Does Personalization Serve Our Preferences or Gradually Create Them?

Personalization can do both. It can recognize an interest we already have and help us find more of it. It can also change what we encounter often enough that our habits, knowledge, and eventually some of our interests begin to move.

The two effects are difficult to separate because a personalized feed does not merely observe our choices. It supplies many of the choices we get to make next.

A system that learns from what we watch also influences what becomes available for us to watch.

First, Personalization Can Be Genuinely Useful

Imagine developing an interest in urban birds. You watch a short video about swifts nesting under a railway bridge. The next day, your feed includes a careful explanation of migration routes and a photographer who documents birds living among city buildings.

Those recommendations serve an interest you brought to the platform. Without them, you might not have known the photographer existed or found the right words to search for the migration study.

There is nothing inherently suspicious about learning from a person’s choices. A good bookseller does it. A friend recommending a film does it. Personalization becomes valuable when it lowers the effort needed to find material that matters to someone.

But the bookseller analogy has a limit. A social feed can make thousands of recommendations, measure the response to each one, and quietly adjust the next selection. Its influence is continuous.

Then the Available World Begins to Narrow

Suppose the bird videos that attract your attention most reliably involve dramatic rescues. The system notices that pattern and shows more rescues. You watch them, partly because they are moving and partly because they are now the bird videos placed in front of you.

After several weeks, the feed contains fewer migration maps, field recordings, and patient observations of ordinary birds. You have not pressed a button saying, “Remove everything except rescue videos.” Yet your visible version of the subject has changed.

If someone then measures your preferences from what you watch, the evidence will appear to confirm the system’s choice. You watch many rescue videos because you like them. You also watch many because they are what you are repeatedly offered.

This is an exposure feedback loop. Past behavior influences recommendations; recommendations influence what behavior is possible; the new behavior becomes evidence for the next round.

What Research Can—and Cannot—Show

Researchers Brandon M. Stewart, Barbara E. Engelhardt, and Allison J. B. Chaney studied this problem in work on algorithmic confounding in recommendation systems. They examined how systems trained on responses to earlier recommendations can create feedback loops, increasing the similarity of what users encounter and reducing the usefulness of recommendations under the conditions they studied.

That research identifies a mechanism. It does not establish that every social feed changes every person’s deepest beliefs, or that every repeated recommendation is harmful. The narrower lesson is enough: once a system affects exposure, the behavior it later observes is no longer an independent record of what the person would have chosen from an unrestricted set.

Another line of research asks whether exploration can improve recommendations. Google researchers studying recommender systems examined accuracy, diversity, novelty, and serendipity rather than treating immediate prediction as the only measure of quality. The point is not that novelty must always win. It is that repeatedly offering only what already looks familiar can prevent a system from learning what else a person might value.

Creation Is Not the Same as Manipulation

Our preferences are never formed in isolation. A teacher introduces a subject. A friend lends us a novel. A radio station plays a musician we have never heard. Exposure can create genuine appreciation.

If the bird feed introduces you to conservation work and you decide to volunteer, the platform has influenced you. That does not mean the interest is false. People routinely become interested in things because they encounter them.

The concern is not that preferences change. It is whether the person can see and shape the conditions of that change. A friend can explain why they recommended a book. You can tell them you want something entirely different. A recommendation feed may offer far less clarity about why one path keeps appearing and another has disappeared.

Influence becomes more troubling when the system repeatedly favors material because it produces measurable attention while the person would prefer a broader or calmer experience.

The User Changes the System Too

It would be inaccurate to describe people as passive objects being molded by a machine. You can search for migration studies, follow a field biologist, choose a chronological feed where available, or mark repetitive rescues as unwanted. You can leave the app and visit a library or a park.

But those actions require noticing the pattern and spending effort to correct it. Watching the next video requires almost none. This difference in effort gives defaults considerable power.

Nor does the system know your entire life. Your interests can change through work, friendship, travel, illness, or curiosity that has never appeared in your viewing history. A feed trained mainly on yesterday’s behavior may be slow to recognize the person you are becoming.

What Would Serving a Preference Require?

A system serving preferences should help people pursue interests they can recognize and revise. It should leave room for discovery beyond the patterns it has already detected. It should let a person say, “I enjoyed that once, but I do not want my feed to revolve around it.”

A system that gradually creates preferences will often look similar from the outside: people watch, return, and watch again. The difference becomes clearer when we ask whether they are encountering a range of meaningful possibilities or being guided deeper into whichever narrow pattern produced the strongest response.

Personalization is neither a perfect mirror nor an irresistible force. It is part of a relationship in which person and system continually respond to one another. The question is who has the better opportunity to redirect that relationship.

And that leads to a sharper boundary: at what point does a recommendation stop feeling relevant and start functioning as manipulation?