaluation, and deliberate opportunities to demonstrate competence without the tool (Stevanovski, 2026).
In August, Ezekiel Emanuel, Abe Baker-Butler, and Neal and Vinod Khosla argued that autonomous AI will likely exceed both physicians and physician-AI teams at five cognitive tasks, and that a human in the loop may make performance worse once AI alone is better (Emanuel et al., 2026). They call for urgent work on medical education, among other areas, and they set radiology aside as an area where AI alone has not reliably outperformed physicians (Emanuel et al., 2026). I find that striking, because Emanuel co-wrote in 2016 that machine learning “could end radiology as a thriving specialty” within 5 to 10 years (Chockley & Emanuel, 2016). It is a Perspective, not a study, by authors with financial interests in AI companies (Wachter, 2026).
Wachter (2026) agrees that AI will often outperform clinicians on a curated set of facts but argues that does not mean clinicians can be replaced. He invokes the doorman fallacy: when automatic doors replaced doormen, people realized the doormen had done far more than open doors, and many properties rehired them (Wachter, 2026). His examples include end-of-life family conversations and finding care for an uninsured patient. He also describes primary care splitting into AI-first care for routine needs and an AI-enabled clinician for complex needs or those with more resources (Wachter, 2026). I want to push on what that could mean: a two-tier system, different from today’s, in which better-resourced patients have a clinician in the mix of AI-assisted care and less-resourced patients get AI alone or mostly alone. Wachter discloses that he advises Curai, in which Vinod Khosla invests.
The Measurement Gap: Most Programs Can’t Currently Close This Loop
Most programs cannot close this loop today, and the reasons are structural.
- Linking training to practice is rare, and the results are mixed. In a 2023 commentary with Lisa Howley, I noted that competency efforts have mostly stayed siloed within training programs and that graduate medical education outcomes have rarely been connected to outcomes the health system or regulators define (Grover & Howley, 2023). It accompanied what it described, to its authors’ knowledge, as the first published study linking GME assessment data to patient complaints: residents rated below “proficient” on ACGME Milestones in professionalism and communication had significantly higher odds of complaints in their first year of independent practice (Han et al., 2023). A 2024 JAMA study of new hospitalists found no association between overall or medical knowledge milestone ratings and hospitalization outcomes, though certification exam scores were associated with lower 7-day mortality and readmissions (Gray et al., 2024).
- Attribution is hard even when downstream data exist. In a national analysis of 57,132 first-year residents, within-specialty variation in milestone ratings was attributed about 35% to residency programs, 22% to trainees, and 2% to medical schools (Santen et al., 2025). That is not a verdict on schools; it shows how hard it is to read a curriculum’s effect off early ratings, and the interpretation is mine, not the authors’.
- Data are fractured. In the AMA webinar (AMA, 2026b), Alexis Rossi cited a MedBiquitous survey finding institutions use an average of 15 distinct systems to support medical education. There is no shared framework comparable to the clinical data standards behind electronic health records, and the one being built is not finished.
What Closing the Loop Actually Requires
Closing the loop takes four things, and the table shows what each changes for an institution.
| Institutional question | Open loop today | Closed loop |
| How are curriculum changes decided? | Faculty intuition or vendor pressure | Downstream performance data |
| What can a program say about its AI training? | That it was delivered | Whether it holds up in practice, in a form deans and accreditors can examine |
| Are faculty ready to teach these skills? | Unevenly; not all can yet | Faculty development, with resources tiered for faculty and learners |
| Does data travel across the continuum? | Many separate systems, 15 on average in one survey (AMA, 2026b) | Shared standards from medical school to practice |
- Competency-level data during training and practice, not just final answers. An illustration that ties back to the opening: a simulator study of several hundred practicing physicians found that palpation force below roughly 10 newtons put clinicians at significant risk of missing deep lesions near the chest wall, and that human observers cannot reliably judge how much force a clinician applies (Laufer et al., 2015). A 2024 follow-up with 152 clinicians found that successful examiners kept force steady and pressed about 15% harder when palpating, while unsuccessful ones did not search with enough force to detect deep masses (Laufer et al., 2024). That is a simulation finding, not a patient-outcome finding, and Carla Pugh noted in the AMA webinar (AMA, 2026b) that such tools do not replace human observation.
- Shared data standards. A common language that people and machines read the same way, so data can travel from medical school to residency to practice.
- A real path from signal to decision. Feed-forward and feedback built in, with a scheduled route for data to reach curriculum committees and program directors instead of piling up unread.
- Trust. Audience questions at the AMA webinar (AMA, 2026b) centered on bias and on prematurely labeling learners. The answers: transparency about what data is used and how, clinicians deciding what the data means, and data used to help people improve, not to threaten them. Measuring the loop is not the same as ranking people.
The Verdict
Whether AI is deskilling clinicians or adding to their toolbox cannot be settled by philosophy or intuition. Only a measurement system can answer it, for the students and residents being trained now and in time for clinicians already in practice, and most institutions do not have one yet.
This calls back to the opening. The argument over what counts as proof for AI mammography and blood-based screening is the same argument education faces: whether clinicians scrutinize a new screening result or an AI flag appropriately is something to measure, not assume. For deans and program directors, that means defensible curriculum decisions, accreditation-linked outcomes such as the ACGME Milestones (Santen et al., 2025), and not building “AI-forward” training on faith. Quantifying clinical reasoning shows one way that measurement can look.
No institution should have to work this out alone. The questions raised here, from what to measure to how to act on it without ranking people, are the ones I want to take further with deans, program directors, and the faculty who teach, and the questions below are a place to start. If you want to take these questions further, I will be discussing them in More Data, Harder Choices: AI, Breast Cancer, and What It Means for the Future of Care, Medical Education, and Patients on Wednesday, October 21, 2026 at 12:00PM EDT/9:00 AM PDT/6:00PM CEST. Register for the webinar.
Frequently Asked Questions
Is AI deskilling students and clinicians, or giving them new tools?
Neither answer can be verified without a measurement system. That is the central point of this piece.
What is a virtuous cycle between clinical practice and training?
A loop where what happens in AI-augmented practice shapes what gets taught, and the performance of the people taught becomes the evidence for what to teach next.
Can institutions currently track whether AI-integrated training produces better clinicians?
Mostly not yet. AMA’s precision education grants are building systems for data-driven learning across the continuum (AMA, 2026a), and AMA does not expect to have solved the problem by the end of the grant period (AMA, 2026b).
What would closing this loop actually require?
Competency-level data, shared data standards, a real path from data to curriculum decisions, and trust among the people being measured.
Who will teach students and residents to work with AI?
Faculty who have these skills themselves. In my view, not all current faculty do yet, so faculty development, with learning resources tiered for faculty as well as for learners, has to run alongside the training of students and residents.
Does measuring learners and clinicians risk mislabeling them?
It can, and audience questions at the AMA webinar (AMA, 2026b) raised that concern. Measurement should be transparent, used to help people improve, shaped by clinicians themselves, and never a substitute for human observation.