Artificial intelligence is no longer arriving in medicine. It has arrived. The American Medical Association’s 2026 Physician Survey found that 81 percent of physicians now use AI in their practices, more than double the 38 percent who reported using it in 2023. In three years it went from a minority tool to a default one, concentrated most heavily in documentation and summarization, the workflows closest to how we think.

The same survey carried a quieter number. Even as adoption surged, 88 percent of physicians said they were concerned about the erosion of clinical skills, a worry most acute among those earliest in their careers. Those two figures belong together. We are adopting a technology at remarkable speed while openly fearing that it may weaken the very judgment medicine depends on.

That fear is well placed, and it points to something the current conversation tends to skip. Before we ask what Artificial Intelligence can do, we have to protect another kind of intelligence first.

The intelligence that must come first

I call it Actual Intelligence, deliberately set against Artificial Intelligence, because the contrast is the whole point. By Actual Intelligence I mean the clinician’s trained capacity to listen, observe, examine, reason, sit with uncertainty, test assumptions, and remain accountable for what happens next. It is not intuition alone, and it is not unassisted brilliance pitted against technology. It is disciplined clinical method, practiced in sequence.

It is the intelligence that hears a patient say something still doesn’t feel right and does not simply reassure because the labs are normal. It is the intelligence that keeps the differential open long enough to ask what does not fit. It is the intelligence that knows a treatment plan is not complete merely because it is medically sound. It is complete only when it can survive the patient’s actual life, with this body, this history, this burden, this uncertainty.

Artificial Intelligence can extend that work. It can organize information, draft notes, summarize records, surface possibilities, and lift documentation burden. Used well, it can even hand some attention back to the patient in the room. But it cannot do the work that gives those outputs meaning. AI does not know what the patient meant to say but could not yet explain. It does not know whether a fluent summary quietly dropped the hesitation, the fear, the changed symptom, the question the patient almost asked. It does not know whether a recommendation fits this patient. Those judgments cannot be outsourced. They are the work of clinical responsibility.

The danger is not error. It is plausibility

This is not the familiar augmentation-versus-replacement debate. Much of the current conversation, rightly, focuses on governance, oversight, and who is accountable when an algorithm is wrong. I want to add something that happens earlier, inside the encounter itself.

The danger is not mainly that AI will be wrong. In many cases it will be directionally useful. The deeper danger is that AI will be plausible. Plausible enough to narrow the differential too soon. Plausible enough to make uncertainty quieter. Plausible enough to turn an incomplete encounter into a confident-looking note. AI arrives with the appearance of completion: fluent language, orderly summaries, ranked differentials, recommendations formatted with confidence. But clarity of output is not the same as clinical truth. A clean summary is not necessarily a faithful one. A reassuring risk score is not necessarily reassurance.

Consider how ordinary this looks. A patient describes dizziness, a heaviness that sleep does not relieve, and a quiet sense of foreboding he cannot name. He has already run his symptoms through an AI checker before the visit, not for a diagnosis, but for language. The labs come back reassuring. When he says it still doesn’t feel right, he is gently encouraged not to over-interpret the tools. The note is accurate. The reasoning is defensible. And yet his uncertainty was answered with a probability rather than explored with a question. The summary organized the story before anyone verified its meaning.

In a system already pressured toward speed, documentation, and closure, that is how plausibility gets mistaken for completion. When a tool offers a coherent first draft, the clinical mind is tempted to edit rather than originate, to confirm rather than re-examine, to accept a frame before testing whether it belongs. This is how the skill erosion that 88 percent of physicians fear actually happens, not through one dramatic mistake, but through repeated small substitutions: fewer independently generated differentials, fewer deliberate re-checks, less tolerance for the open question.

The order is the safeguard

The principle is simple, and it is a safety requirement rather than a sentiment: Actual Intelligence first, Artificial Intelligence second. The clinician listens, examines, reasons, and names what is still uncertain. Then AI can test completeness, organize information, prompt what may have been missed, and support the documentation. And when an AI-generated addition does not fit the patient, it should be disregarded, not politely retained because it is plausible, efficient, or well phrased. That act of disregarding is not resistance to technology. It is clinical judgment, and it is only possible if the clinician has done the thinking first.

This reframes the goal. The measure of good care will not be how seamlessly technology is integrated. It will be how deliberately judgment, verification, and accountability are preserved.

A prior question

This is one of the principles I develop in Care-Full Medicine, which is not anti-AI. It asks a prior question: What clinical method is AI being asked to amplify?

If the method is compressed, rushed, closure-driven, built around throughput, AI will scale the compression, making a thin encounter look complete. If the method is Care-Full, if listening stays active, reasoning stays understandable, uncertainty stays named, and the plan is tested against the patient’s real life, AI will strengthen it.

This is why the 81 percent figure is, by itself, neither good news nor bad. Adoption is now a settled fact; the sequence is not. An 80-percent-adopted technology layered onto a compressed method will scale that compression across the whole system. The same technology layered onto a Care-Full method could make medicine more attentive than it has been in years.

The future of medicine will not be decided only by how powerful Artificial Intelligence becomes. It will be decided by whether Actual Intelligence remains first.

Alan P. Feren is a retired surgeon, independent physician, health care consultant, and patient advocate with more than 50 years of experience in clinical practice, system leadership, and health care innovation. Formerly in academic and community surgical practice, he has worked across the evolving landscape of managed care and clinical governance.

In the 1990s, Dr. Feren co-authored clinical guidelines that evolved into what is now MCG Health, now used by more than 80 percent of U.S. health plans and over 3,100 hospitals. He has advised health technology startups, helped shape managed care policy, and served as a clinical content developer for health care technology platforms.

His work centers on restoring shared understanding between clinicians and patients in an era defined by speed, fragmentation, and technological mediation. Drawing on both professional experience and his own journey as a complex patient, he writes about transparency, accountability, and the disciplined methods that make medical care trustworthy. He is a contributor to KevinMD and a podcast guest. More information is available at mypersonaladvocate.net and on LinkedIn.