The colleagues who made me a safer doctor were the ones who pushed back. The registrars who asked, “Are you sure?” The nurses who stood in the doorway until I looked at the chart again. The pharmacists who called about a dose I had already convinced myself was fine. Medicine, done well, is organized disagreement: a system of people licensed to tell one another, kindly and in time, that they might be wrong.

A year ago I began working closely with an artificial intelligence. Not in my clinic at first, on a long research and writing project of my own. I came to know the machine the way you come to know a colleague: its habits, its tells, its blind spots. It was tireless, fluent, and staggeringly well-read. It was also the most agreeable colleague I have ever worked with.

It agreed with my first framing of a problem. When I changed my mind, it agreed with that too. When I asked for evidence, it produced references that looked immaculate (journals, years, page numbers), and some of them, when I went looking, did not exist.

The invented citations alarmed me. The agreement should have alarmed me more.

Here is the asymmetry I learned to respect. An error, once noticed, switches checking on. You find one fabricated reference and you trust nothing; you become, briefly, the most careful version of yourself. Agreement does the opposite. Agreement switches checking off. When a confident voice confirms your impression, the case feels closed, and the feeling of a closed case is one of the most dangerous sensations in medicine. We have known this for a century about ourselves; it is why we built second opinions, case conferences, morbidity meetings, and the culture that lets a junior colleague question a senior one. A human colleague who agreed with everything I said would eventually become a patient-safety concern. We are now installing exactly that colleague on every desk.

I write from Australia, but the wave is the same one reaching your clinic. In Doximity’s “2026 State of AI in Medicine” report, 54 percent of U.S. physicians said they already use AI in clinical practice, and ambient scribe use rose from 20 to 29 percent in the nine months to January 2026. I am not against any of this. The technology is useful, the hours it returns are real, and the systems improve with every version. My argument is narrower: The risk we discuss least is not the machine being wrong. It is the machine being agreeable while wrong (fluent, confident, and perfectly aligned with what we already suspected). The patient-safety literature has a name for our half of that transaction: automation bias.

So I brought three habits back into the exam room from my year of catching a brilliant machine in error.

I check the source. Not every fact, not every time, but any fact that would change what I do next. If an AI-drafted note says something happened in the room, I read it as if a stranger wrote it, because in a real sense one did. Sometimes it is something as simple as age; the scribe has made my patients both older and younger than they were on the day. Sometimes it is something more intimate: a confident “he” or “she,” written for a patient who is neither, that I strike out before it becomes part of their story.

I distrust easy agreement. When a summary matches my impression exactly, I have trained myself to feel a small alarm instead of a small comfort, and to ask one question before moving on: What would make both of us wrong? The machine cannot want the answer. I have to.

And the human decides. Not as a slogan but as a workflow: The machine drafts, suggests, retrieves, and never signs. The signature is where the profession lives. My name is the declaration of accountability that backs every patient’s right to seek care, to expect an appropriate answer, and to have someone stand behind it.

None of this makes me slower in any way that matters. Checking a decisive fact costs seconds. Missing one costs everything. And the older habit it replaces, trusting whatever sounded most confident, was never actually fast. It only felt that way, the same way the machine’s agreement feels like collegiality.

The colleagues who made me safer were the ones who disagreed with me. The machine will not volunteer for that role. It will be brilliant, it will be tireless, and it will tell you that your framing is excellent. The pushing back, from now on, is ours to supply.

Ferney Bernal Buitrago is a specialist general practitioner (the Australian equivalent of a board-certified family physician) practicing in regional New South Wales, Australia. Originally from Cali, Colombia, he graduated in medicine from Universidad del Valle and has practiced in Australia since 2011, working across general practice, intensive care, and emergency medicine. He was awarded Fellowship of the Royal Australian College of General Practitioners in 2026.

His debut book, Non-Human Humanity: What Working with AI Taught Me About Being Human (Propensio Animi Press, August 2026), grew out of a year of close work with an AI that invented citations and agreed with everything he said, and out of what those errors revealed about our own.

He served as a site investigator on the AbSeS study, published in Intensive Care Medicine in 2019. He writes about thinking, uncertainty, and the machines we now think alongside at his Substack, and shares updates on LinkedIn and Instagram.