Never-Skilling: The Silent AI Risk Decision Facing PA Leaders

A new Nature Medicine paper names a danger that lands on every table doctoral PAs sit at, and the window to lead it is open right now.

Matt Bell, DMSc, PA-C, CAQ-HM -- Chief AI Officer, WNC VA Health Care System

Author's Bio:

Matt Bell, DMSc, PA-C, CAQ-HM, is a Physician Associate with nearly three decades in internal and hospital medicine who currently serves as Chief AI Officer and the Chief PA for the Western North Carolina VA Health Care System. He works at the intersection of healthcare AI governance, clinical operations, and education, and is an accepted presenter at the upcoming 2026 PAEA Education Forum. He is a doctoral-prepared PA and a member of the Academy of Doctoral PAs.

A student is presenting a case.

She is good. The differential is broad and well-structured. The workup is efficient. The management plan is current, evidence-aligned, and articulated with a confidence most learners take years to earn. The preceptor is, frankly, impressed.

Then the preceptor asks her to set the laptop aside and reason aloud, just her, the history, and the patient in front of her. 

And she stalls.

Not because she is unprepared, but because the reasoning she delivered so fluidly was never fully hers. The structure was AI model-generated, as was the pathway she followed. What looked like clinical judgment was a refined echo of the model’s thinking. Strip away the AI tool, and the depth of independent reasoning is thinner than anyone in the room assumed. The student was not lazy. The model was not wrong. But something that was supposed to be built during training had, quietly, never been built at all.

That gap has a name: Never-Skilling.

A New Word for an Old Fear

In my previous article, I alluded to what many in the Healthcare AI world have been discussing in the background over the past few months.  Shortly after I published that blog, a new Nature Medicine article featured a Perspective by Ke, Jin, Ong, and colleagues that further elevated concerns about never-skilling.¹

The authors draw a distinction that matters more than it first appears. Medicine has long worried about de-skilling, which is the erosion of competencies in experienced clinicians who lean on automation until the old skill fades. That is a real risk, and the paper documents it using the example of endoscopists who routinely used AI-assisted colonoscopy, who showed a 6-percentage-point lower adenoma detection rate when they later worked without the AI.³ The skill was there, however, disuse dulled it.

The authors name a second risk, mis-skilling, in which a trainee uncritically absorbs a model's errors and biases (internalizing flawed reasoning, including race and gender-based associations), as fact.

The heart of the paper is the third category, where PA leaders should focus. “Never-skilling” is not the loss of a skill but the failure to build it in the first place. When AI assumes the cognitive burden during the period when trainees should be developing clinical reasoning and critical thinking (building the differential, framing the plan, and working through uncertainty before seeing the answer), the underlying reasoning may never take hold. The trainee logs the hours but bypasses the mental work those hours are meant to provide, and that gap follows them into early practice. The result is what the authors call false proficiency: competence that holds only in the presence of AI and collapses without it. The central line of the paper deserves to be quoted directly: “a copilot is only useful if trainees first learn how to be pilots.”¹

Why This Is Not the Calculator Panic

A reasonable leader's first instinct is healthy skepticism. We have heard this before. Imaging would kill the physical exam. The electronic health record would erode clinical memory. Calculators would ruin our ability to do mathematical equations on our own. Each anxiety came and mostly went without lasting harm. Why should AI be different?

The authors anticipate this objection and address it rigorously, which makes the paper credible rather than alarmist. AI differs from prior technologies in two specific ways.¹

First, the older tools changed the kind of thinking required; however, they did not eliminate the need for thinking. A CT scan still takes real expertise to read. Lab values still have to be read and incorporated into the overall clinical picture. The EHR organizes the data but does not draw conclusions for you (though this is also changing with AI, which can lead to another potential problem, which we can cover in another blog if there is interest). A large language model can do the whole chain on its own, from the history to the differential to the final plan. That is not a shift in thinking; it is a stand-in for it.

Second, the timing is different. Overreliance on imaging tends to appear after foundational skills are established. AI enters at the very start of training, before clinical reasoning has taken shape. As the authors put it, the question is no longer whether trainees develop these skills and then lose them, it is whether they develop them at all.¹

I want to be precise here, because precision is the difference between leadership and hype. The authors are clear that never-skilling is a risk model, not an established fact. We don't yet have the long-term studies in clinical trainees to prove it. The signals are indirect: the colonoscopy data,³ a study in which students who were given an unrestricted AI tutor scored 17% lower on a later unaided exam, with the biggest drops among the students who were already struggling, ⁴ and a pre-print study linking AI-assisted writing to weaker memory of what people had just written.⁵ Again, suggestive, not conclusive.

That caveat is not a reason to wait. It is the reason to lead now, while the standards are still unwritten.

The Reframe: This Is Not a Student Problem

Here is where the PA profession, and doctoral PAs in particular, need to hear something plainly.

It is tempting to file never-skilling under "education," in other words, a problem for program directors and curriculum committees to sort out. That reading is too small, and it lets most of us off the hook that we are already on.

Never-skilling is a leadership problem with a student-shaped symptom.

The harm is decided long before it shows up in a struggling learner. It is decided in the room where a health system chooses which AI tools go live on the hospital wards that a PA student will rotate through. It is decided in the boardroom where governance defines what "competent" means in the presence of AI. It is decided in the consulting engagement, where a standard of care is drafted, and in the courtroom, where, eventually, someone argues about who was accountable when an AI-dependent clinician was handed a case that the AI could not handle.

Doctoral PAs sit at all of these tables as faculty, preceptors, chief PAs, and process improvement leaders; as entrepreneurs building the tools; and as legal and regulatory consultants guiding the institutions that deploy them. We are not bystanders to this risk. In each of these roles, we are actively shaping these decisions, often without realizing it.

The Sequencing Test: Three Questions for Any Table

The central insight of the Nature Medicine paper is that the danger lies not in AI itself, but in its sequencing: deploying answer-generating systems before foundational competence has developed.¹  The same tool that undermines a novice can amplify an expert.

The leadership question is not whether to use AI, but in what order to introduce it. Drawing on the paper’s framework, I developed a brief tool, the Sequencing Test, which I now teach and share with students and colleagues. It consists of three questions:

1. Build or borrow? Is the tool strengthening the user’s reasoning, or replacing it with ready-made answers? The paper’s key distinction is between answer-delivery AI (a ward-round assistant that produces the differential) and learning-oriented AI (a tutor that asks questions and forces the learner to reason it out). ¹ Both have value. Only one is safe for those who have not yet built the skill.

2. Can the user check it? The rule for safe AI use is simple: verify, then trust. But you can only check something if you already know enough to judge it. As the authors put it, a trainee with no clinical foundation of their own can't check the AI, they can only take its word for it.¹,² Before you roll out a tool, ask whether the person using it could actually catch it being wrong.

3. Does it survive the unplug? If the tool suddenly wasn’t available (due to an outage, downtime, or while covering a shift at a rural site that never had it), would the work still be safe? That's the false-proficiency test. If the honest answer is “no,” you haven't deployed a copilot. You've deployed a single point of failure with a clinician attached to it.

These three questions travel. They work for a preceptor deciding what a student may use on a rotation, for a governance committee scoring a procurement, for an entrepreneur designing a product, and for a consultant defining a standard. The risk is the same, and the same test applies to all four tables that doctoral PAs sit at.

Four Tables, One Accountability

For the faculty member and preceptor: the question is no longer whether to allow AI, but where in the sequence it belongs. The paper’s three‑phase model: establish AI‑independent competence first, then teach calibration through adversarial practice (where students must catch deliberately planted AI errors), then integrate collaboration under supervision, is the most concrete curriculum guidance available.¹ It pairs with the DEFT‑AI supervisory framework from the New England Journal of Medicine, a structured way to turn an offhand “I asked ChatGPT” into a teaching moment.² Across health-professions education, AI literacy and responsible use are rapidly becoming core competencies rather than optional electives, and PA programs cannot be an exception. Yet neither framework was developed for PA education, and that is the gap we are now responsible for closing.

For the clinical leader: every tool you bring onto a unit becomes part of the hidden curriculum your students and new graduates absorb. Workflow integration and education are not separate decisions. They are the same decision, made once, with downstream effects on everyone who learns in that environment.

For the C-suite advisor and governance member: the issue is this: current definitions of a “competent” clinician say nothing about how to demonstrate that competence when AI is involved.¹ The FDA does not address educational use in its oversight of clinical decision support.¹ That leaves a gap in governance, and gaps are filled by whoever shows up. A PA who can clearly articulate the risk of never-skilling in a procurement discussion is contributing expertise that most others in the room are lacking.

For the entrepreneur and the legal or regulatory consultant: the paper lays out what's at stake downstream; licensing, liability, and supervision costs, whether credentials still mean the same thing across settings, and a workforce that quietly splits in two: clinicians who can practice without AI, and clinicians who can't.¹ Every one of those is a business question and a standard-of-care question. The builders and consultants who understand never-skilling before it becomes case law will be the ones who set the language everyone else ends up using.

The Opening

Two-thirds of U.S. physicians reported using AI in 2024, a 78% increase in just one year.⁶ Yet fewer than 15% of students and faculty report formal AI expertise or proficiency using it.¹ AI adoption is sprinting while AI preparation takes a leisurely walk. That gap is not simply a problem to lament; for the PA profession, it is an opening, and one we should move quickly to claim.

My first piece for this community argued that healthcare AI needs clinicians who can connect innovation to frontline reality, and that PAs are built for the role. Never-skilling is the first concrete test of that claim. It is not a distant, technical concern. It is a decision being made today, quietly, in classrooms, clinics, and boardrooms, about whether the next generation of clinicians will be able to think without a machine in the loop.

The invitation in my last article was to step into the conversation. This time, the message is more direct: you are likely already in it. If you teach, precept, govern, advise, or build, you are already determining how AI is sequenced into the formation of clinicians. The only question is whether you do so deliberately or by default.

The doctoral PAs who learn to ask, “Build or borrow? Can the user check it? Does it survive the unplug?” will not only protect their learners, they will also become the people their institutions rely on when standards are written. That window is open now, but it will not stay open long.

The views expressed in this article are strictly those of the author and do not represent the Department of Veterans Affairs or the U.S. government.

References

1. Ke Y, Jin L, Ong JCL, et al. AI-induced never-skilling in medical education. Nat Med. Published online May 22, 2026. doi:10.1038/s41591-026-04438-y

2. Abdulnour REE, Gin B, Boscardin CK. Educational strategies for clinical supervision of artificial intelligence use. N Engl J Med. 2025;393(8):786-797. doi:10.1056/NEJMra2503232

3. Budzyń K, Romańczyk M, Kitala D, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10(10):896-903. doi:10.1016/S2468-1253(25)00133-5

4. Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI without guardrails can harm learning: evidence from high school mathematics. Proc Natl Acad Sci USA. 2025;122(26):e2422633122. doi:10.1073/pnas.2422633122

5. Kosmyna N, Hauptmann E, Yuan YT, Situ J, Liao XH, Beresnitzky AV, Braunstein I, Maes P. Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv. Published online June 10, 2025. Revised December 31, 2025. doi:10.48550/arXiv.2506.08872

6. Henry TA. 2 in 3 physicians are using health AI -- up 78% from 2023. American Medical Association. Published February 26, 2025. Accessed May 25, 2026.