The Internet Doesn't Just Learn Your Taste. It Helps Shape It.

Imagine Spotify recommends a song you never searched for.

You let it play. Maybe you listen to the end. Maybe you save it.

The explanation seems obvious: Spotify has discovered something about your taste.

That may be true.

But the song was not something you independently found and presented to the system as evidence. The platform placed it in front of you. Your response occurred inside an environment it had already helped arrange.

The play became a signal about preference.

But it may also have become part of the process that strengthened one.

We usually imagine recommendation systems as mirrors. They watch what we choose, learn what we like, and reflect it back to us.

The simplest model of personalization looks like this:

You have a preference.
The algorithm detects it.
The algorithm recommends more of it.

In this model, your taste exists first. The system merely becomes better at reading it.

But every recommendation changes something before the next choice is made:

The menu.

One song is placed before another. One film is ranked above another. One video appears while thousands of other possibilities remain unseen.

This does not make the resulting choice unreal. Nor does it mean a system can make people select anything it wants.

It means the choice was made from a field of possibilities that had already been filtered, ranked, and arranged.

A restaurant menu does not create your hunger. It does not invent all your tastes. But it affects what you notice, what you compare, and what you never consider ordering.

A personalized digital menu goes further.

After you choose, it remembers.

The system does not observe preference directly. It observes signals: a play, a skip, a replay, a save, a completed episode, a rating, or a few extra seconds spent watching. Spotify, for example, says listening, skipping, and saving can help train its recommendation systems and influence the taste profile used for future recommendations. 

Those actions are interpreted as evidence about what should appear next.

The structure looks less like a mirror and more like a loop:

The system learns from your behavior.

But the behavior it learns from happened inside a menu the system had already helped construct.

This distinction would matter less if recommendations merely predicted what people were already going to do.

But changing the recommendation environment can also change what people go on to consume.

In a randomized field experiment on Spotify, researchers gave users podcast recommendations. One group received recommendations personalized using its previous music-listening history. A control group received podcasts that were popular among people in the same demographic group.

The personalized recommendations increased podcast streams by an average of 28.9 percent.

They also changed the structure of that consumption. Individual users listened to a less diverse range of podcasts, even as listening across the entire group became more diverse. The researchers also found signs that the recommendations affected listening in parts of Spotify that the experiment had not directly changed. 

The experiment did not prove that Spotify permanently rewrote anyone’s taste. It studied podcast consumption, not the slow construction of a musical identity.

But it demonstrated something narrower and still important:

Change the recommendation environment, and you can change the behavior later recorded as evidence of preference.

That is one half of the loop.

The other half begins when the behavior returns to the system.

Netflix describes recommendation as a process in which the system selects what to show, then receives feedback through actions such as playing, skipping, rating, completing, or adding a title to a list. Those signals can be translated into rewards and used to train the policy responsible for later recommendations. 

At Netflix, the loop is not a metaphor.

A recommendation produces a response. The response becomes training data. The training data helps shape the next recommendation.

But Netflix’s engineers also acknowledge that the meaning of a signal is not always clear. Watching ten minutes of a film might mean someone disliked it. It might also mean they were interrupted. Completing a series may suggest enjoyment, but completion followed by a negative rating tells a different story. Optimizing too heavily for a simple measure such as clicks can reward attention-grabbing choices without improving long-term satisfaction. 

The system does not uncover a single, clean object called “your preference.”

It builds a working representation from behavior that must first be interpreted.

On TikTok, the same loop becomes easier to notice because the distance between signal and response can be much shorter.

TikTok’s own explanation says a new user may begin with selected interest categories or a generalized feed of popular videos. Early likes, comments, replays, follows, and other interactions then help refine what appears later. The company also acknowledges that personalization can produce an increasingly homogeneous stream and says it deliberately introduces varied content and controls such as “Not Interested” to interrupt repetitive patterns. 

Your behavior shapes the feed.

The feed shapes your next opportunity to behave.

But the loop can also be redirected.

YouTube’s recommendation research reveals another complication: implicit selection bias.

The system does not receive reactions to every video a person could possibly have watched. It receives reactions to the smaller set of videos that were selected, ranked, and presented. 

A click is real behavior.

But it is not a neutral survey of every available possibility.

The system is not learning from every choice you might have made. It is learning from the choices available inside the environment it helped construct.

This is why recommendation systems do not need to create a preference from nothing in order to influence it.

They only need to make some options easier to encounter, observe what happens, and use the resulting behavior to shape the next set of options.

A pattern the system initially detected can become a pattern it repeatedly reinforces.

None of this means your preferences are fake.

People do not enter recommendation systems as blank profiles waiting to be written.

We bring memories, relationships, cultural influences, needs, moods, curiosity, and years of experience with us. We search directly. We reject recommendations. We discover things outside platforms. Our interests change when our lives change.

Nor does every interaction reveal a durable preference.

Exposure can change without changing consumption.

Consumption can change without becoming taste.

Taste can change without becoming identity.

Recommendation systems can also weaken their own loops by introducing unfamiliar material, prioritizing discovery, accepting explicit negative feedback, or allowing users to reset or exclude parts of their histories. Spotify and TikTok both describe mechanisms intended to introduce variety or give users greater control over future recommendations. 

The question is therefore not whether algorithms discover preferences or create them.

They can participate in both processes—but not equally, not for every person, and not in every system.

An algorithm may detect an existing interest, make it easier to act on, record the resulting behavior, and return a stronger version of the same interest. In other cases, it may introduce something genuinely new. Sometimes the recommendation is rejected, the user changes direction, or the system interprets the signal incorrectly.

The relationship is neither pure control nor pure reflection.

We enter the loop with preferences. The system responds to them, and then we respond to the environment it creates.

The next time Spotify recommends a song you never searched for, the familiar explanation may still be correct:

It has learned something about you.

But that is no longer the whole picture.

The system is also learning from choices made inside an environment it helped arrange.

Maybe your preferences are shaped not only by what you choose, but by what you are repeatedly given the chance to choose.

The Engagement-Diversity Connection: Evidence from a Field Experiment on Spotify — David Holtz, Benjamin Carterette, Praveen Chandar, Zahra Nazari, Henriette Cramer, and Sinan Aral. 

Humans + Machines: A Look Behind the Playlists Powered by Spotify’s Algotorial Technology — Spotify Engineering. 

Recommending for Long-Term Member Satisfaction at Netflix — Netflix Technology Blog. 

How TikTok Recommends Videos #ForYou — TikTok Newsroom. 

Recommending What Video to Watch Next: A Multitask Ranking System — Google Research.