There used to be a very simple technology for discovering something new. It worked like this: you stood in front of a shelf full of records, movies, or books without really knowing what you were looking for, and you let your eyes stumble across something. The stumble was the whole mechanism. You picked up a record because of its cover, without knowing the artist, without any reason an algorithm could reconstruct afterward, and sometimes that record changed something in you. Most of the time it didn’t. But the system never needed to succeed very often to justify itself. It only needed to leave the door open.
That door closed a long time ago, and the strange thing is that nobody held a funeral. We closed it ourselves, gladly, in exchange for something that seemed strictly better: no longer wasting time on things we probably wouldn’t like.
The Deal We Were Offered
At its core, the logic behind recommendation algorithms is both honest and perfectly reasonable. Take everything we already know about you—what you’ve watched, how long you stayed, where you stopped, what you’ve consistently ignored—and use it to predict what you’re most likely to enjoy next. The better the prediction, the less friction, the less wasted time, and the more immediate satisfaction.
Netflix says that around eighty percent of the hours people spend watching content now come from personalized recommendations rather than direct searches. That’s not a trivial statistic. It means that, for most of us, we no longer decide what to watch. We simply confirm what the system had already decided we were probably going to want.
It sounded like an unbeatable deal. Who would miss the minute spent browsing a shelf if a system could save it?
The problem is that the minute we thought we were saving wasn’t an inefficiency in the process.
It was the process.
What the Algorithm Can’t Do—By Design
Here’s the point that almost nobody discusses with the seriousness it deserves: recommendation systems need data about you in order to work. And the only data they can ever have is data about what you’ve already done.
Structurally, every recommendation is therefore an extrapolation from your past.
It can never offer you something that truly lies outside who you already are, because it has no other raw material to work with. At best, it gives you an intelligent variation of yourself. At worst, it gives you an increasingly narrow repetition of the same thing, wrapped in different packaging.
More than a decade ago, the term filter bubble was coined to describe exactly this phenomenon: the more accurately a system predicts what you want to see, the more it quietly narrows the range of things you actually end up seeing.
This isn’t a flaw in the system.
It’s the system working exactly as it was designed, optimizing the metric it was asked to optimize—which is almost never “broaden this person’s horizons” and almost always “keep them looking at the screen a little longer.”
And that’s where the irony I find most interesting appears.
The platforms know this.
They know it so well that they’ve built simulations of serendipity to compensate for it. Playlists called Discover Weekly. Rows labeled Something Different for You. Carefully curated mixes promising surprise.
But let’s call things by their name.
That isn’t serendipity.
It’s prediction disguised as accident.
The system is still calculating, using the same logic as always, the optimal variation of the familiar that you’ll accept without ending your session. We call it discovery because we need to believe we’re still discovering something.
But once an accident has been calculated, it stops being an accident.
Why Nobody Held a Funeral
Nobody mourned this because there was never a precise moment when serendipity died. It wasn’t an event. It was a slow replacement, one shelf after another, one record store after another, until one day the entire mechanism was gone and its replacement felt so natural that we barely noticed anything was missing.
It’s much easier to grieve something that disappears overnight than something that dissolves so gradually you never consciously realize you’ve lost it. We mourn what we remember losing. This disappeared slowly, across generations of teenagers who never knew the shelf in the first place.
And the consequence isn’t merely cultural, in the boring sense that “we don’t discover new music anymore.” It’s deeper than that. Genuine chance—the encounter with something no system could have predicted because not even you could have predicted it—is one of the few reliable ways we have of truly changing direction. Not the comfortable variation within what’s already familiar, but the turn that wasn’t contained in any previous history because it didn’t come from any previous history.
A system trained exclusively on your past can never offer you that, no matter how sophisticated it becomes.
The question worth asking isn’t whether recommendation algorithms are good or bad.
They’re extraordinarily good at what they do.
The more uncomfortable question is this:
If we let them decide almost everything we consume, where will the next thing that changes us come from, if by definition it has to be something no system trained on who we already are could ever have recommended?