Netflix Figured You Out in 2009. Your Doctor Is Still Guessing.
Somewhere around 2009, a computer figured out that you'd probably like the show you binged last weekend. It did this with a few hundred ratings, a mountain of strangers' ratings, and math that fits on a napkin. Meanwhile, your doctor — a trained professional with a decade of education — prescribed you a drug tested on people who look nothing like you, at a dose picked for the average participant, based on an outcome measure you've never heard of.
One of these systems is dramatically better at handling uncertainty than the other. It's not the one with the white coat.
The Problem Netflix Had to Solve (and Medicine Mostly Avoids)
Netflix's core challenge was brutal: almost nobody rates anything. The average user has watched hundreds of hours of content and rated a handful of titles. This is called sparse data, and it's the same problem your doctor faces when trying to figure out whether a statin will actually help you, specifically.
The difference is what happens next. Netflix attacked sparsity with collaborative filtering — the insight that you don't need to know much about a movie to predict whether you'll like it. You just need to know that 10,000 people with eerily similar taste loved it. Your data gets borrowed strength from everyone else's.
Medicine has the same raw material. Millions of patients, decades of records, endless variation in who responds to what. But instead of borrowing strength across all of it, the standard clinical trial takes a few thousand people, averages them, and hands your doctor a single number: "reduced risk by 21%."
Twenty-one percent for whom? The trial won't say. Netflix would never ship a product that way.
Cold Starts: How Both Fields Handle a Stranger
Every recommender system has a cold-start problem: a brand-new user with no history. Netflix handles it with a quick onboarding — pick a few shows you like, and the algorithm has enough to start making educated guesses, refining itself with every click.
Medicine's version of a cold start is a new patient. And medicine's onboarding is... a clipboard and a family history question. After that, you're treated largely the same as everyone else your age and sex, because that's how the evidence was generated. There's no equivalent of "rate these five shows" — no systematic loop where your early responses to treatment sharpen the model for you personally.
The wild part? The infrastructure for that loop exists. It's called an adaptive trial, and it lets researchers update the study design as data comes in. They're used constantly in tech, occasionally in oncology, and almost never anywhere else. Running a trial that learns as it goes is still considered exotic in most of medicine. Running one that learns as it goes is the entire business model of every app on your phone.
What a Trial Could Borrow Tomorrow
Here's a short, realistic wishlist:
- Heterogeneity as a feature, not a bug. Recommender systems treat differences between users as signal. Clinical trials treat them as noise to be averaged away. Stratified analyses exist but are usually an afterthought.
- Continuous updating. Your Netflix profile today is not your profile from 2019. Your treatment plan, meanwhile, might be based on a trial that finished enrolling when flip phones were cool.
- Embracing the long tail. Collaborative filtering finds niche gems for unusual users. Trials routinely exclude anyone unusual — multiple conditions, odd ages, complicated meds — then wonder why the results don't generalize to the messy patients doctors actually see.
None of this requires magic. It requires accepting that the average patient doesn't exist.
Where the Analogy Snaps in Half
Before we hand the stethoscope to Silicon Valley, some honesty: the comparison only stretches so far, and it snaps at exactly the point that matters.
If Netflix recommends a dud, you lose an evening. You click off, the algorithm logs it, the system improves. The feedback loop is instant, free, and harmless. If a treatment recommendation is wrong, you might lose years. The feedback loop is slow, expensive, and sometimes irreversible. There's no "thumbs down" button for a side effect that shows up a decade later.
This is why medicine can't just run infinite experiments on individuals the way tech does. Ethical review boards exist for good reason. Exploration — the engine of every recommender system — is a core moral principle of the algorithm and a core ethical violation of the clinic. Netflix must show you weird stuff to learn what you like. A doctor cannot try random things to learn what works on you.
There's also the measurement problem. "Did you finish the season" is a clean signal. "Did the drug improve your quality of life" is a fog. Recommenders optimize a proxy that's nearly the real thing; medicine optimizes proxies that are often poor stand-ins for what patients actually care about.
The Takeaway
The lesson isn't that your doctor should be replaced by an algorithm. It's that medicine has been slower to adopt statistical machinery that's been sitting on the shelf for fifteen years — largely for cultural and regulatory reasons, not mathematical ones. The math of borrowing strength from crowds works whether the crowd is rating documentaries or responding to chemotherapy.
Your watchlist is a small-sample problem solved beautifully. Your healthcare is the same small-sample problem, solved the way it was in 1998. That gap is one of the great underreported stories in modern statistics — and closing it won't take a breakthrough. It'll take a willingness to admit the average patient was a myth all along.