At 11:40 p.m. on April 14, 1912, the Titanic hit an iceberg. She didn’t go under until 2:20 a.m. That’s two hours and forty minutes in which roughly 2,200 people lived between the impact and the catastrophe, and for most of that time the majority of them didn’t believe anything serious had happened.
You can see why. The lights stayed on. The cabins were warm. The deck felt level. A steward telling you to put on a life vest and climb into a small open boat, then be lowered sixty feet to the freezing Atlantic, sounded like the irrational option. The ship was the safe place. Everyone knew that.
So the early lifeboats went down half empty.
Lifeboat No. 1 had room for 40 and reportedly left with 12. Lifeboat No. 6 had room for 65 and carried about two dozen. By the time the bow was visibly under water and disbelief turned into panic, the last conventional lifeboat had already been lowered, at about 2:05. Fifteen minutes later there was no ship.
The people on board didn’t fail to notice danger. They spent the only time in which noticing could still have changed anything.
Medicine is in that interval now
AI has already hit the profession. It’s changing how patients get information, how we search the literature, how notes get written, how differentials get generated, and how decisions get explained to the people they affect.
I’ve been writing about this for a while. We published the first paper on AI in ObGyn in the American Journal of Obstetrics & Gynecology more than three years ago, and again a few months ago, and more than ten articles in between. I also spend a good part of my week vibecoding clinical tools with it, which is not something I expected to be doing in my seventies.
And yet a large share of physicians, obstetricians and gynecologists very much included, behave as if nothing structural has changed. They warn about the dangers of AI and stop there. They dismiss it as unreliable. They find one hallucination and treat it as if it settles the whole question. Some ridicule the colleagues who are learning to use it. Most assume that thirty years of training will keep being enough without learning how this particular tool works, and they’re waiting for a hospital, a professional society, a regulator or a medical school to tell them what to do.
The exam rooms are full. The ORs are running. The lights are on. It feels safe to stay where we are.
Our patients aren’t waiting for us
Patients are already using AI to read their lab results, their ultrasound and pathology reports, their medication lists and the treatment plan we handed them. Pregnant women are asking it about induction, cesarean, prenatal testing, fetal anomalies, vaccines and birth plans, usually the night before the visit.
Some arrive with information that’s wrong or dangerously oversimplified. Others arrive with sharper questions than we’re used to: a paper from last month, a management option we hadn’t mentioned, a risk estimate we hadn’t calculated. Both kinds of patient need a clinician who can tell which is which.
That’s the actual problem. Patients are using these systems while many clinicians can’t judge the output’s accuracy, can’t see what it left out, can’t correct it, and don’t use it themselves. A physician who refuses to learn AI isn’t protecting anyone from AI. He’s handing interpretation to the patient, the algorithm, a Facebook group, or whichever commercial product gives the most confident answer.
Our professional organizations haven’t helped. There’s no clear AI competency curriculum. There are the usual panels and slide decks at the annual meeting, where few people learn much, and then everyone goes home.
Clinical authority used to rest partly on the fact that the physician controlled access to medical information. That’s over. What’s left is the ability to evaluate information more critically than the patient can, apply it to this patient in front of you, say plainly where the uncertainty is, and explain why you’re recommending what you’re recommending.
Competence isn’t trust
I’m not asking anyone to accept AI output at face value, let a chatbot diagnose patients on its own, or replace the exam and the relationship with a screen. Competence means knowing enough to use the tool safely and to recognize when it shouldn’t be used at all.
In practice that means knowing what generative AI can and can’t do, how persuasive its fabrications are, how to check a claim against the primary source, how the answer changes when the prompt is missing clinical context, where bias enters through training data, how confidentiality gets breached, how automation bias erodes your own judgment, where AI helps (documentation, patient education, evidence retrieval, translation), when a human has to review, and who’s accountable for the decision at the end. The last one has an easy answer. You are.
Nobody needs to become a programmer. Every physician needs to be able to supervise a tool that’s already influencing their patients and their practice.
Why obstetrics can’t afford to wait
Obstetrics is unusually exposed to confident, incomplete advice. Decisions are time-sensitive, risk estimates depend heavily on which population you’re in, the evidence is often thin, and every recommendation lands on two patients at once.
A generic AI answer on induction at 39 weeks, TOLAC, fetal growth restriction, aneuploidy screening, home birth or a medication in pregnancy can sound authoritative and still omit the one detail that decides whether it applies. If our patients are reading those answers, we have to be able to spot false certainty, missing contraindications, outdated guidance and invented references. We also have to learn where the tool helps us: finding evidence faster, preparing individualized counseling, checking whether a patient understood, surfacing the question nobody thought to ask.
Refusal isn’t a safety strategy.
The obligation has changed
Medicine has absorbed disruptive technology before. What’s different this time is speed, access, and the fact that patients adopted it first. AI didn’t need hospital installation or a department chair’s approval. It’s on every phone in the waiting room.
Whether AI enters clinical medicine stopped being a question a while ago. The open questions are whether clinicians will use it competently, whether institutions will teach it, and whether professional organizations will set standards before bad habits harden. Clinical AI competence belongs in medical school, residency, maintenance of certification, credentialing and patient-safety programs, and the teaching should cover how to interrogate and challenge the tool, not just how to operate it.
AI isn’t replacing clinical judgment. It’s changing what competent clinical judgment consists of.
The Titanic story isn’t about AI sinking medicine.
It’s about what people do with two hours and forty minutes when the lights are still on.
The worst outcome would be to wait until AI’s influence can no longer be ignored and then start preparing in a panic, lowering boats in the dark.
The iceberg has already been struck.