The Curated Public: How Algorithms Govern What Becomes Visible
Who decides what becomes visible before choice begins.
Abstract
Digital platforms do not simply help people navigate an abundance of information. They decide which fragments of that abundance become visible, in what order, and under what conditions. Yet claims that algorithms directly determine what users think often outrun the evidence. Recommendation systems influence exposure and attention more reliably than belief, and their effects vary across platforms, populations, and political contexts. This essay argues that the democratic problem is therefore not algorithmic mind control but privately governed visibility. Platforms can shape what appears relevant, credible, urgent, or socially normal before conscious judgment begins. Algorithmic literacy is an important civic defense, but it cannot correct structural asymmetries on its own. Meaningful reform requires user choice, independent scrutiny, public-interest access to platform data, and institutions capable of contesting how attention is allocated.
Choice Begins Before the Click
Opening a social media application feels like an act of choice. We scroll, pause, follow, ignore, search, and leave. Nothing forces us to read one post rather than another. The interface presents a continuous sequence of small decisions, each apparently our own.
But the most consequential decision has already been made. Before a user sees the first item, a system has selected it from a much larger field of possibilities. It has estimated what the user may watch, share, buy, resent, or return for. It has also decided what will remain absent.
This does not make choice unreal. It makes choice conditioned. In a post-perception era, that conditioning matters twice: before users can ask whether something is authentic, a ranking system has already decided whether it will enter their field of perception at all.
The distinction matters because public debate often swings between two exaggerated positions. One treats recommendation algorithms as machines of psychological control, capable of inserting beliefs into passive minds. The other argues that users remain free to click elsewhere, so the platform is merely reflecting preferences that already existed. Neither account is adequate. People are not programmable objects, but neither do they choose from a neutral or complete menu.
Algorithmic power operates earlier. It shapes the environment in which judgment takes place. It affects which subjects seem important, which voices appear popular, which claims are repeated, and which alternatives must be actively sought. The central democratic question is not whether algorithms can make everyone believe the same thing. It is who governs the conditions under which something becomes visible, salient, and credible at all.
What Recommendation Systems Actually Do
There is no single “algorithm.” Search ranking, social feeds, video recommendations, advertising systems, and content moderation perform different functions. They use different signals and pursue different combinations of objectives. Treating them as one machine obscures both how they work and where responsibility lies.
Still, most large-scale recommender systems share a basic structure. They collect behavioral signals, such as viewing time, clicks, reactions, follows, skips, and reports. Models then predict which items are most likely to produce selected outcomes. Those outcomes may include engagement, retention, user satisfaction, safety, advertising value, or some weighted combination of them. The system ranks content, observes the response, and uses that response as new training data.
This feedback loop operates through three linked forms of influence.
The first is exposure: the system changes what a person encounters. The second is salience: repetition, ordering, and social signals affect what feels important or normal. The third is adaptation: users alter their behavior in response to the environment, while the environment is simultaneously adjusted in response to them. Adaptation is not merely a third channel; it is the cumulative mechanism through which the first two compound over time.
These mechanisms should not be confused with persuasion. A prominent item may be rejected, mocked, or forgotten. The empirical record is mixed. Research on Facebook found that both algorithmic ranking and users’ own choices reduced exposure to ideologically cross-cutting material, with individual selection playing a substantial role1. Later large-scale experiments found that reducing exposure to like-minded sources altered what users saw but produced no measurable change across several attitudinal outcomes2.
Other studies have found effects under different conditions. A 2026 randomized field experiment on X compared algorithmic and chronological feeds over seven weeks. Among users who had previously used a chronological feed, switching to the algorithmic feed shifted some political attitudes and account-following behavior. Among users who had previously used the algorithmic feed, switching to a chronological feed produced no comparable attitudinal effect. Neither intervention detectably changed affective polarization or self-reported partisanship. In that setting, the algorithm promoted more conservative content and more posts from political activists while demoting posts from traditional news media. Users newly exposed to it were more likely to follow conservative political activists, altering even the chronological feed generated by the accounts they now followed3. The asymmetry suggests that removing a ranking mechanism does not necessarily undo the behavioral changes it helped produce.
The apparent contradiction is instructive. Algorithms do not have one universal political effect. Results depend on the platform, the ranking model, the available content, the user population, the period studied, and the outcome being measured. The responsible claim is not that algorithms always polarize or persuade. It is that they possess the capacity to reorganize exposure at scale, and that this capacity can produce persistent behavioral effects and, under some conditions, attitudinal effects.
Beyond the Filter Bubble
The “filter bubble” remains an appealing metaphor because it gives a simple shape to an invisible process. A person is surrounded by congenial information and protected from disagreement. Yet the evidence does not support such a uniform picture.
Some users inhabit highly partisan information environments. Many do not. People may encounter opposing views and become more hostile rather than more moderate. In a well-known field experiment, exposure to messages from political opponents increased polarization among Republican participants, while the effect among Democrats was smaller and not symmetrical4. Exposure alone does not create deliberation. The manner, source, and emotional framing of disagreement matter.
The bubble metaphor also places too much emphasis on isolation. A person need not be sealed off from alternative views for algorithmic curation to matter. A feed can include multiple perspectives while repeatedly presenting one as mainstream, another as ridiculous, and a third as too marginal to show. It can widen the range of content while narrowing the range of content treated as worthy of attention.
The more useful frame is to ask how visibility is governed. Scholars have long observed that platforms construct regimes of visibility: Taina Bucher showed over a decade ago that algorithmic power on social platforms operates less through permanent surveillance than through the threat of invisibility, the constant possibility of disappearing from view5. The question pursued here shifts the perspective from how that regime is experienced to how it is governed: by whom, under what objectives, and with what accountability. Platforms exercise power by continuously allocating attention among competing people, claims, and events. This resembles editorial judgment, but with several important differences. The ranking is personalized, automated, dynamically revised, commercially embedded, and largely inaccessible to outsiders. Two citizens can use the same service on the same day and receive different accounts not only of what an event means, but of whether the event appears at all.
A perfectly shared public sphere never existed. Newspapers, broadcasters, parties, and social groups have always filtered reality. What changes is the scale, speed, opacity, and intimacy of the filtering. Algorithmic gatekeeping can generate millions of individualized editorial environments without leaving a stable public record of why a particular item was promoted.
The result is less a set of sealed bubbles than a public divided by unequal maps of relevance.
Why Engagement Changes Public Speech
Platforms do not need to prefer falsehood for falsehood to prosper. They need only reward the behavioral features that false or inflammatory material can exploit.
Emotion is one of those features. Brady and colleagues found that moral-emotional language was associated with greater diffusion in political messages on Twitter, although the study measured sharing patterns rather than proving that a ranking algorithm caused the effect6. Later work showed that social reinforcement can teach users to express more moral outrage when such expression is rewarded7.
Novelty is another. Vosoughi, Roy, and Aral’s analysis of Twitter cascades found that false news traveled farther, faster, deeper, and more broadly than true news in the dataset they studied. False news was also perceived as more novel than true news, which helps explain why people were more likely to share it. Human sharing behavior, rather than automated bots alone, was central to the difference8.
Neither finding proves that all platforms maximize outrage or misinformation. It does show why systems trained on observable reactions can privilege destructive content without being explicitly instructed to do so. If indignation holds attention, optimization can amplify it.
This changes more than individual information diets. It changes the incentives of public speech. Politicians, journalists, activists, influencers, and ordinary users learn which forms of expression travel. Nuance becomes harder to distribute. Certainty outperforms hesitation. The most uncharitable representation of an opponent can become the most efficient version of that opponent.
A democratic society can survive disagreement. It depends on disagreement. What it cannot easily sustain is a communication environment that repeatedly converts disagreement into identity threat while making the conversion profitable.
Still, caution is necessary. Digital media can also increase political participation, widen access to information, and provide channels for dissent, especially where traditional institutions are closed or captured. A systematic review of 496 studies found both democratic benefits and harms, with effects varying by political context and by the outcome examined9. The problem is not that digital communication is inherently anti-democratic. It is that its governing infrastructure is optimized and altered by private actors whose choices can reshape public life without equivalent public accountability.
Algorithmic Literacy, Without the Alibi
Citizens need a practical understanding of these systems. They should know that a feed is a prediction, not a mirror; that popularity can be manufactured or selectively displayed; that repetition is not corroboration; and that the absence of a story from a feed says little about its importance.
Such literacy has at least three dimensions.
Procedural literacy concerns how ranking works: behavioral data become predictions, predictions become ordering, and reactions return as data.
Evidential literacy concerns the content itself: who produced it, what evidence supports it, what incentives shaped it, and whether independent sources confirm it.
Strategic literacy concerns action: deliberately consulting sources outside a habitual feed, using non-profiled or chronological options where available, separating discovery from verification, and recognizing when an emotional reaction is being converted into an engagement signal.
These skills matter. But literacy can become an alibi for institutional failure. A person cannot continuously audit a system that runs thousands of hidden ranking decisions on their behalf. Nor should the public be expected to compensate, through vigilance alone, for designs created by organizations with vastly greater data, expertise, and experimental capacity.
The asymmetry is fundamental. The platform can test countless interface variations across millions of users. The individual sees only one interface and rarely knows that an experiment occurred. Telling users to “think critically” is therefore necessary but structurally inadequate. It protects agency at the margin while leaving the architecture untouched.
Governing Visibility
Regulation has already moved beyond abstract calls for transparency. In the European Union, the Digital Services Act10 requires very large online platforms and search engines to undergo independent audits at least annually and, where they use recommender systems, to provide at least one option not based on profiling. Under specified conditions, platforms must also grant regulators and vetted researchers access to data needed to study systemic risks, and authorities may request explanations of the design, logic, functioning, and testing of algorithmic systems.
These are significant advances, but compliance is not accountability. A platform may describe its “main parameters” without showing how they interact. An audit may confirm that a risk process exists without proving that it protects democratic discourse. A non-profiled feed may offer formal choice while remaining buried or inferior by design.
Visibility is governed in a second sense as well. A 2025 preprint by Zach Bastick describes control over who is allowed to observe and measure platformed publics: paywalled APIs and restricted research tools have enclosed the data on which public-interest scrutiny depends11. The two questions are complementary. This essay asks what platforms make visible to citizens; that work asks who is allowed to see the platform itself. Both are preconditions of accountability, and both are currently decided in private.
A stronger framework should test algorithmic systems against four principles.
Legibility: users and overseers should be able to understand the main forces shaping what is shown, not merely receive a generic description of personalization.
Contestability: users should be able to inspect, modify, or reset the signals used to personalize their feeds; researchers and public authorities should be able to challenge systemic ranking effects, risk assessments, and the evidence used to justify them. The two levels are distinct because a person cannot easily contest what they never knew was withheld: invisibility rarely produces an individual decision that can be appealed.
Plurality: people should be able to choose among genuinely different modes of ordering, including options that do not depend on behavioral profiling. Choice between interfaces controlled by the same provider is not full pluralism, so interoperability and independent recommendation layers also deserve consideration.
Accountability: those who allocate attention at societal scale should be required to demonstrate what they optimized, what risks they measured, what trade-offs they accepted, and what changed after harm was identified.
The object of regulation should not be a mythical neutral algorithm. No ordering system is neutral. A chronological feed favors frequent posters and the rhythms of publication. A popularity-based feed favors existing attention. Human editors bring their own institutions and biases. The democratic objective is not neutrality, but governable power: visible objectives, contestable decisions, credible evidence, and consequences when declared safeguards fail.
Conclusion
Algorithms do not remove human agency, and they do not need to.
Their power lies in governing the field before agency is exercised. They select from abundance, assign prominence, learn from reaction, and alter the environment in which the next reaction occurs. Sometimes this changes attitudes. Sometimes it changes only exposure or behavior. Sometimes it produces no measurable political effect at all. The uncertainty does not make the power trivial. It makes independent scrutiny indispensable.
The language of mind control is therefore both inaccurate and politically unhelpful. It invites easy dismissal because people know they are not automatons. The more defensible concern is also the deeper one: a small number of private systems now mediate visibility for large parts of public life, while citizens have limited knowledge of their objectives and limited ability to contest their operation.
Algorithmic literacy can help people recover a measure of distance from the feed. It cannot, by itself, rebalance the institutions behind it. That requires rights, audits, access to evidence, plural modes of recommendation, and public authorities capable of questioning not only whether a system follows its own rules, but whether those rules are compatible with democratic life.
The stakes rise further when perception is no longer sufficient to distinguish what is authentic from what is fabricated. Curation operates one step earlier still: it decides what is perceived before perception can be tested at all. In a post-perception era, the question of who governs visibility precedes the question of what is true.
The question is not simply whether algorithms shape what we think. It is who decides what receives the chance to be thought about.
Sources
- Bakshy, E., Messing, S., & Adamic, L. A. (2015). Exposure to ideologically diverse news and opinion on Facebook. Science, 348(6239), 1130–1132. ↩
- Nyhan, B., et al. (2023). Like-minded sources on Facebook are prevalent but not polarizing. Nature, 620, 137–144. ↩
- Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E. (2026). The political effects of X’s feed algorithm. Nature, 652, 416–423. ↩
- Bail, C. A., et al. (2018). Exposure to opposing views on social media can increase political polarization. Proceedings of the National Academy of Sciences, 115(37), 9216–9221. ↩
- Bucher, T. (2012). Want to be on the top? Algorithmic power and the threat of invisibility on Facebook. New Media & Society, 14(7), 1164–1180. ↩
- Brady, W. J., et al. (2017). Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences, 114(28), 7313–7318. ↩
- Brady, W. J., et al. (2021). How social learning amplifies moral outrage expression in online social networks. Science Advances, 7(33), eabe5641. ↩
- Vosoughi, S., Roy, D., & Aral, S. (2018). The spread of true and false news online. Science, 359(6380), 1146–1151. ↩
- Lorenz-Spreen, P., Oswald, L., Lewandowsky, S., & Hertwig, R. (2023). A systematic review of worldwide causal and correlational evidence on digital media and democracy. Nature Human Behaviour, 7, 74–101. ↩
- Regulation (EU) 2022/2065 of the European Parliament and of the Council of 19 October 2022 on a Single Market for Digital Services. ↩
- Bastick, Z. (2025). Governance of visibility: How social media interfaces allocate power. Preprint. doi:10.13140/RG.2.2.26918.18245. ↩