Keep your head down. Don’t ask too many questions. Just finish your note and move on.

I heard a version of this command at every level of training and well into my attending practice, from colleagues, seniors, teachers, and administration. Most delivered the advice like it was some sort of hard-won wisdom, though it wasn’t. It was the sound of people who repeated the system’s rules because it felt like control.

Lately, I’ve noticed this same attitude creeping into the conversation regarding AI in medicine.

Don’t fight it, it’s inevitable. Better to be at the table than late to the party. Same cadence, delivery, and lack of understanding. Most of these people have no more say over how these tools are being built and deployed than any of us did over how our residency programs were run.

We keep talking about AI in medicine like it’s a new beast to apprehend and understand. In reality, it’s the same problem in a new skin, moving much faster than the versions before it.

Look at the shape of control in medicine over the last few decades, and a pattern becomes obvious long before anyone even mentioned large language models. The Match tells you where you’ll train, with no leverage to negotiate and severe consequences to walking away. Productivity metrics tell you how quickly a clinic visit should occur. RVU models tell you what your time is worth, determined by someone who has never sat face-to-face in a clinic with a patient. Prior authorizations tell you which treatment you’re allowed to provide, regardless of what your clinical knowledge and the standard of care dictate.

Here’s what’s different about this round: Artificial intelligence tools now sit, or are about to, directly inside the zone of clinical judgment and decision-making, the one territory that used to be ours by default. The modern electronic health record governs our schedules. Insurance determines our offerings. But the actual thought — that used to be the physician’s alone, for better or worse.

AI won’t ask permission to enter that space. It will simply be there, generating notes and follow-up before you’ve even finished shaking your patient’s hand. And the tools were built, trained, and deployed almost entirely without physicians at the helm. According to the AMA’s own 2026 survey, 81% of physicians are already using AI professionally — more than double the rate three years ago — while 85% say they want a real voice in how it gets adopted in their own practices. Take in these two numbers together to appreciate the actual story: Adoption raced ahead of consent.

At this point, most AI tools are being used in the name of efficiency. AI can help write notes for a system that none of us like in order to free up time to see a few more patients that we won’t get compensated for at nearly the same rate as our predecessors. The degree of improved efficiency remains an area of quantification: A JAMA study published in April reported that ambient AI scribes saved about 16 minutes of documentation time per eight hours of patient care. This comes out to saving about two minutes per hour of patient care.

And the days are coming when AI is meaningfully involved in clinical care, beyond just transcribing our own words, and the follow-up recommendation deviates from the standard of care. Who will be blamed for the missed diagnosis: the trillion-dollar AI company trained on the whole of human knowledge, or the physician whose name is at the top of the chart?

In January, the FDA narrowed the definition of what counts as a regulated device in clinical decision support, meaning that a wide swath of AI tools face less federal oversight as long as the clinician independently reviews the recommendation. The result is decreased scrutiny of the tool, combined with increased liability for the person clicking “accept.” This is not a story of technological triumph. This is the same organizational instinct that has governed medicine since before I began medical school in 2013: Risk will be shouldered by the people with the least amount of power to resist it.

Liability will get the most attention because there will be lawsuits at the end of it. But it’s a symptom, not the disease. The real question isn’t who takes the blame when judgment fails. It’s whether physicians are still the ones exercising judgment at all.

Physicians still pursue deep mastery in an age when most facts are a simple search away — because mastery was never really about the facts themselves. It was about independence. A physician who has truly internalized the craft doesn’t need to be told what to think about something new. They can sit with an unfamiliar approach, evaluate it at a distance, and decide for themselves whether it holds up, on their own timeline. That’s the difference between evaluating a tool and simply being issued one. Most of us are not afforded that luxury anymore.

This isn’t an argument against using the tools. I use AI mostly for automating repetitive tasks and helping with time management, and it’s often genuinely useful when outputs are critically evaluated appropriately.

But there remains a real risk in accepting deferral without evaluation: the difference between reaching for a tool because I’ve decided it’s useful versus reaching for it because it has become the path of least resistance. An institution only needs to make the second version of reaching easier than the first, often enough, for long enough, that physicians stop trusting their own judgment and become reliant on a shortcut. Nobody will strip our autonomy from us. We’ll just stop reaching for it.

I’m not saying physicians should bury our heads in the sand about AI. That won’t help, and pretending otherwise is its own kind of denial.

But we — or you, since I’m no longer practicing — should refuse the framework that comes attached to it: that AI adoption is inevitable, someone else will ride in to insist on more oversight and better protocols, and physicians should simply be grateful for the efficiency AI has granted us while quietly absorbing whatever risk trickles down.

AI is a mirror. It’s showing us, with unusual clarity and speed, exactly how little say we’ve had in our profession for a very long time, and how practiced we’ve become at simply complying anyway.

The reflex was trained into us. It can be untrained. Just don’t wait for AI to do it for you.

Frances Mei Hardin, M.D., is a writer, consultant, and former ENT surgeon.