Think about where the last song you played came from.
If you searched for it, you chose it. If an app put it on the first screen, the story is a little different. You liked it and listened to the end, and that record became raw material for the next recommendation.
The feeling of liking it is real, and the song may well be good. Still, here is the question I want to ask: if that song had never shown up on your screen, would you have come to like it?
We have assumed that taste is something we pick with
The word "taste" carries an old assumption. Somewhere inside us there is already an outline of what we like, and we go through the world finding things that fit it. Running a finger along the spines in a bookshop until it stops on one is the picture most people have in mind.
In that picture, taste comes first and choice comes second.
Recommender systems unsettle the order. Once the options number in the millions, nobody can scan them from the beginning. Something has to narrow the field first. You choose inside the narrowed list, the choice is recorded, and the record shapes the next list. It becomes hard to say whether taste produces choices or choices produce taste.
I am not saying this loop is bad. A bookseller's recommendation and an album lent by a friend run through the same loop. What changed is the scale, the speed, and the fact that the inside of the loop is hard to see.
A window-side shelf of books and records with a few pulled slightly forward, an imagined sceneView original
Recommendation does real work
First, a look at how widely the technology is used.
In a 2015 paper, Netflix's Carlos Gomez-Uribe and Neil Hunt describe the recommender system as a key pillar of the product, helping members find something to watch in every session. At the time of writing, they say, Netflix had more than 65 million members streaming more than 100 million hours a day. This essay stops at scale and standing. I did not find a figure in that paper for what share of viewing recommendations drive, so I will not offer one.
Spotify's researchers start from a similar premise. The 2020 Web Conference paper by Anderson and colleagues notes that the short-term benefits of recommender systems are well known while their long-term effects are less understood. So the paper looks at what recommendations do to the breadth of music people listen to.
Recommendation is not decoration. It takes part in what people watch and hear, which makes the question about taste a fair one.
People who see others' choices choose differently
The first thing a claim like "recommendation changes taste" needs is an experiment. One that gets cited often is the 2006 study in Science by Matthew Salganik, Peter Sheridan Dodds, and Duncan Watts.
The researchers built an artificial music market. A total of 14,341 participants downloaded songs they had never heard. Some could see what earlier participants had downloaded, and some could not. When the strength of social influence went up, both inequality and unpredictability of success went up with it. Success was only partly determined by quality, too. The best songs rarely did badly and the worst rarely did well, but anything in between could happen.
This was not a test of recommendation algorithms. It shows that the same song can have a different fate once other people's choices are visible. A modern recommendation list gathers many people's choices and shows them to you, so the study is a reasonable place to begin understanding the effect. Going from there straight to "algorithms steer taste" would be too far, though. What the paper supports is that choices respond to social signals.
Easy recommendations tend to narrow
What happens to the taste of someone who receives a lot of recommendations? The Spotify research offers one clue.
The team built a way to measure distance between songs from listening records and scored how diversely each user listens. They found two things. High diversity was strongly associated with long-term metrics such as conversion to paid plans and retention. And listening driven by recommendations, meaning streams programmed by an algorithm, was associated with lower diversity. According to Spotify's research blog, most premium users listened more diversely in organic streams they had chosen themselves and less diversely in programmed ones. When users became more diverse over time, they did so by moving away from algorithmic listening and toward listening they had chosen.
There are cautions. The authors say the findings are correlational and fit several causal stories. People who love music may simply listen widely and stay longer, or they may find variety by routes other than recommendations. The diversity measure also relies on how likely songs are to share a playlist, a principle some of Spotify's own recommenders use, and that is worth keeping in mind when reading the results.
One sentence from this research is the one I take away. Recommendation is good at hitting what you need right now, and that same skill can sit close to narrowness. The paper names the same tension: recommending what suits an immediate need pulls in a different direction from helping users stay diverse over time. A randomized experiment in the paper found that recommendations worked better for users with lower diversity, which adds to the tension.
Where does taste remain?
Read together, the three studies sketch this picture. Recommendation operates at large scale and does intervene in choices. Choices change when other people's choices are visible. Programmed listening tends to be narrower than listening a person selects. None of this lets us say taste is a product of algorithms. The evidence does not go that far.
Even so, the assumption inside the word "taste" needs a small repair. Taste does not sit finished inside us waiting to find a match in the world. It grows among the things we run into, and what we run into is its raw material. If the hand that picks the material is not ours, it is fair to ask how much of the finished flavor is.
That leaves a distinction. The feeling of liking is an outcome, and how we met the thing is a path. We look at the outcome and call it taste. When the path is hidden, giving reasons becomes harder as well. Answering "why do you like this?" with "it kept coming up" is honest, but it is a thin account of taste.
A window-side table with an empty notebook, a pencil and headphones, an imagined sceneView original
Three ways to keep a taste your own
I do not want to tell anyone to stop using recommendations. Songs and essays that I would never have found otherwise exist because of them. Still, I think there are ways to keep at least part of the path in our own hands.
The first is to write down where you found things. When a song or video lands, add one line on where it came from: the recommendation screen, a person, or your own search. After a while you can see which doorway most of your taste comes through.
The second is to look for a share of your material outside recommendations. In the Spotify study, users who became more diverse increased the listening they chose themselves. That should not be stretched into a prescription, but it works as a hint about direction. Once a week, choose from a place where people did the gathering: a library shelf, a record shop, someone's hand-made playlist.
The third is to put your dislikes into words. A recommender will gladly multiply what you like, but it cannot write out why you dislike something. Once the reason is a sentence, it is at least your own judgment.
The same question turns up at work
The question does not end with music apps. Teams that bring AI into their work often notice something. After using a tool that drafts documents for a while, the tone and structure of their documents start to converge, and the pile of documents drifts toward an average that belongs to no one. Whether to treat that as a matter of taste or a matter of quality differs by team, but the fact is the same: the choosing happens inside candidates someone else put forward.
This essay does not cite research that measures that effect. The studies above concern listening and choice in recommendation settings. I have no evidence here that work documents show the same effect. Still, I think the habit of asking who is supplying the candidates is useful on either side.
The question that remains
Back to the opening question. If the song had never appeared on your screen, would you have come to like it?
Probably not. But the same is true of a book picked up in a corner of a shop or a record a friend put on. No taste forms without an encounter. What changes is how visible the party designing the encounters is.
So I do not think protecting taste means cutting off recommendations. It is closer to remembering where you met things. If you attach a path to the feeling of liking, the taste is at least one whose origin you know.
References
- Carlos A. Gomez-Uribe, Neil Hunt, "The Netflix Recommender System: Algorithms, Business Value, and Innovation," ACM Transactions on Management Information Systems 6(4), 2015. https://doi.org/10.1145/2843948
- Matthew J. Salganik, Peter Sheridan Dodds, Duncan J. Watts, "Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market," Science 311(5762), 2006. https://doi.org/10.1126/science.1121066
- Ashton Anderson, Lucas Maystre, Ian Anderson, Rishabh Mehrotra, Mounia Lalmas, "Algorithmic Effects on the Diversity of Consumption on Spotify," The Web Conference 2020. https://doi.org/10.1145/3366423.3380281
- Spotify Research, "Algorithmic Effects on the Diversity of Consumption on Spotify," 2020. https://research.atspotify.com/2020/12/algorithmic-effects-on-the-diversity-of-consumption-on-spotify
