In Defense of Recommendation Algorithms

The recommendation algorithm has become the agreed-upon villain of the internet, blamed for radicalization, addiction, polarization, and the general degradation of public sanity. Some of that blame lands. But the case is usually made against a fantasy alternative, a purer feed that we supposedly lost, and the fantasy does not survive contact with how these systems actually came to exist.

The honest history is that the unsorted feed was tried and people hated it. The early chronological timeline, every post from everyone you followed in reverse order, sounds clean until you follow more than a few dozen accounts. Then it becomes an unreadable torrent in which the thoughtful thing your friend wrote at noon is buried under two hundred low-effort posts by the time you look. Users did not experience the raw firehose as freedom. They experienced it as noise, and they abandoned the products that insisted on delivering it. Ranking was not imposed on a happy public. It was built because the public could not use the alternative.

Once you accept that some ordering is unavoidable, the interesting question changes. It is no longer sorted versus unsorted, because unsorted is not a real option at scale. It is which objective the sorting serves. An algorithm that ranks for time spent will learn to feed outrage and cliffhangers, because those hold attention. An algorithm that ranks for whether you found something valuable a week later would behave very differently. Same machinery, different target, opposite character. The technology is not the problem. The thing it is told to maximize is.

This distinction matters because the popular remedy, tear out the algorithm and go back to chronological, aims at the wrong layer. It attacks the existence of ranking rather than the choice of what to rank for, and it would deliver users back into the firehose they already rejected once. The reforms that would actually help are less dramatic and more specific: change the objective, expose the controls, let people tune the feed toward what they endorse rather than what they compulsively click.

None of this absolves the companies that tuned their systems for engagement and looked away from the results. That was a real choice with real damage. But the useful critique is precise about where the damage comes from. The recommendation algorithm is a tool for making an unmanageable amount of information usable, which is a genuine service, and the problem was never that it sorts. The problem is what it was asked to sort toward, and that is a decision, not a law of the technology.