How AI Is Actually Changing Healthcare (Beyond the Hype)

My grandmother waited four months for a diagnosis. Four months of tests, referrals, waiting rooms, more tests, a specialist who was booked six weeks out, a second opinion because the first specialist wasn't sure. By the time she had an answer, the condition had progressed from easily treatable to manageable-but-complicated.

This was in 2019. Not the 1970s. In a major Indian city with access to good hospitals. The bottleneck wasn't technology — it was human bandwidth. There simply aren't enough specialists to see every patient promptly, and diagnostic expertise is concentrated in urban centers while patients are everywhere.

AI won't fix institutional healthcare problems. But it's already reducing the specific bottleneck my grandmother hit — the gap between "something might be wrong" and "here's what it is." And that gap, measured in weeks or months, is where conditions worsen, anxiety compounds, and treatment windows narrow.

Diagnostic AI That's Actually Working

The headline-grabbing claims about AI diagnosis tend to be overstated. "AI is as accurate as doctors!" typically means "AI performed comparably to radiologists on a curated dataset under controlled conditions." Real-world performance is messier and less impressive.

But some applications have crossed from research into genuine clinical use, and the results are meaningful:

Diabetic retinopathy screening. Google's system for detecting diabetic eye disease from retinal images is deployed in clinics across Thailand and India. It screens patients who would otherwise wait months for an ophthalmologist, catching conditions early enough for intervention. This isn't replacing doctors — most of these patients had no access to a specialist in the first place. The AI is providing screening capacity that didn't exist before.

Chest X-ray triage. Several AI systems can now flag potentially critical findings in chest X-rays — pneumothorax, large masses, obvious fractures — within seconds. The radiologist still reads the image and makes the diagnosis. But the AI sorts the queue: critical findings move to the top, obviously normal images move to the bottom. In emergency departments where radiologists review hundreds of images daily, this triage function reduces the time to diagnosis for urgent cases from hours to minutes.

Pathology. AI-assisted pathology tools help pathologists identify cancer cells in tissue samples. Not replacing their judgment — augmenting their attention. A pathologist examining a tissue slide might need to scan thousands of cells. AI can highlight regions of interest, acting as a tireless second pair of eyes. Studies show that pathologist plus AI consistently outperforms either alone.

Drug Discovery: Where AI Might Change Everything

Traditional drug discovery follows a brutal pipeline: identify a biological target, screen thousands of compounds to find one that interacts with it, test that compound endlessly, fail most of the time, start over. The average new drug takes 10-15 years and costs over $1 billion to develop. Most candidates fail. The success rate from initial compound to approved drug is around 5-10%.

AI is being applied at every stage of this pipeline, and in my view, this is where AI's healthcare impact will ultimately be largest — not in diagnosis, but in creating treatments that wouldn't otherwise exist.

AlphaFold — DeepMind's protein structure prediction tool — solved a problem that had stumped biology for fifty years: predicting how a protein folds into its three-dimensional shape from its amino acid sequence. This matters because a protein's shape determines its function, and understanding function is essential for designing drugs that interact with it. AlphaFold has now predicted structures for nearly every known protein. The database is free. Researchers worldwide are using it to accelerate their work.

Companies like Insilico Medicine, Recursion Pharmaceuticals, and Isomorphic Labs (a DeepMind subsidiary) are using AI to generate novel drug candidates computationally — testing millions of virtual compounds against targets before ever touching a test tube. Insilico claims to have brought a drug from target identification to Phase 1 clinical trials in under 30 months, roughly one-third the typical timeline.

What AI Can't Do in Healthcare

It can't provide empathy. It can't read the room when a patient's body language contradicts their verbal responses. It can't make the judgment call that a technically correct treatment plan is wrong for this particular patient because of their life circumstances, values, or unstated preferences.

It can't handle the novel. AI systems are pattern matchers — extraordinarily powerful pattern matchers, but pattern matchers nonetheless. When a patient presents with something genuinely unusual, the AI has no pattern to match. Human doctors can reason from first principles, combining knowledge across domains to approach unfamiliar territory. AI, currently, cannot.

I think the right framing is: AI as a highly competent but narrowly focused colleague. It will handle the routine exceptionally well, freeing human doctors to spend more time on the complex, the unusual, and the human elements of care that no algorithm can replicate.

My grandmother's four-month wait might have been two weeks with AI-assisted triage and screening. Not because AI would have replaced her doctors, but because it would have handled the routine steps faster, moving her to the right specialist sooner. That's not a revolution. It's a meaningful improvement, and in healthcare, meaningful improvements save lives.