Spotify knows my taste better than most of my friends.
And somehow, despite that, I keep discovering music I love by accident. By word of mouth. By a stranger's playlist. Almost never through the algorithm.
I started wondering why.
Recommendation algorithms are built to predict what you'll like.
Not what's good for you. Not what will expand you. What you'll like — which means what you've already liked, adjusted slightly.
Netflix recommends shows similar to shows you watched. Spotify recommends songs similar to songs you played on repeat. Amazon recommends products similar to products you bought.
It's a mirror. A very accurate, very sophisticated mirror.
The problem with a mirror is that it shows you yourself.
Over and over. Slightly varied. But always you.
The algorithm that knows you best is also the algorithm that traps you most effectively in who you already are.
There's a term for this: the filter bubble. Coined by Eli Pariser back in 2011, it describes how personalization creates an information environment shaped around your existing preferences, slowly filtering out things that differ from what you already know and believe.
Spotify builds a music bubble. Twitter builds a political bubble. YouTube builds an ideological bubble.
You asked for personalization. You got isolation.
The algorithm didn't radicalize anyone. It just found out what they already believed and fed them more of it, forever.
What recommendation systems optimize for vs. what might actually help you:
Optimized for: clicks, watch time, replays, return visits
Not optimized for: growth, surprise, challenge, real satisfaction
Nobody measures: "did this change how you think?"
Some companies are trying to fix this. Spotify's "Discovery Weekly" was an attempt. But even that eventually narrows — it learns what kinds of discovery you accept, and stops offering anything too far outside it.
The real fix is you.
Go somewhere the algorithm didn't send you. Ask a person. Read something from outside your feeds. Follow someone whose recommendations confuse you.
The algorithm will never recommend what you don't already know you want.
That's its design. Not its failure.