The Algorithm Has No Taste

A recommendation algorithm does not know what is good. It knows what people clicked on before and predicts what you will click on next. These are not the same thing, and the confusion between them is producing cultural environments of a particular kind of badness.

Let me be precise about what I mean. When Netflix recommends a show, it is not asking: is this well-made? Does it treat its characters with intelligence? Will you be glad you watched it in six months? It is asking: given the behavior of users similar to you, what has the highest probability of being played for more than seventy seconds, thus registering as a view?

This is a reasonable thing for a business to optimize for. It is not a reasonable thing to mistake for curation.

The problem is that the scale of algorithmic recommendation has displaced older systems of cultural discovery — critics, bookshops, word of mouth, the serendipity of a record store browser — without reproducing their functions. The older systems were not perfect. They were gatekept, biased, and often wrong. But they were epistemically different in a way that mattered: they were not optimization processes. They were opinions.

An opinion can be argued with. It has a position, which means it can be right or wrong, defensible or indefensible. The algorithm has no position. It has a loss function.

I think about this when I look at what recommendation systems have done to music. Spotify's Discover Weekly is genuinely useful — it surfaces music I would not have found on my own, and some of it is excellent. But the long-term effect on the music ecosystem is to advantage music optimized for streaming: shorter songs, hook-in-the-first-ten-seconds structures, less dynamic range, more sonic consistency. These are not aesthetic choices made by artists. They are adaptations to an engagement function.

Books are heading the same way. Amazon's recommendation system advantages certain genres and certain cover aesthetics and certain opening-chapter pacing conventions. Not because these are better books, but because they are books that perform well against the metric of: did the person who downloaded the sample go on to purchase the full book?

The most insidious version of this is what happens to news and nonfiction. Recommendation systems in news optimize for what people engage with emotionally, which is not the same as what is true, important, or worth your time. We have been running this experiment for fifteen years and the results are in: people are more anxious, more angry, more certain of things that are wrong, and less trusting of institutions than they were before.

I am not saying technology caused this. I am saying technology accelerated it, and built business models on top of it.

The alternative is not a return to a gatekept past. It is to develop something we do not yet have very well: systems and practices for algorithmic curation that optimize for quality rather than engagement, and that are transparent enough to be argued with.

Some things exist. Letterboxd is a recommendation system built around explicit ratings from real humans who have opinions. Goodreads, despite its problems, has the seed of something similar. Substack's discovery layer is built around editorial judgment made by editors you choose to trust.

These are small and imperfect. But they are epistemically different from engagement optimization. They have taste, or something close enough to argue with.

That is what we need more of.