AI Is Already Smarter Than Your Doctor. That’s Not the Scary Part.

I wanted AI to solve this problem.

For years, I had been taking complicated patient histories, laboratory findings, symptoms, medications, supplements and years of prior treatment and trying to turn all of it into a coherent clinical plan. A complex case could consume three hours of research before I was comfortable deciding what mattered and what should happen next.

Then generative AI arrived.

My first thought was the obvious one: finally.

Here was a technology that could consume enormous amounts of information, recall medical knowledge faster than I ever could and find relationships across a case in seconds. I assumed I could put an interface around a large language model, build a few safeguards and be finished.

That would have been fabulous.

Instead, roughly a year and $1 million later, with software engineers working alongside a 10-person clinical team, we were still building.

The reason was not that AI wasn’t intelligent enough.

It was that intelligence was the easy part.

I Was Getting a Different AI Than Everyone Else

My first clue came before we ever tried to build clinical software. I kept noticing that I could use the same AI tool as someone sitting next to me and get dramatically better results.

Why? Because I wasn’t simply asking a question and taking the answer. I was arguing with it. Challenging assumptions. Supplying context it had missed. Correcting errors. Changing the frame. Pushing again.

Technically, we were using the same tool. Functionally, we weren’t.

That is an interesting productivity problem when you are writing an advertisement. It becomes a very different problem when the subject is a patient.

I have worked with more than 70,000 practitioners. Someone can be brilliant clinically and still be terrible at getting useful information out of an AI system.

Those are different skills.

Yet much of healthcare’s answer seems to be that clinicians just need to learn “prompt engineering.”

I don’t think that is a serious solution.

If a clinical AI system requires a physician to become a sophisticated prompt engineer before the system is dependable, we have not finished building the clinical AI.

The Research Is Already Telling Us Something Uncomfortable

A 2024 randomized clinical trial gave 50 physicians difficult diagnostic cases. Physicians using GPT-4 scored 76 percent; those using conventional resources scored 74 percent, a difference that was not statistically significant.

Then researchers tested GPT-4 alone. It beat the physicians using conventional resources by 16 percentage points.

A 2026 trial in Pakistan found that, after AI-literacy training, physicians with GPT-4o access scored 71.4 percent on diagnostic reasoning versus 42.6 percent for those using conventional resources. But GPT-4o alone scored 82.9 percent.

A 2026 UK replication found a similar gap: physicians improved with AI assistance but still scored more than 21 percentage points below the LLM working by itself.

We need to say the uncomfortable part out loud: on certain cognitive tasks, AI is already better than an individual physician.

No physician has read every medical paper or can remember every obscure disorder, interaction or contraindication the instant it becomes relevant. AI can process enormous amounts of information without getting tired or running out of time.

The mistake is assuming we can put a chatbot in front of a clinician and call the problem solved.

The Empty Prompt Box Is the Problem

Imagine a physician uploads a laboratory report containing 150 markers and asks an LLM to analyze it.

The answer comes back beautifully organized. There is only one problem: the system didn’t reliably extract all 150 markers.

General-purpose LLMs are not deterministic clinical document parsers. Tables and dense medical reports remain particularly difficult. The frightening part is that the final response may not announce, “I missed 50 values.”

It may simply answer.

Now imagine one of those missing values requires urgent medical evaluation.

That was not a limitation I was willing to build around. We had to create different extraction technology to handle laboratory reports accurately and reproducibly before the clinical reasoning portion of our system could begin.

That experience changed the way I think about AI safety.

An AI system that tells you it does not know something is inconvenient.

An AI system that does not know what it missed is dangerous.

Why We Had to Build Beyond AI

Clinical work requires something general-purpose generative AI does not naturally provide: reproducibility.

We needed control over what went in and what came out: reliable data extraction, structured knowledge, deterministic logic, safeguards and systems capable of evaluating relationships across a patient’s history.

Clinical reasoning is not simply a collection of if-then statements. Sometimes the signal is the relationship between several markers. Sometimes five individually unremarkable findings become important when viewed together.

Engineers need rules explicit enough for software to execute. Clinicians spend their lives saying, “Yes, except…”

Human biology is full of exceptions, modifiers, context and competing explanations. At times, building the system felt as though one group communicated in sign language while the other spoke Klingon.

The answer was never “AI” or “not AI.” It was figuring out exactly where each belonged.

AI Can Read the Chart. Humans Can Read the Room.

Used correctly, AI can expand a clinician’s memory, accelerate research, surface uncommon possibilities and recognize patterns across huge amounts of information.

That begins to look less like replacing the practitioner and more like building a bionic one.

But the patient is not a clinical vignette.

A practitioner hears the hesitation before the answer. She notices when the story changes. She learns that one patient habitually minimizes symptoms while another reports every sensation at maximum intensity. She knows when “I’m doing fine” very clearly does not mean fine.

That information may never exist in the chart.

Human interaction in medicine is not merely a pleasant layer we add after the machine has done the intellectual work. It is another source of clinical data.

AI can read the chart. Humans can read the room.

The best clinical systems will need both.

Stop Asking Whether AI Will Replace the Doctor

We are asking the wrong question.

The future will not necessarily belong to the company with the cleverest chatbot or smartest foundation model. The important innovation is increasingly happening around the model: how information is captured and verified, what knowledge is supplied, what the system is allowed to infer, where deterministic controls take over and where a human must remain in the loop.

Raw intelligence is becoming abundant.

Reliable systems for applying that intelligence are much harder to build.

And we should not solve that problem by demanding that every clinician become an amateur AI developer. The complexity should live inside the technology, not inside the doctor’s prompt.

The physician of the future will not beat AI on memory. That contest is already becoming pointless.

Instead, we should build systems that give clinicians the memory, processing power and pattern recognition of the machine while preserving the context, judgment, skepticism and human connection of the practitioner.

We call that the bionic practitioner.

I suspect patients will eventually call it something much simpler.

Their doctor.