Platforms and Algorithms of Attention

Attention is scarce, and the feed is a selection machine

There is far more online content than anyone can inspect. Selection is therefore unavoidable. A chronological feed selects mainly by time. A search engine ranks results in relation to a query and other criteria. A social platform may rank according to the probability of a click, watch time, reaction, share, return visit or some combination of many signals.

A recommender system is therefore not a neutral pipe through which an already-formed public opinion simply flows. It is a mechanism for allocating a scarce resource: human attention. Even if the platform did not create the content, it can determine which item receives the first, second or hundredth chance to be noticed at all.

That power is not inherently malicious. Without ranking, many large platforms would be almost unusable. The problem begins when we do not know what the system optimises, which behaviours count as success, or whether users can meaningfully change the mode of selection.

One useful provocation carried into this article from the Green Book is the criticism of assuming that a private medium automatically speaks for the public. The platform version is sharper still. A trending list, recommendation page or personalised feed does not necessarily show what “society thinks”. It shows the output of selection rules, user signals, network structure and the business design of a particular system.

This does not make trends fictional. They may measure a real rise in attention or activity. But they measure activity inside a particular system, using criteria someone has chosen. What becomes visible may represent a broad group, a highly active minority, or a topic that recommendation systems further amplify.

We therefore need to distinguish three things: what people believe, what people actually post or click, and what the platform selects from those actions for further visibility. Collapsing the three can turn an algorithmically curated stream into a false mirror of society.

What does a recommender algorithm actually optimise?

A recommender system does not normally search for an abstract category called “the best content”. It works with one or more measurable objectives. It may estimate the probability that we will click, watch longer, comment, buy, return or rate an item as relevant. The balance between these signals can vary by service, user and business model.

This matters because the objective function shapes the environment. If retention is central, the system searches for what retains. If satisfaction is central, different signals are needed. If conversion is the goal, the ranking problem differs from a system whose purpose is to answer a question as efficiently as possible.

Digital sovereignty does not require every user to read the source code of every model. It does require an intelligible answer to a basic question: which parameters matter most in deciding why I see this item, and what can I change? Article 27 of the EU Digital Services Act translates precisely this kind of logic into a transparency obligation.

Algorithms often infer preference from what we do. But a click is not the same as a considered desire. We may open content from anger, shock, curiosity or to verify something we reject. Long viewing time can mean fascination or confusion. Sharing can express approval or warn other people.

Research therefore distinguishes revealed behavioural signals from what users say they actually value. A study auditing engagement-based ranking on Twitter/X, for example, found that compared with a reverse-chronological baseline it amplified emotionally charged and out-group-hostile political content, while users did not necessarily say they preferred the political tweets selected by the algorithm.

That does not prove that every system always rewards the most conflictual content. It reveals a more general problem: a proxy is not the goal itself. A system that measures a person mainly through behaviour may optimise whatever provokes a response rather than what the person would choose, on reflection, as a good use of attention.

From observing behaviour to a feedback loop

“Who Controls the Data About Us?” treated data as a record of behaviour. This article adds the loop. The system records where we pause and uses that signal to change what it serves next; the new selection affects the next behaviour, which then becomes another signal for the system.

Personalisation is therefore not merely a mirror of past preference. It can become a dynamic relationship between the user and the environment, with each influencing the other. The user trains the system through action, while the system changes the set of options among which the user acts.

That loop can be useful: it can surface music, specialist sources or people we would otherwise miss. But it can also narrow exploration if the system concludes too quickly what we want and repeatedly returns similar stimuli. The important question is whether a user can interrupt, reset, broaden or redirect the loop.

“How Media Shape What We Think About” examined the agenda a medium places before an audience. On a platform, two people can open the same application at the same moment and receive radically different information environments. Personalisation disperses the front page, but it does not necessarily disperse the power to construct it.

This matters for common discussion. People may disagree not only about interpretations of the same events but already about which events exist in their everyday field of attention. Algorithmic ranking can increase or decrease certain forms of exposure, although effects also depend on whom users follow and how their social networks are structured.

Research on Facebook and Instagram during the 2020 US election found that moving participants to reverse-chronological feeds substantially changed platform use and content exposure, yet did not significantly change several measured political attitudes over the three-month study period. This is an important limit: algorithms can strongly shape information experience without that automatically implying total control over belief.

Amplification is not the same as programming a person

The strongest explanation is often the most tempting: the algorithm programs us. That formula is too crude. Exposure also depends on our initial choices, social network, language, location, interest groups, timing and the behaviour of other users. Research commonly reveals an interaction between algorithmic selection and human behaviour rather than one-way remote control.

Rejecting determinism does not make amplification irrelevant. If a system places one signal in front of people a million times more often, it changes the ratio between what exists and what becomes visible. That can affect reach, revenue, status, political agenda and the likelihood that a behavioural pattern is repeated.

A more defensible formulation is: a platform need not determine what we think in order to change the probabilities of what we see, how much time we give it and which responses are rewarded with additional visibility. That is substantial power and it deserves transparency and meaningful user choice.

When the interface competes for time, the algorithm is not the only issue

Attention is shaped not only by content order but by interface design. Infinite scroll removes a natural stopping point. Autoplay reduces friction between one item and the next. Push notifications carry the demand for attention out of the application and into everyday life. Visual emphasis and social cues can increase urgency.

The OECD and the US Federal Trade Commission use the language of dark patterns for forms of digital choice architecture that can steer, deceive or impair autonomous decision-making. The European Parliament’s work on addictive design has likewise focused on the attention economy and features intended to extend engagement.

Caution is still necessary. Long use is not automatically evidence of harm, and infinite scroll is not unlawful in every context. But where commercial success depends heavily on additional minutes, interactions or ad impressions, users should understand that their ability to stop may sit in direct tension with the service’s optimisation objective.

The business model helps reveal whose attention creates value for whom

Digital platforms are not all financed in the same way. Some depend mainly on advertising, others on subscriptions, commissions, sales, licences or combinations of these. It is therefore inaccurate to claim that every platform always maximises time spent. Sometimes conversion, retention, task completion or answer quality is more important.

Still, the business model helps explain incentives. Where more attention means more ad impressions and more behavioural data, there is a strong reason to optimise retention. A subscription model may align incentives more closely with satisfaction, yet subscription services can also use designs that complicate exit or encourage over-consumption.

This article therefore avoids the slogan “if it is free, you are the product”. The sharper question is: which outcome creates revenue, which metrics does the system reward because of that, and where can the provider’s interest diverge from the user’s long-term interest?

Transparency is not source-code disclosure but the ability to understand and influence

Publishing the entire source code of a recommender would not by itself give most users meaningful control. Modern systems are complex, change continuously and combine models, data and business rules. More useful transparency is functional: which main parameters affect recommendations, why an item was suggested and which settings a person can alter.

The DSA therefore requires platforms using recommender systems to explain their main parameters in plain language and describe options for influencing them. For very large platforms it also requires at least one recommender option not based on profiling. That does not solve every problem, but it changes the principle: personalisation should not be the only gateway into the information environment.

By early 2026, regulatory scrutiny had also extended to designs that may foster compulsive use. In February 2026 the European Commission announced a preliminary finding concerning TikTok’s addictive design. Because the finding was preliminary, it must not be presented as a final legal determination. It is relevant as evidence that oversight now includes the technical organisation of attention itself.

A practical audit of attention sovereignty

We can audit a platform without guessing at secret intentions. Instead of asking “is the algorithm manipulating me?”, we can ask concrete questions about selection, objectives, transparency and exit.

Twelve useful questions are: (1) who determines the order of content, (2) what main objective the system optimises, (3) which of my signals most affect recommendations, (4) whether I can see why a specific item was recommended, (5) whether I can choose a chronological or other non-profiled feed, (6) whether I can reset or correct my interest profile, (7) how easily I can disable autoplay, notifications and other retention mechanisms, (8) whether I can clearly signal “show me less of this”, (9) whether advertising is clearly distinguishable, (10) whether independent researchers can examine systemic effects, (11) what happens to my reach or access if I reject personalisation, and (12) how easily I can move my information relationships elsewhere.

A good result is not a feed without algorithms. Chronology is itself a selection rule, and wholly manual selection can also create a narrow environment. The stronger standard is: does the user understand the basic logic of selection, have real alternatives, and remain able to alter or leave the system without disproportionate loss of access to people, information or audience?

“Who Controls the Data About Us?” asked who controls data. This article showed how those data, together with ranking, recommendations and interface design, can shape visibility and the use of attention. We now have two layers of digital power: a system may know a great deal about a person and at the same time significantly influence what that person sees next.

Attention, however, is not yet the hardest form of digital dependency. A person can close a feed and still retain the capacity to pay, prove identity or access essential services. The stakes rise when a single digital gateway becomes a condition for normal participation in everyday life.

“Digital Identity and Payments as Control Points” will therefore examine digital identity and payments as control points: interoperability, exclusion, single points of failure and the possibility that a technical or institutional decision changes not merely what someone sees but what they can do at all. This article leaves behind a simpler rule: attention remains part of freedom when people have a real ability to decide who directs it and by what rules.