I’ve sat in the room where these systems get tuned. Nobody writes “make people angrier” into the spec. What they write is “increase session length” and “increase return rate,” and then they let a model figure out the fastest path to those numbers using your own behavior as the map. It turns out the fastest path, reliably, across billions of people, runs through anger, fear, and outrage — not because engineers are evil, but because those emotions produce a measurably longer scroll than contentment does.
That’s the whole mechanism. There’s no ideology in the code. There’s a scoreboard, and outrage wins the scoreboard.
The part people skip is that the algorithm needed something to learn from, and that something was you — every video you lingered on a half-second longer, every post you tapped to expand, every rage-quit you did anyway after already watching the whole thing. You trained the system that trained you back. It’s not a hypnotist implanting beliefs. It’s a mirror that got very good at figuring out which of your existing impulses to feed first.
Which means the fix was never going to be “better content moderation.” Moderation removes the worst outputs. It doesn’t touch the incentive that’s still running underneath, optimizing for the same thing it was optimizing for yesterday.
You can’t out-argue a recommendation engine, but you can starve its training data. The thing that changes what you get recommended isn’t complaining about the algorithm — it’s what you actually do with your thumb in the next ten seconds. Every “keep watching” that isn’t a deliberate choice is a vote for more of exactly that. The app didn’t radicalize you against your will. It got very precise, very fast, about what you’d already decided — without deciding — to keep giving it.
