A finding from our own data that I have been thinking about.

When a patient loses decision-making capacity, the proxy they chose picks the treatment the patient would have picked 44.7% of the time. We gave a language model the patient’s own documented personal values, and it matched 72.6%.

The temptation is to read that as the algorithm knowing the patient better. But is it? If you strip the person’s own value profile, the model is back to the performance of family members, and it cannot infer values from the medical chart at all.

One of the books that formed me, long before medicine did, was Milorad Pavić’s Dictionary of the Khazars: a novel about long-gone Khazar people who survived only in other people’s books. Three lexicons, Christian, Islamic, Hebrew, were telling one story about them and never agreeing, while the Khazars themselves left not a single page of their own.

That is what the model met in the medical chart: a person recorded entirely in other people’s languages. It read every entry and could not find the person.

Pavić knew one more thing. Two of his characters dream each other, each awake inside the other’s sleep, so that a man’s death happens, most fully, in somebody else’s life.

That is what a proxy is: the one who is awake in your dying.

My father died in December of 2010. He spent a year failing, losing weight and forgetting himself. The last months were agony in his own bed, lost to himself and to us. My mother alone at the bedside in Bosnia, with no support but her neighbors, who would come running at two in the morning when she needed help lifting him back into bed.

He was fiercely independent, yet deeply frightened of illness and of dying. We never spoke of either. As his body and his identity kept disappearing, in his lucid moments he kept reminding me to be there for my mother and my sister. It comforted him to know I had proved my strength to him. He carried his regrets openly, the things he felt he had not done right. And in the long passages of time when his mind was leaving him, his body took over, and the fixation became the failing systems of that body. The pain of the body took over the person.

We never had the conversation. I stepped in, because he had asked me to be the strong one. My mother paid the physical price of being there.

Mientras que viva tu padre no estás en el mundo sola. While your father lives, you are not alone in the world.

So what would an algorithm have done for my father, when even we did not know what to do? And how many of us have the courage to ask that question of ourselves, or of the people we love?

The numbers have been telling us this for twenty years. When researchers pooled every study that asked surrogates to predict what patients would choose, across sixteen studies, more than 2,500 patient-surrogate pairs and nearly 20,000 paired answers, the surrogates were right 68% of the time. And here is the finding that should stop us: neither being formally chosen by the patient nor having talked with them beforehand improved that accuracy. Meanwhile only about one in three American adults has completed any advance directive at all.

We built an entire ethics of substituted judgment on a prediction that fails one time in three, made by people we mostly never prepared, about wishes we mostly never wrote down.

We may be told, by newly produced evidence, that algorithms are more precise. In a randomized trial published in 2024, a language model working alone outscored the physicians who were using their conventional resources, and not narrowly. When licensed clinicians blindly compared chatbot answers with physician answers to patient questions, they preferred the chatbot nearly four times out of five, and rated its answers empathetic almost ten times as often as ours. We may think, indeed, that AI is catching diagnoses we miss, avoiding errors we make, out-warming us on the page.

But patients do not present as written text or data tables or images on a screen. They arrive in their full human complexity, overwhelmingly unique every single time, often not knowing where to look or what to do. As a newly minted physician you learn to absorb that complexity: to find the patterns running closest to the common denominator, to fit the assembled codes of diagnosis, to reduce the complexity while expanding the humanity of the moment.

Yet the hardest thing in medicine is not always to be correct. But how often do we actually know what correct means? We lack fundamental understanding of so many biological processes, and of the several thousand rare diseases we have catalogued, fewer than one in twenty has an approved treatment.

The hardest thing in my thirty years of practice was to resolve the deepest desire not to be correct: the wish to say, you are fine, this is not what you fear, let me absolve you of pain and worry. You see the dark raven circling, and you accept the pain of the truth, and you serve it for the first time. You are deep in that moment, immersed in duende, accepting that you do not know why the universe has chosen this person for this suffering. Accepting is part of the human job. Holding the moment, attending to a fellow human soul, is the work we do every day, and we need to do it well, because there is nobody else to do it.

Years ago, a young mother was my patient in the ICU, her brain severely injured by a catastrophic complication of childbirth. Every data element, every image of her brain, every medical book pointed to the inevitable: she would not recover, and she would never be herself. Her father, himself a physician who had spent a career on the delivering side of exactly this conversation, with families of young people struck by unimaginable and random events, was on the other side now. He was on the receiving end, and he was not ready. Nobody is ever ready. I was not ready: I wanted to believe in the randomness of good as much as I knew the randomness of pain, and I wanted to give her as much time as she needed. Down the hall, her newborn was growing into a new human. The algorithm reading her data would have been certain that her case led to one outcome. None of us knew her values: she went to give birth thinking of baby food and toys. What evidence would the algorithm have produced for her? Keep supporting her body in the ICU, waiting for her brain to come back, if ever? Meet every new complication, every failing organ, or at some point accept and surrender? Many doctors were involved in her care, with differing opinions. Every algorithm was certain. After many months she recovered enough to leave the ICU for rehabilitation, and in time she was caring for her child on her own. She was, in the end, the miracle.

The machine is never better at knowing us. It may be better at applying what has already been said, and only because what is being said is sometimes unbearable to hear: for a child, for a wife, for a husband who needs to say goodbye and may need more time to do it.

So we arrive back at the same place. If talking beforehand does not make our predictions better, then prediction was never what the conversation was for. The weight returns to us, to physicians, and to a conversation almost nobody has time for: the one where a patient says, in their own words, what matters to them, before anyone needs to know, so that the weight is lifted from the people for whom that conversation will one day be impossible. Or the weight of enduring uncertainty for as long as it takes, because sometimes the right thing is not to decide but to wait: to be present, witnessing, feeling, hoping, believing.

What would it take, then, to make that conversation a standard part of care, rather than a form? It would take teaching it: at the bedside, under supervision, done badly and then better, the way every clinical skill is learned. We are redesigning medical education right now around what machines can do. This is the part they cannot do, and the more of medicine they absorb, the more deliberately we will have to form physicians who can sit down, ask what matters, and stay for the answer. Perhaps that is where the redesign should begin.

(Some details in the patient story have been altered to protect privacy.)

This essay first appeared in Nonalgorithmic on Substack. Subscribe there to receive new essays each Friday.