The World Through a Screen: How Algorithms Change the Picture Available to Us

Epigraph

“He who knows only his own side of the case, knows little of that.”

John Stuart Mill, On Liberty, chapter II. Short quotation checked against the English text.

What this article is about

One person opens their phone and sees a succession of threats. Another receives advice, successful projects, and funny videos. Both may be talking about the same day in the same city, yet the picture available to them differs considerably. This does not necessarily mean someone is lying or inhabiting an entirely closed information world. Several selections stand between an event and a screen: what was published, whom we followed, what the platform allowed, how it ordered posts, and what held the reader's attention. The algorithm participates, but does not explain the process completely. To discuss its influence, we must distinguish displayed material, attention, belief, and action. This article examines those transitions, compares specific experiments, and suggests a way to check our own information routes. The aim is not a view without intermediaries. It is to understand how mediation works and avoid treating a convenient selection as a complete measure of the world.

I. Selection begins before the algorithm

A screen displays an already selected part of what happens. Before ranking, someone must notice an event, turn it into a post, and decide it deserves publication. Other events remain conversations, private photographs, working documents, or receive no record at all. Then come subscriptions, language accessibility, moderation, and platform format. A feed therefore does not become a complete picture of the world even with an ideal ordering of posts.

Imagine a neighborhood where a library opened, rents rose, and a parking dispute occurred. The dispute is easy to film in a minute. Understanding rising rents requires comparing contracts and speaking to different people. The library's benefits emerge gradually. Differences in speed and ease of representation can affect publication before any personalized recommendations. If the feed subsequently contains many conflicts, we cannot automatically attribute the entire imbalance to one program. We need to examine the available messages and how they were organized.

This does not absolve the platform of responsibility. Earlier selection and algorithmic selection can reinforce each other. But a precise explanation asks more specific questions: which messages were available, which were excluded, which gained an advantage, and who made those decisions? The word “algorithm” often covers a business objective, a moderation rule, a human habit, and a predictive model operating together. Conflating them makes it harder to understand what can change.

In everyday reading, distinguishing an absent event from an absent report helps. An acquaintance's silence about something important to you does not establish indifference. A topic's absence from your feed does not establish its rarity. A screen makes things accessible; it does not register everything that happens. Sometimes the first correction to our picture requires no battle with technology: we turn to a source where the relevant kind of event is routinely recorded.

II. A recommendation predicts a response, not truth

A recommendation system selects and orders material according to its design. A user's likely interest and a message's reliability are different properties. A system may consider both, but the first does not establish the second. A post's popularity, its accuracy, and its usefulness for a particular decision also differ. This is an important boundary between interface convenience and grounds for knowledge.

Meta's explanation of June 29, 2023 describes multiple systems working separately and together, rather than one mechanism governing the whole platform. The company describes system cards, processing steps, and available customization controls. This is primary evidence of the product's stated design. It is not an independent audit of all outcomes or a statement of every version's current parameters. For our purposes, the distinction suffices: explaining a technical process and evaluating a social effect require different data.

Suppose someone often pauses at alarming headlines to check whether a threat concerns their family. The observed pause does not automatically tell us whether they wanted more such content. A reaction may express interest, fear, disagreement, or a need to check an error. If a measure does not distinguish these reasons, the system uses a coarser signal than human intention. This is a logical limitation of measurement, not a claim about every developer's hidden motives.

Users can also mistake their behavior for a transparent expression of preference. “I don't like this” does not always match where our time goes. Observation is needed: what do I open, why, and what do I do afterward? A workable reading route requires these distinctions. They do not make someone responsible for the environment's design, but show where choices are available and where the product itself needs to change.

Recommendation and search deserve distinction, although both use algorithms. In search, someone usually asks a question and receives ordered answers; in a feed, the next question often arises from suggested material. This is a working distinction between routes, not a rigid boundary across all products. Search can suggest new interests, and a feed can serve a specific task. But the question's initial role changes what we check: am I finding grounds for a claim already identified, or gradually adopting an agenda I did not previously choose?

Imagine opening a phone to check the library's hours and, twenty minutes later, discussing someone else's conflict. There is no obligation to blame yourself for weakness: the environment offered another route, and the topic may have been interesting. Yet the initial task remains unfinished. Returning to it lets us evaluate the tool by its purpose. Enjoyment alone does not measure a platform's usefulness. One route can excel at discovering topics and poorly support finishing a bounded task. Choosing a more direct source then adjusts how we work; it does not reject digital life.

III. Displayed, noticed, and accepted are different outcomes

To establish a feed's influence, we must identify the outcome that changed. Showing material does not ensure attention. Attracting attention does not ensure persuasion. Changing a survey answer does not ensure a changed action. Each connection needs examination; studying one transition does not automatically establish the entire chain. This is why apparently conflicting reports about algorithms sometimes concern different quantities.

In an imagined scene, someone sees a controversial post, forwards it to a friend, and writes, “Look how ridiculous this is.” The system may register a view and a share. An observer who misses the comment might conclude the user supports the content. Another person says nothing but revises a decision. Public activity and private assessment can diverge in either direction. A precise conclusion requires more than an interaction counter.

Timescale also matters. One exposure, several weeks of use, and years within an environment are different interventions. A short experiment can reveal some changes and miss others. That is no reason to disregard it: it answers a bounded question, usually clearer than whether the internet has damaged society. Research restraint is useful when it specifies measurement rather than refusing to discuss consequences altogether.

Everyday discussion can follow the same order. Instead of “this platform makes me worse,” distinguish: concentrating is harder after reading; conflicts have increased; I assess risk differently; the available arguments have changed. Some points permit personal observation; others need external checking. One person does not represent all users, but their experience does not become worthless because it is not universal. It supports a bounded decision about personal use, not a final verdict on the digital world.

IV. An experiment can change a feed without changing beliefs

A null result for beliefs cannot be retold as the absence of any algorithmic influence. In Brendan Nyhan and colleagues' 2023 Nature study, involving 23,377 US participants during the 2020 election period, exposure to politically like-minded sources was reduced by roughly a third. Content exposure changed, but no measurable effect appeared on eight preregistered attitudinal measures. This is the result of a specific intervention, not a declaration that all feeds are neutral.

That result is more demanding, but more useful, than a slogan. If changing available information did not change the selected attitudes during that period, the simple hypothesis “remove some like-minded content and polarization will fall” received no support. Other mechanisms, longer effects, and differences between people and outcomes remain possible. Choosing among them requires further data. The absence of an effect does not identify the correct alternative story.

This also challenges everyday intuition. We sometimes imagine any disagreement can be resolved by presenting the right selection. But beliefs connect to previous experience, relationships, and a source's significance. New messages can broaden knowledge without changing political identity. Understanding another argument need not produce agreement. Demanding immediate change as proof of openness replaces inquiry with pressure.

Evaluating technology benefits from retaining several measures. Did source diversity change? Did a claim's origin become clearer? Was verification quicker? Did attitudes toward people with other views change? Choosing one measure makes sense for a specific task, but is dangerous as a universal substitute for quality. Finding information faster may help without changing a worldview. A major change of opinion may be worse if a convincing error caused it.

V. A different experiment finds a different transition

Germain Gauthier and colleagues' Nature article of February 18, 2026 describes a seven-week experiment with US users of X in July–September 2023. Switching from a chronological to an algorithmic feed shifted several measured political opinions toward more conservative positions. Switching in the opposite direction produced no comparable effect. No significant change appeared in affective polarization or self-reported partisanship. The authors also investigated account following as a possible part of the mechanism. These results concern a particular platform, period, and participants.

Comparison with the Facebook study does not justify choosing the convenient result and declaring the other wrong. The interventions, environments, and measured consequences differ. One changed the proportion of sources; the other changed the feed mode, with different starting conditions. The difference raises substantive questions about accumulated experience and the sequence of exposure. It weakens overgeneralization whether the formula defends platforms or condemns them.

The asymmetry matters. A simple analogy offered by the authors of this article can help: a recommended route brought someone to a new library, where they obtained a library card. Removing the recommendation does not revoke that card. The analogy illustrates how contacts might persist after changing modes; it does not replace the study's data or describe every user. An intervention and its reversal need not be mirror operations.

A practical decision may therefore extend beyond one switch. Someone can change their feed mode while retaining subscriptions, response habits, and conversational circles. Checking changes in the available picture should include these conditions. This guarantees neither changed opinions nor a single correct information diet. It identifies the part of the environment actually changed, avoiding the conclusion that all effort is pointless because one small step failed to produce a desired outcome.

VI. Chronology also organizes attention

Chronological order can offer a useful alternative to ranking, but remains an order of selection. It favors recent and frequent publication. Followed accounts determine the participants; login time determines the accessible segment; format affects noticeability. There is no view outside every condition. Still, knowing the rule can help: the reader can more easily understand why one post appeared above another.

Imagine an expert publishing a considered analysis weekly while a commentator posts brief reactions several times daily. The latter may occupy more of a chronological stream without personalized amplification. A reader seeking understanding may need a separate route to the expert material. This addresses attention, not the human worth of either author. Publication frequency does not measure quality or always express a wish to dominate.

Plato raises a related problem in Phaedrus, 275d–e: written speech appears responsive but repeats itself when questioned. Today a message can receive comments and its author can answer. The ancient argument therefore cannot apply literally to every modern interface. Yet distinguishing an accessible form from living knowledge remains useful. Opening a text does not ensure access to its grounds, and pressing a reaction button is not a conversation.

A more substantive criterion for an information route is whether it reaches an argument. Is the original material linked? Do we know who speaks and on what grounds? Can we read fuller context? These features do not guarantee truth, but enable checking. Choosing chronology or recommendations then becomes a configuration for a specific task: discovering something new, following acquaintances, researching a question, or checking a disputed claim. Different tasks may require different routes.

VII. Another opinion does not always make us more attentive

Information diversity should be considered alongside the form of encounter. Christopher Bail and colleagues' 2018 study tested exposure to opposing political messages through Twitter. Under the selected conditions, the result was not universal moderation; Republicans showed increased conservatism. This does not argue for closed information worlds. It limits the simple assumption that adding the opposite camp to a stream suffices. Context and format matter.

Imagine a comments dispute. Someone reads five harsh replies and concludes they have encountered the opponent's argument. Instead, they met particular participants constrained by reply length and mood. Elsewhere, the same question might be posed far more precisely. Selecting a strong objection differs from repeatedly reading irritating examples of the other side. The second habit can sustain the belief that the first task is already complete.

Mill's On Liberty, chapter II, urges us to learn the grounds of another position. His philosophical argument was no social-media experiment. It supplies a normative standard of intellectual honesty: reconstructing an objection in a form its advocate recognizes. An algorithm may locate such a text, but does not guarantee the corresponding reading. Participation requires time, a clear question, and the ability to delay reacting.

Practical choices therefore concern contact quality. Sometimes one substantial work is more useful than following dozens of contentious accounts. Sometimes direct conversation clarifies what a feed only aggravates. Yet no one must endure aggression indefinitely to prove broad-mindedness. Openness can examine arguments and set boundaries around interaction at once. It loses meaning when meeting someone is replaced by showcasing their worst replies or an obligation to tolerate any treatment.

VIII. A feed makes the particular look general

How often we encounter reports does not equal how prevalent the events are. Users usually do not know the underlying population, how many posts were excluded, or why similar examples recur. “This is everywhere now” therefore needs separate checking. The impression may be accurate, but its strength cannot establish accuracy. Distinguishing messages from independent cases is especially important.

Imagine seeing seven videos of an unusual conflict, then discovering they all use one clip posted by different accounts. Seven encounters represent one event and several editorial interpretations. Even different events do not establish prevalence without an appropriate basis for comparison. Reporting, official statistics, and personal accounts answer different questions. Rather than choosing whom to believe in general, ask what each source can establish.

Repetition also shapes urgency. An undated post may seem new; archival video enters today's sequence. Checking time restores distance. An old event does not become unimportant merely through age: it may still require investigation or remembrance. The error lies in changing its status, presenting the past as a newly discovered current threat.

Emotion retains a place in this reading. Anxiety, indignation, or compassion neither make a message false nor deprive someone of reason. But emotional significance does not replace provenance. We can acknowledge being moved and delay sharing until the date, author, and complete passage are clarified. Such a pause matters especially when consequences concern a real person. It requires no complete knowledge of the world; it prevents unchecked availability from becoming a confident accusation.

IX. Checking our route does not promise complete objectivity

A small audit of an information route should match the question. Local news benefits from dates, primary documents, and independent witnesses. A philosophical argument requires the complete case and a strong objection. Rest does not oblige us to research every video. Trouble begins when leisure quietly becomes the basis of an important decision.

Choose one topic and trace the last three messages: who reported them, where the information came from, what you checked, and why you opened them. This record neither reveals the platform's internal design nor establishes an algorithm's causal effect. It shows an accessible part of your own work. The messages may prove independent, or acquaintances may all repeat one unopened original. That distinction alone changes confidence.

Do not confuse diversity with artificial balance. A well-supported claim does not require equal numbers of denials for symmetry. A strong counterargument reveals limits; a random objection merely takes space. Nor are two sources independent because their logos differ. They may use one recording, press release, or sample. Checking provenance helps more than mechanical counting.

Another audit element is reading's consequence. Did it produce a distinction, a suitable question, or checked information—or only readiness for more debate? The latter need not be bad: people need entertainment and belonging. But if the goal was a decision, return to the original question. Sometimes closing the stream and finishing one text suffices; sometimes a person or document outside the platform is needed. Changing routes matters more than imagining one ideal feed can perform every task for us.

Readers remain responsible when sharing. A public retelling increases access for people who may never see the original context. “Just sharing” does not remove participation. We can retain a question's status: identify what remains unknown, avoid adding accusations, and link the full material. If clarification comes later, make it accessible to the same audience. This is ordinary care in communication, not a demand never to err.

Advice derived from a feed also needs scrutiny. Someone sees a vivid example and advises an acquaintance to change jobs, reject collaboration, or make a purchase. A message now supports an action with a real cost. Return to the acquaintance's conditions: does the situation apply, what resources exist, what do they want, and what alternatives are available? Another case can suggest a useful question, not a ready-made decision. Screens broaden our examples; they do not make lives interchangeable or authorize confident invention of missing facts.

X. Environment and participation change together

Discussion becomes poorer when it retains just two characters: an all-powerful platform and an entirely guilty user. Responsibility is distributed. Developers choose objectives and limits, organizations choose access and transparency rules, authors shape messages, and users select some reading routes. Their opportunities are unequal. A personal technique does not replace public demands concerning the environment, and a systemic problem does not cancel available individual steps.

Transparency matters when tied to a checkable question. Which signals are used? Which controls actually change selection? How can researchers evaluate consequences? How can mistaken decisions be challenged? An attractive explanation is insufficient, but “black box” should not end analysis either. Between total inaccessibility and complete disclosure are concrete information and checks that improve users' position.

For someone opening a phone in the morning, the starting point is modest: this message sequence tells me what was available to me, not how to measure the whole day. Choose one question worth leaving the stream for. The library, rent, and parking dispute remain different events with different grounds for knowledge. They need not receive equal screen space, but two disappearing cannot establish their nonexistence.

A good digital environment helps us find, distinguish, and check; good participation uses that help without surrendering all judgment. The world through a screen is inevitably limited. Limits become less dangerous when their workings can be discussed, researched, and changed. We can then debate particular consequences and choose effective measures while respecting what remains unknown. That is harder than blaming one program, but leaves more actual points of action.

What this means in practice

For several days, choose one bounded information question: a change in a local service, for instance, or an argument your acquaintances discuss. Save three encountered messages, dates, and original links. Note whether search, a subscription, recommendation, or another person brought them to you. These facts do not explain every reason for display, but create a checkable map of access.

Find a source capable of establishing the relevant fact. For a rule, this might be an official text; for research, the publication itself; for a philosophical claim, the author's work. Check whether several messages use one original. Separate confirmation, interpretation, and unresolved questions. No large report is required: a few sentences keeping those levels distinct suffice.

If different feed modes are available, change one and observe what appears. Do not call this an established experiment on political beliefs: a personal trial does not control every condition. Examine a particular outcome—whether finding material becomes easier, independent sources appear, or interruptions decrease. Account for subscriptions and login time separately. One switch does not automatically change the entire accumulated environment.

Finally, decide beforehand when reading will end. What knowledge suffices for the chosen decision? Without that condition, the stream can continue after clarification. If new uncertainty appears, formulate it separately. Returning to the original question lets entertainment remain entertainment, inquiry remain inquiry, and an important decision rest on grounds actually gathered for it. This practice seeks a clearer route, not perfect protection against every influence.

Key quotation

“… put a question to one of them, the speaker always gives one unvarying answer.”

Plato, Phaedrus, 275d, on written speeches. Short passage checked in Benjamin Jowett's English translation. Modern ways of responding differ; the quotation retains the question about the boundary between a message and knowledge.

Connections between traditions

Comparison	Similarity	Boundary
Plato, Phaedrus, 275d ↔ a feed message	An accessible sign does not ensure access to its grounds.	An online author can answer; the ancient critique of writing does not describe every interface.
Mill, On Liberty, II ↔ engaging with objections	Knowing a position requires encountering its strongest grounds.	A normative argument does not establish an algorithm's effect.
Nyhan et al. 2023 ↔ Gauthier et al. 2026	Field interventions distinguish environmental changes from attitudes.	Platforms, interventions, and outcomes differ; no universal effect follows.
Bail et al. 2018 ↔ Mill's argument	Encountering another position requires attention to its form.	The Twitter experiment neither refutes free discussion nor guarantees harm from every encounter.
Meta's technical description ↔ independent assessment	Both support an intelligible discussion of selection.	Design documentation does not replace examination of social consequences.

Primary sources and research

1. Meta. Introducing 22 system cards, June 29, 2023. Stated distinctions between systems and descriptions of customization controls; dated documentation.
2. Nyhan B. et al. Like-minded sources on Facebook are prevalent but not polarizing, Nature, 2023. Field reduction in exposure and limits of conclusions about attitudes.
3. Gauthier G. et al. The political effects of X's feed algorithm, Nature, 2026. Experiment conducted in 2023; switching asymmetry and separate measures of opinions and polarization.
4. Bail C. A. et al. Exposure to opposing views on social media can increase political polarization, PNAS, 2018. Limits of automatic moderation through opposing content.
5. Plato. Phaedrus, 275d–e, Benjamin Jowett's translation. Philosophical distinction between external text and knowledge.
6. John Stuart Mill. On Liberty, chapter II. Argument about knowing the grounds of an opposing position.

Tags

algorithms, screen, recommendations, ranking, feed, attention, information, sources, reliability, political attitudes, polarization, experiment, causality, chronology, subscriptions, context, digital environment, transparency, Mill, Plato, Meta, Facebook, X, Twitter, source independence, date, fact checking, media literacy, everyday decision, responsibility

About the authors

Andrey Zarinsh is the author and founder of United Wisdom. He explores inner transformation, the unity of traditions, intercultural dialogue, and contemporary spiritual discipline. Instagram: @andrey.zarinsh.

Maria Evilkut is a coauthor of United Wisdom. Alongside her research, she seeks to test these ideas in practice and observe how principles of unity and mutual understanding work in real life between people and cultures. Instagram: @ilkutki.