Medicine is not a benchmark, and patients are not vignettes

Today, a new article appeared in JAMA from billionaire Vinod Khosla: "Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care?"

The authors argue we're getting closer to the day when AI is better than doctors. AI can already generate differential diagnoses, recommend diagnostic tests, retrieve enormous amounts of medical evidence and identify guideline-concordant treatments with impressive accuracy.

These capabilities are real, but they raise a more fundamental question: is performing medical tasks the same as practicing medicine?

Then, the authors go one step further. As AI becomes better than physicians, they argue, physicians may actually make it worse. Eventually, autonomous AI could outperform not only doctors, but doctors using AI.

It's an interesting argument. But it also confuses something medicine has struggled with for a long time: Knowing the evidence is not the same as knowing what to do with a patient.

AI is getting very good at medical problems

There is little reason to doubt that AI will outperform physicians at many bounded clinical tasks. Give an AI a complete case and ask for the differential. Ask it which tests should be ordered. Ask it what NCCN or ASCO recommends.

Ask it to summarize every relevant trial published during the past five years.

AI may soon do all of these things better than any individual physician. That's extraordinary, but it's not the hardest part of oncology.

Patients don't arrive as clinical vignettes

A clinical vignette gives you the information that matters. Patients don't.

They arrive with incomplete histories, conflicting records, subtle symptoms, comorbidities, family dynamics, previous toxicities, financial constraints, transportation problems and preferences that may not become apparent until halfway through a conversation.

Sometimes, the most important piece of information isn't in the chart at all.

An experienced oncologist notices that the patient who said he tolerated his last cycle "fine" is now walking more slowly into the room. Or that the daughter is answering questions the patient previously answered himself. Or that a treatment that is technically appropriate is becoming increasingly unrealistic for the patient sitting in front of them.

This isn't trivia around the edges of medicine. They are medicine.

Guideline-concordant isn't always the right answer

AI will become exceptionally good at knowing guidelines. That is valuable.

But oncology doesn't become difficult when there is one clearly superior, guideline-supported treatment. It becomes difficult when there are three reasonable treatments.

When the trial population doesn't look like your patient.

When renal function is marginal.

When the patient is 84 instead of 64.

When the disease is responding, but the toxicity is accumulating.

When the evidence changed six weeks ago.

When the guideline says either option is appropriate, but an experienced oncologist strongly prefers one of them in this particular situation.

Or when the technically "correct" treatment isn't the treatment the patient wants.

Guidelines provide the map, but clinical judgment helps you navigate the terrain.

The hardest oncology decisions live in the grey zones

Consider two patients with exactly the same cancer. Same stage, same biomarkers, same previous treatment, and same scans.

An algorithm might reasonably recommend the same next therapy for both. But one patient wants every possible month of disease control, regardless of toxicity. The other wants to attend her daughter's wedding without spending the preceding six weeks exhausted from treatment.

Those preferences don't modify the medical decision. They are part of the medical decision.

Practicing oncology isn't simply about choosing the therapy with the highest expected response rate. It's about understanding which outcome matters most to the patient and then applying the evidence accordingly.

AI can know every trial, but it still has to interpret the patient.

There's another distinction that matters. Medical knowledge exists in papers, but clinical expertise does not. Experienced oncologists accumulate thousands of small lessons that never appear in the literature:

• Which toxicities become difficult in actual practice

• Which patients tolerate a regimen better than expected

• When to dose reduce

• When to wait

• When a technically reasonable treatment isn't worth pursuing

• When something about a patient's presentation simply doesn't fit

This is why fellows still ask attendings what they would do. Why community oncologists call disease specialists. Why difficult cases go to tumor boards. And why two excellent oncologists can read exactly the same literature and make different recommendations.

Clinical judgment is not imperfect retrieval of medical information; it is the interpretation of that information through experience.

Maybe the problem isn't keeping physicians in the loop

The JAMA authors raise another fascinating possibility. If AI eventually becomes more accurate than physicians, they argue, physicians may degrade its performance by incorrectly overriding its recommendations.

This is possible. But there's also another interpretation: perhaps we simply haven't learned how physicians and AI should work together.

Most doctors practicing today were never trained to collaborate with an AI system. We haven't developed norms around when to trust it, when to challenge it, when to request additional evidence, or when human judgment should take precedence. These workflows are just beginning to emerge. And the AI systems keep changing.

The first generation of physician-AI interaction may tell us less about the limits of collaboration than about how primitive our collaboration models are.

Medicine is not chess

The article also draws an analogy to chess. Once computers became sufficiently good at chess, humans eventually stopped improving the machine. The machine was simply better.

But medicine is different. In chess, the board is visible, and the rules are fixed.

The available moves are known. And everyone agrees on the objective: to win.

Medicine doesn't work that way. The information is incomplete, the probabilities are uncertain, the available options change.

And most importantly, not all patients define "winning" the same way. One patient prioritizes survival. Another prioritizes independence. Another prioritizes avoiding hospitalization. Another simply wants enough time to see the birth of a grandchild.

In an AI system, there is no single, objective function that captures these considerations.

The future isn't physician vs AI

AI is going to transform medicine. There are enormous parts of clinical practice that machines will eventually do better than humans. AI should remember every drug interaction, monitor every lab result, and know every study. It should recognize patterns that clinicians miss. It should make sure important details aren't missed by the care team.

And clinicians should use it. But the goal shouldn't be to determine whether the machine wins or the doctor wins.

The goal should be to build a system where each party does what it does best.

AI can provide extraordinary intelligence. Physicians provide context, experience, accountability and judgment. And patients provide the goals that ultimately determine what good medicine means.

The real opportunity

The most interesting question isn't whether AI can replace the oncologist.

It's what happens when every oncologist has access to AI and the accumulated reasoning of other experienced physicians. Imagine being able to instantly retrieve the evidence, and then see how leading specialists actually apply it.

Not just: what does the guideline recommend?

But: what would you do for this patient, and why?

That's where clinical decision support is heading. Because the future of medicine isn't humans competing against machines, it's better tools helping doctors to make better decisions for the patient sitting in front of them.