Outside the Training Data

Epistemic status: The history comes from the published surveys, the hospice movement’s own record and Medicare’s numbers, all sourced below. What I say about how models learn is my understanding as an operator who builds with them, stated plainly so an engineer can argue with it. Section 07 is the strongest case against all of it.

In 1961 a Chicago doctor named Donald Oken asked 219 physicians what they did when they found cancer. Nine in ten said they usually didn’t tell the patient. Some changed the diagnosis they gave, so the word cancer never came up.

Sixteen years later a research team gave the same questionnaire to 699 doctors at a university hospital. Of the 264 who answered, 97 percent preferred to tell. The authors called it a complete reversal, which is about right.

01

Sure, Then Sure of the Opposite

The reversal isn’t the strangest thing in the second paper. In both surveys, the reason doctors gave most often for their policy was their own clinical experience. In 1977 they also gave more credit to what medical school and their hospital training had taught them. Their training had changed, and their answer changed with it.

One question, asked twice: do you tell the patient it’s cancer?

1961preferred not to tell90%

1977preferred to tell97%

1961: 219 physicians in Chicago. 1977: the same questionnaire to 699 doctors at a university hospital, 264 answered. Both times, the reason given most often was the doctor’s own clinical experience.

The authors closed with a line I keep coming back to. Physicians, they wrote, were still basing their policies on “emotion-laden personal conviction.” So the answer flipped, but the doctors were just as sure the second time, and for the same kind of reason.

The question that sent me back to these two papers came from an emergency physician who builds clinical tools. Take a model and train it on everything medicine believed in 1972. Then ask it how to care for a patient who is dying.

02

The Model From 1972

1972 sits in the middle of the first reversal. A model trained on that year’s medicine would have read decades of doctors explaining why they kept the diagnosis to themselves, and only a few years of the other side. It would lean toward the old answer, because the old answer was most of what had been written down.

On dying, it wouldn’t have to lean. The other side barely existed. America’s first hospice was still two years away. That year Elisabeth Kübler-Ross, whose On Death and Dying had come out three years before, testified at the first national hearings on death with dignity. Medicare wouldn’t pay for hospice care for another eleven years.

So ask the 1972 model how to care for a dying patient and you’d get the standard of the day, fluent and complete. It wouldn’t be lying. Almost everything it had read agreed with it. And the work I do every day, hospice, wouldn’t be in the answer at all, because in 1972 it was barely in anything.

03

The Heresy

Modern hospice started in London, with someone who had worked at the bedside before she was a doctor. Cicely Saunders qualified as a nurse in 1944 and worked as a medical social worker before she went to medical school. She qualified as a doctor in 1957. Ten years later she opened St Christopher’s Hospice. Every hospice I’ve run traces back to what she started.

In 1963 she brought the idea to Yale. Florence Wald, the dean of Yale’s School of Nursing, took it seriously enough to resign the deanship and build an American version.

On June 11, 1974, Connecticut’s Commission on Hospitals and Health Care approved construction of the first hospice facility in America, in Branford. The building cost 2.6 million dollars. That year Wald, with two pediatricians and a chaplain, founded Connecticut Hospice.

What they built ran against how America handled dying. As late as 1980, 54 percent of Americans who died, died in a hospital, and 17 percent died at home. And in 1961, nine in ten of the doctors Oken asked weren’t telling patients the diagnosis. Hospice turned that around. The patient knew what was happening, she could be at home, and the family was part of the care.

04

From Heresy to Half

In 1982 Congress created a Medicare hospice benefit. It started paying in 1983, and a budget law made it permanent a few years later.

Here is where it stands now. In 2024, 1.82 million Medicare beneficiaries received hospice care from 6,706 hospices, and Medicare spent 28.3 billion dollars on it. Of the Medicare beneficiaries who died that year, 52.9 percent used hospice, a new high. In 2010 it was 43.8 percent. HealthView is one of those 6,706 hospices.

Hospice, from heresy to half

Outside the consensus

1. 1957Cicely Saunders, a nurse first, qualifies as a doctor.
2. 1963She brings the idea to Yale.
3. 1967St Christopher’s Hospice opens in London.
4. 1972Kübler-Ross testifies at the first national hearings on death with dignity. The year the model is trained on.
5. 1974Connecticut approves America’s first hospice facility, in Branford. Connecticut Hospice is founded.

Inside it

1. 1982Congress creates a Medicare hospice benefit.
2. 1983Medicare starts paying for it.
3. 201043.8 percent of Medicare decedents use hospice.
4. 202452.9 percent, a new high: 1.82 million beneficiaries, 6,706 hospices, 28.3 billion dollars.

Fifty years ago this was a nurse’s conviction and a plan for a building in Branford. Now more than half of the people on Medicare who die spend the end of their lives in it. A model trained in 1972 would have called that a fringe idea, because in 1972 it was one.

05

What a Model Learns

I build with these models, so here is the plain version of how I understand them. A language model learns from what people wrote, and it learns most from what they wrote most. That’s why it’s useful. Ask it a question medicine has settled and you get the settled answer, because the settled answer is everywhere in what it read.

It’s also why it can’t see a reversal coming. A heresy is the thing almost nobody has written yet. In 1972 the case for hospice had been made in a handful of lectures and books. The case for the way things were done was nearly everything else. A model weighs an idea by how often it appears, and nearly all the weight was on one side.

So a model’s confidence tells you how many people agreed, which isn’t the same as whether they were right. Nine doctors in ten agreed in Oken’s survey. Train on them and you get nine-in-ten confidence in an answer the profession dropped within twenty years. A model can learn what a profession agreed on. It can’t learn the disagreement nobody has written down yet.

Doctors learn this way too, which is what that detail in the 1977 survey was saying. They gave their training more credit for their answer than the 1961 doctors had.

06

The 1961 Question, Again

The question from 1961 is back, about a machine this time. In 2023 Stanford tried a model that drafts replies to patient messages inside the health record. 162 clinicians used it, and a clinician reviewed and edited every draft before the patient saw it.

Should the patient be told a machine wrote the first draft? A team at Duke asked 1,455 people. They preferred the AI drafts, mildly. But when they were told AI was involved, their satisfaction dipped slightly.

That’s the 1961 argument with a new subject. Telling makes people feel a little worse, so don’t tell. The Duke team measured the cost of telling and came down where the 1977 doctors did. They recommended keeping the disclosure anyway, to uphold patient autonomy.

California has answered in law. Since January 2025, an AI-generated message to a patient about their clinical care has to say so and tell them how to reach a human, unless a licensed clinician read and reviewed it first. With that review, the patient doesn’t have to be told.

Other states have gone further. Texas requires providers to tell patients when they use AI in diagnosis or treatment. This year Rhode Island became the first state to require notice when AI documents a patient visit, and Louisiana passed a similar law. Illinois, Maine and Colorado require a patient’s consent before AI is used to support their therapy. Utah requires disclosure in what it calls high-risk interactions, and health counts. By my count, at least eight states now require telling patients about AI in some part of their care.

States that require telling patients about AI in their care

StateWhat it requiresSince

CaliforniaAn AI-written message about a patient’s clinical care says so and tells them how to reach a human, unless a licensed clinician reviewed it first2025

TexasTell patients when AI is used in diagnosis or treatment2025

UtahDisclose AI in high-risk interactions, and health counts2025

IllinoisA patient’s consent before AI supports their therapy2025

Rhode IslandNotice when AI documents a patient visit2026

LouisianaNotice before AI records a visit to write it up2026

MaineA patient’s consent before AI supports their therapy2026

ColoradoA patient’s consent before AI records their therapy2026

8 now, of the 15 the Ledger call needs by the end of 2028

So far the states are landing where the 1977 doctors did.

07

The Strongest Case Against This

Four objections, as strong as I can make them.

One. Models aren’t frozen in 1972. They get retrained, and you can hand one this month’s journals. It would learn a reversal as soon as it was published. True, and that’s the limit I’m describing. A model learns a reversal once enough people have written it down, which is after it happened. An operator needs to see it while it’s still a minority view, and a minority view is a small part of any training set.

Two. I picked two reversals that won. Plenty of bedside convictions deserved to lose, and a model that sides with the consensus is right most of the time. So is a doctor who does. Agreed. The question is which parts of the consensus rest on evidence and which rest on habit, since the habits are the ones that tend to flip. The evidence share is smaller than most people assume. A review of 6,329 recommendations in American and European cardiology guidelines from 2008 to 2018 found that 8.5 percent of the American ones rested on the top level of evidence, multiple randomized trials or one large one. 41.5 percent rested on the lowest level, limited data or expert opinion.

Three. Maybe the reversal on telling had little to do with the bedside. Maybe it came from lawyers, from patients demanding their rights, or from treatments that made a cancer diagnosis less hopeless. Maybe. I’m not claiming nurses did it alone. My claim is narrower. Whatever caused it, a model trained before it couldn’t have seen it coming.

Four. AI isn’t cancer. Keeping a diagnosis from a dying man is a lie about his life. Not mentioning that software drafted a message his doctor reviewed is closer to not mentioning the dictation software. This is the one I take most seriously, and part of it is right. The stakes aren’t the same. But the argument for keeping quiet is the same one, that telling makes people feel worse, and that argument lost the last time. The researchers who measured the cost of telling still recommended doing it.

08

What to Do With It

If you run a care company, here is what I take from this.

Sort your defaults. Every company runs on assumptions nobody has checked in years: how we schedule and what we say at the bedside. Some rest on evidence, and some on how it has always been done. The second kind is where the next 1961 is hiding, and it’s also where a model will sound most sure of itself, because how it has always been done is what fills the training data.

Use the models where the consensus is solid. Documentation, coding, scheduling, eligibility, the parts of the work where this year’s right answer will still be right next year. Be slower to let one settle a question your field is still arguing about.

Keep the bedside in the room. Hospice was started by people who had stood at the bedside, a nurse who became a doctor and a nursing dean who gave up her deanship. In a care company, the nurse or aide who says we’re doing this wrong is the earliest warning you have. Whether she says it depends on your culture. If speaking up has ever cost someone at your company, she won’t, and you’ll find out late. When a bottleneck or something that went wrong is seen as an asset, and a nurse leads the fix, the fix usually corrects itself, because it’s in the best interest of the staff and the patient.

And when you’re unsure whether to tell a patient something, tell them. The doctors got there by 1977. The researchers measuring AI disclosure got there in 2025.

09

The Ledger

Every field note that makes a prediction puts a dated line on the public ledger. This is the tenth. My bet is that telling wins again, one state at a time.

10

Ways I’m Wrong

The model argument is a simplification. Models are more than vote counters, and a good one will make the case for hospice in 1972 terms if you ask it to. My claim is narrower. Asked for the best answer, a model gives you the consensus, and in 1972 the consensus had no hospice in it.

The two surveys aren’t the same doctors. Oken asked physicians in Chicago. The 1977 survey went to a university hospital, and only 264 of the 699 answered. Some of the reversal could be who answered. A swing from ninety percent one way to ninety-seven percent the other is too big for that to be the whole story.

The prediction could miss for reasons that have nothing to do with patients. Late last year an executive order directed the Justice Department to challenge state AI laws it considers burdensome, and drafts in Congress would preempt some state AI laws. If Washington takes this away from the states, the count stalls and I’ll score a miss. I could also be counting generously, since some of the eight cover a narrow slice of care, like therapy sessions.

And the biggest one. AI disclosure is the 1961 question in a new setting, which makes it easy to spot. The next real reversal in my field is probably something nobody is arguing about yet, and that means my best guess about it is the one most likely to be wrong.

I run my companies on plenty of things I’m as sure of as those doctors were in 1961. Some of them are wrong, and I don’t know which ones. If I find out early, it’ll probably be because one of our nurses told me.