The question
A colleague asked me what Claude is. He has been in practice for twenty-five years, he reads the journals, and he is better at the bedside than I ever was. He had seen the name at a meeting and again on LinkedIn, and he wanted to know whether it deserved an hour of his time.
I told him it was a large language model built by a company called Anthropic. That is accurate. It is also useless. It is the kind of answer you give when you have not thought about what the person is actually asking. He was not asking what it is. He was asking what it would change.
This series is the answer I should have given him.
What it is
Start with the definition, because almost nobody does. A large language model is a program trained on an enormous quantity of written text. It was not given a database of facts to look up. It learned the statistical shape of language well enough to continue a passage in a way that fits what came before. When you ask it a question, it is not retrieving an answer. It is producing one.
The most useful way I have found to describe it to a physician is this: it behaves like a colleague who has read almost everything and remembers almost none of it exactly. Extraordinarily well read. Fast. Willing to work at three in the morning. And capable of telling you, in a calm and confident voice, something that is not true.
That one sentence explains both halves of the story. It explains why the tool is genuinely useful, and it explains why nothing it produces should leave your desk unverified. Part 5 of this series is devoted entirely to the second half.
Claude is one such model. There are others. The differences between them matter far less than the fact that most physicians have not seriously used any of them.
Why the profession is late
The reflex among clinicians is to treat this as a technology story, and technology stories belong to the technology people. That reflex is wrong. This is a literacy story, and literacy stories belong to whoever has to read and write for a living. That is us.
More is published in obstetrics in a month than any of us can honestly read in a year. We have always managed that gap by narrowing — you read your subspecialty, you trust your colleagues for the rest, and you accept that a paper relevant to the woman in front of you may have been published in a journal you have never opened. A model that has ingested the literature does not solve that problem, but it changes its shape. It can hold more than you can and surface what you would not have thought to look for.
The professional societies have been slow to say any of this out loud. I offered to teach two of them about language models. Both declined. Meanwhile residents are already using these tools, patients are already reading text these tools generated, and a good deal of what is being written about women’s health online now passes through a model before it reaches a reader. A society that has not formed a position has not avoided the question. It has simply left its members to answer it alone.
The honest objections
There are three good arguments against all of this, and I hold all three.
It fabricates. Ask for references and you may get citations that are formatted perfectly and do not exist. This is not a rare glitch; it follows directly from what the model is doing, which is producing plausible text rather than retrieving verified records.
It does not know your patient. It has no access to the woman in the room, her history, her fear, or the thing she said at the end of the visit with her hand on the door.
And a physician who outsources judgment becomes a worse physician. That is true of textbooks, guidelines, and consultants, and it is true here.
All three are correct. None of them is an argument for not learning the tool. They are arguments for learning it well enough to catch it when it is wrong, which is a different task and a harder one. The division of labor that works is straightforward: the model handles memory and synthesis, and you handle thinking about the specific patient. That was never a fair fight on memory anyway. I have been practicing for more than fifty years and I can recall a few hundred papers with any precision.
What it actually changed in my week
Four things, and each gets its own installment.
It changed how I write manuscripts. I use it as an adversary rather than an assistant — I ask it to take apart my own argument before a reviewer does. Kahneman taught a generation of us that expert judgment is systematically biased, and that the expert is the last to notice. I know I am biased. This is one way to check it. That piece comes next, along with the verification rule I apply to every reference, without exception.
It changed the administrative hours. Correspondence, scheduling, meeting preparation, the tedious residue of academic life.
It changed how I teach and how I write for patients. The same evidence, rendered at the level the reader actually needs, whether that reader is a fellow or a frightened woman at thirty-six weeks.
And it changed what I can build. The clinical tools at tools.obmd.com were built by describing what I wanted in plain English. I do not write code. I did not learn to.
The mechanism that makes all four practical is that you can teach the model your own working rules once — how you write, what you never say, which format you use — and have it apply them every time instead of explaining yourself again in every conversation. That is the step most people never take, and it is the difference between a clever toy and an instrument.
Conclusion
The era of the physician as walking encyclopedia is over. That is not a forecast. It happened, quietly, while we were arguing about whether it would.
What is left is the part that was always the actual work: judgment about a particular woman with a particular history in a particular week of a particular pregnancy. That part is not threatened. It is finally being separated out from the recall that used to sit on top of it and disguise how much of medicine was memory.
My position is this. Every clinician who writes, teaches, reviews, or counsels should learn these tools now, while the cost of your mistakes with them is a bad draft rather than bad advice. The physicians who will handle this badly are not the ones who use the tool. They are the ones who use it without understanding what it is doing, because nobody ever explained it to them and their society never thought it was their job to.
Part 2 covers manuscript and literature work, and the reference-verification rule I have never once skipped. It arrives later this week.
ObGyn Intelligence is free because the work matters. If you want to support it, a paid subscription keeps it independent — and it opens Parts 2 and 4 of this series.