AI is already smarter than your doctor. That’s not the scary part. – Unite.AI

I wanted AI to solve this problem.

For years I have taken complicated patient histories, lab results, symptoms, medications, supplements, and years of previous treatments and tried to turn it all into a coherent clinical plan. A complex case might take three hours of research before I felt comfortable deciding what mattered and what should happen next.

Then came generative AI.

My first thought was the obvious one: finally.

This was technology that could consume enormous amounts of information, call up medical knowledge faster than I ever could, and find relationships in a case in seconds. I figured I could build an interface around a large language model, build some safeguards, and be done.

It would have been fabulous.

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

The reason wasn’t that the AI ​​wasn’t smart enough.

The thing was, intelligence was the easy part.

I was getting an AI that was different from everyone else

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

Why? Because I wasn’t just asking a question and accepting the answer. I was debating it. Challenging hypotheses. Provide the context he was missing. Error correction. Change the frame. Pushing again.

Technically we used the same instrument. Functionally, we weren’t.

This is an interesting productivity problem when 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 clinically brilliant and still be terrible at getting useful information from an AI system.

These are different skills.

Yet much of the healthcare industry’s response seems to be that doctors just need to learn “ready engineering.”

I don’t think it’s a serious solution.

If a clinical AI system requires a doctor to become a sophisticated timely engineer before the system becomes reliable, we are not finished building clinical AI.

The research is already telling us something uncomfortable

A 2024 randomized clinical trial provided 50 doctors with difficult diagnostic cases. Doctors using GPT-4 scored 76%; those using conventional resources scored 74%, a difference that was not statistically significant.

So the researchers tested only GPT-4. It beat doctors using conventional resources by 16 percentage points.

A 2026 study conducted in Pakistan found that, after AI literacy training, doctors with GPT-4o access scored 71.4% in diagnostic reasoning compared to 42.6% for those using conventional resources. But GPT-4o alone scored 82.9%.

A 2026 UK replication found a similar gap: doctors improved with AI assistance, but still achieved more than 21 percentage points below the LLM working alone.

We have to say the uncomfortable part out loud: On some cognitive tasks, AI is already better than a single doctor.

No doctor has read every medical article or can remember every obscure ailment, interaction, or contraindication the instant it becomes relevant. AI can process massive amounts of information without getting tired or running out of time.

The mistake is assuming that we can put a chatbot in front of a doctor and say the problem is solved.

The problem is the blank request box

Imagine a doctor uploads a lab report containing 150 markers and asks an LLM to analyze it.

The response comes back well organised. There’s just one problem: the system didn’t reliably extract all 150 markers.

General purpose LLMs are not deterministic clinical document parsers. The tables and busy medical reports remain particularly difficult. The scary part is that the final response may not announce, “I lost 50 values.”

He could just answer.

Now imagine that one of these missing values ​​requires urgent medical evaluation.

It wasn’t a limitation I was willing to work around. We had to create different extraction technology to handle lab reports accurately and reproducibly before the clinical reasoning part of our system could begin.

That experience changed the way I think about AI safety.

An AI system telling you it doesn’t know something is inconvenient.

An AI system that doesn’t know what it missed is dangerous.

Because we had to build beyond AI

Clinical work requires something that general-purpose generative AI doesn’t naturally provide: reproducibility.

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

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

Engineers need rules that are explicit enough to allow the software to run. Doctors go through life saying, “Yes, except…”

Human biology is full of exceptions, modifiers, context, and conflicting explanations. At times, building the system seemed like one group was communicating in sign language while the other spoke Klingon.

The answer has never been “AI” or “not AI”. It was a matter of understanding exactly where each belonged.

Artificial intelligence can read the graph. Humans can read the room.

When used correctly, AI can expand a doctor’s memory, speed up research, surface unusual possibilities, and recognize patterns across massive amounts of information.

This starts to look less like replacing the practitioner and more like building a bionic one.

But the patient is not a clinical vignette.

A practitioner feels hesitation before responding. He notices when the story changes. He learns that one patient habitually minimizes symptoms while another reports every sensation to maximum intensity. He knows that “I’m fine” clearly doesn’t mean fine.

Such information may never exist in the chart.

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

Artificial intelligence can read the graph. Humans can read the room.

The best clinical systems will need both.

Stop wondering if AI will replace the doctor

We are asking the wrong question.

The future won’t necessarily belong to the company with the smartest chatbot or the smartest foundation model. Important innovation increasingly happens around the model: how information is acquired and verified, what knowledge is provided, what the system can infer, where deterministic controls take over, and where a human must remain in the loop.

Raw intelligence is becoming abundant.

Reliable systems for applying such intelligence are much more difficult to build.

And we shouldn’t solve this problem by demanding that every doctor become an amateur AI developer. The complexity should lie in the technology, not in the doctor’s suggestions.

The doctor of the future will not beat artificial intelligence on memory. That contest is already becoming pointless.

We should instead build systems that give doctors machine memory, processing power, and pattern recognition, while preserving the practitioner’s context, judgment, skepticism, and human connection.

We call him the bionic practitioner.

I suspect patients will eventually name it much more simply.

Their doctor.