It is 7:14 a.m. Two people stand in the same tram, a few meters apart, and both open their phones. Within five minutes, the first sees an attack in the street, a political argument, an economic threat, and a video claiming that “the world has completely lost its mind.”
The second sees a French bulldog, a new artificial-intelligence tool, a cheap flight, and a short story about someone who changed careers.
None of those posts has to be false.
And yet after a few weeks, the two may come away with entirely different ideas of what the world is like, what people are dealing with, what we should fear, and where society is heading.
Facts alone did not create the difference.
Their selection did.
A recommendation algorithm is not an editor in chief in the human sense. It has no convictions, conscience, or political views of its own. Functionally, it performs tasks that were the privilege of editors for most of history: it selects candidates, ranks them, repeats some, pushes others aside, and never shows the vast majority.
And it does this separately for every person.
The Editor No One Appointed
A traditional newsroom works with limited space. A front page can hold only a few stories. An evening broadcast lasts only a few dozen minutes. The editor decides what is important, timely, trustworthy, or interesting.
A digital platform has the opposite problem: content is not scarce. There is too much of it.
Without selection, a social network, video platform, or online store would be unusable. A recommendation system is not an add-on. It is the basic infrastructure of attention.
TikTok official guidance says that the For You feed ranks videos using signals including user interactions, video information, and account settings. Watching a longer video to the end can be a stronger signal than the viewer and creator being in the same country.1
YouTube likewise cites watch and search history, subscriptions, likes, dislikes, “not interested” choices, and satisfaction surveys. The home page relies mainly on watch history, while the “next video” recommendation draws heavily on the video a person is currently watching.2
The details differ across platforms. The basic logic is similar:
to estimate, from an enormous number of possibilities, what is most likely to be relevant to a particular person at a particular moment.
The Algorithm Does Not Know What Matters
A machine cannot optimize an undefined sentence such as:
“Show a person what matters most in their life over the long term and also benefits society.”
First, someone would have to decide what “important,” “benefits,” and “over the long term” mean. Then those values would have to become measurable signals.
This is where one of the biggest differences between a human goal and a machine goal emerges.
Truth is a value.
Trustworthiness is a value.
Well-being is a value.
Plurality is a value.
For a model, it must become a number, category, prediction, or rule.
This does not mean that platforms optimize only time or click counts. Current systems use multiple signals, including feedback, safety, and stated satisfaction. They nevertheless always work with measurable proxy indicators.
A watched-to-completion video is not the same as a valuable video.
A comment is not the same as agreement.
A long pause is not always interest. It may be shock, disgust, or an attempt to understand what a person is seeing.
But to the system, every observable behavior is a trace.
The key problem: an algorithm does not directly see the meaning content has for a person. It sees a proxy—behavior that may be related to that meaning but may not represent it perfectly.
The Algorithm Learns from You. You Learn from the Algorithm.
Personalization is often described as a one-way process:
You have interests → the system learns them → it shows you suitable content.
In reality, the relationship is bidirectional.
The algorithm watches what you do. It changes the next selection accordingly. The new selection affects what you pay attention to. Your new reaction becomes another signal. And the cycle continues.
If you pause at a video about a crime case, the system may try another. If you pause again, the category gains weight. After a week, the feed may seem as if crime has become the main concern of society.
Maybe society has not changed.
What changed is the probability that you will receive this particular topic.
This second effect is often overlooked.
The algorithm is not just an editor of finished content. It gradually influences what content comes into existence at all.
TikTok creator guidance openly states that watch time feeds into recommendations and that it is beneficial to hold attention from beginning to end.1 That is rational advice. It is also a cultural incentive.
When a fast hook works, more fast hooks are made.
When conflict drives reactions, more conflict is made.
When nuance slows the pace, it may be cut—not because it is untrue, but because it loses the competition for continued attention.
A Major Intervention Can Be What We Do Not See
When an editor publishes a biased headline, we can criticize it.
With an algorithmic feed, part of the editing is invisible.
We do not know how many alternative posts could have occupied the same place. We do not see the article that scored a few points lower. We do not know that someone else was offered the opposite interpretation of the same event at the same moment.
Absence is the hardest thing to analyze.
We cannot disagree with an argument we were never given.
We cannot correct an error we do not know exists.
We cannot notice that a topic is not a society-wide obsession—but merely the dominant theme of our own feed.
This is close to the classic media agenda-setting idea: media do not have to tell people directly what to think; they can strongly influence what people think about. A personalized feed takes this logic further because the agenda no longer has to be shared.
The Bubble Is Not Just the Algorithms Work
The story of the “evil algorithm” has one advantage: it leaves us innocent.
But people chose like-minded friends, newspapers, and communities long before social networks. We follow accounts that interest us. We click topics that confirm our fears or identity. We can block an opponent or skip an uncomfortable opinion.
The algorithm often does not create preferences from nothing. It amplifies, refines, and automates a selection in which the user participates.
A major Facebook study published in Nature in 2023 found that most content seen by American adults came from politically like-minded sources, although political and news content made up only a small share of total exposure. Experimentally reducing content from like-minded sources by roughly one third changed the composition of the feed but produced no measurable changes in eight prespecified measures of attitudes and polarization.3
A related experiment with chronological feeds on Facebook and Instagram during the 2020 US election also substantially changed what people saw and how they used the platforms, but found no detectable changes in the measured political attitudes.4
This is an important brake on a simplistic claim:
“Just turn off the algorithm and society will stop polarizing.”
It is not enough.
But these studies do not support the opposite conclusion either: that algorithms have no political or social effects. They measured specific platforms, interventions, periods, and outcomes.
Why the Research Seems to Contradict Itself
In 2026, Nature published an independent randomized study of the algorithmic feed on X. Active users in the US were assigned to an algorithmic or chronological feed for seven weeks.
Turning on the algorithmic feed increased engagement and shifted some attitudes about political priorities and current events in a conservative direction. The study found no significant change in reported partisan identity or affective polarization.5
That same year, Nature published an audit of TikTok during the 2024 US election. Researchers ran 323 audit experiments with controlled accounts and collected more than 280,000 recommendations. They found systematic asymmetries in political exposure between accounts trained on Democratic and Republican content.6
But this audit did not measure whether real peoples views changed. The authors also state that the data cannot determine precisely whether the difference was caused by algorithmic rules, the available content supply, or another part of the system.
Another field experiment published in Science in 2025 reranked posts on X in real time according to whether they expressed antidemocratic attitudes and partisan animosity. Increasing or reducing exposure to such content shifted evaluations of the political opposing side by more than two points on a 100-point scale.7
No Secret Plan Is Necessary
The algorithm does not have to “want” to polarize society.
There does not have to be a person who issued the command:
“Show everyone a more extreme world.”
A systematic effect can arise without systematic ill intent.
All it takes is a goal that meets human psychology, the available content, and the business model in a particular environment. If one type of post reliably generates watching or interaction, it may gain distribution even when its social effect was not an explicit goal.
It is equally important to acknowledge the other side.
Personalization is not inherently bad.
It helps people discover small creators, expert material, music, communities, job opportunities, or topics that would never reach the screen in a chronological ocean. TikTok acknowledges the risk of a homogeneous “filter bubble” and describes interventions intended to introduce more diverse content and interrupt repeating patterns.1
So the question is not:
“Algorithms: yes or no?”
The question is:
“According to what values does it edit, what do we know about that editing, and how much control do we have over it?”
Europe Begins Demanding a Different Kind of Editing
The European Digital Services Act—the DSA—treats recommendation systems as a matter of public interest, not merely as an internal product secret.
Very large online platforms and search engines must be transparent about recommendation systems, assess systemic risks, provide vetted researchers with access under certain conditions, and offer at least one recommendation option not based on user profiling.8
As part of the European implementation of the DSA, TikTok announced an option to turn off personalization; its non-personalized For You feed then uses popular content from the region and the world, while followed-account feeds are ordered chronologically.9
But even a “non-personalized” feed is not without an editor.
Popularity is an editor.
Chronology is an editor.
Followed accounts are an editor.
The difference is whether a person knows the selection principle and can switch between principles.
How to Take Back Part of the Editing
For most people, leaving the algorithmic environment entirely is unrealistic—and often undesirable.
It is far more practical to stop treating the feed as a neutral window.
A Small Audit of Your Own World
- Distinguish between “it is everywhere” and “it is everywhere in my feed.”
- For an important topic, actively seek out a source instead of waiting for a recommendation.
- Use “not interested” options, topic blocking, and history management.
- Occasionally switch to a chronological or non-profiled feed, if available.
- Keep direct sources outside the social network: newsletters, RSS, and institutional websites.
- Notice which emotion keeps you on the screen—interest, fear, or anger.
- Compare your first picture of an event with one source you do not usually follow.
- From time to time, deliberately look for a topic the algorithm had no reason to offer you.
TikTok lets users refresh the feed, use “not interested,” manage topics, and filter keywords. YouTube lets users delete or turn off history, fundamentally changing the basis of personalization.2 These tools are not perfect, but they remind us of something important:
the user is not just an audience. The user is also a source of editorial signals.
The World Does Not Fit in a Feed
The algorithm solves a real problem.
Among millions of possibilities, it must choose a few.
Without it, we would get lost in digital space.
But every simplification of the world is also an interpretation of it.
Order creates importance.
Repetition creates normality.
Omission creates invisibility.
Personalization creates the feeling that what I see often is seen often by everyone.
A serious danger may not be that an algorithm directly puts a particular opinion into peoples heads.
Its power is far subtler when it prepares a different list of questions for each person to think about.
We do not need to demonize algorithms.
We do need to see them for what they functionally are:
as editorial systems with goals, blind spots, and consequences.