Algorithmic Recommendations

There is a question that seems simple. Why does YouTube show you exactly this video. Why does TikTok show exactly this clip. Why does Spotify play exactly this song. Most people answer — because they like it. Because the algorithm has studied their tastes. Because the platform wants to make their experience better. That is not the answer. That is what the platforms want you to think. The real answer is different. And it changes everything.

The metric the algorithm optimises for. Not the one they talk about publicly. Every recommendation algorithm optimises for a specific metric. This is a mathematical fact — you cannot optimise for a «good experience», because that is not a number. You need a concrete number. What the platforms say publicly. YouTube — «we optimise for user satisfaction». TikTok — «we show content that interests you». Spotify — «we help you discover music you’ll like». What they actually optimise for. Before 2016, YouTube optimised for the number of clicks. This created a clickbait industry — sensational headlines, deceptive thumbnails. Users complained. The reputation suffered. In 2016, YouTube changed the metric. The new goal — watch time. Total viewing time. Not clicks — minutes. It would seem better. If a person watches for a long time — that means they like it. In practice — watch time is maximised through emotional engagement. Content that keeps you at the screen longer is not necessarily the content that brings benefit or gives pleasure. It is content that provokes strong emotions. Any of them. Admiration, anger, anxiety, fear, indignation. An anxious person who watches videos about threats and disasters for three hours creates more watch time than a calm person who watched a useful lesson and closed the tab. The algorithm does not know that the first is feeling anxiety. It sees only the number. TikTok optimises for completion rate — the percentage of videos watched to the end — and repeat views. This leads the platform to favour short videos with an unexpected ending. The brain wants to close an unfinished gestalt — so it watches to the end. The algorithm knows this. Spotify optimises for listening time and number of sessions. This means the algorithm prefers music that creates a state in which a person keeps listening. Background music. Playlists for work. Relaxing tracks. Not because it is the best music — but because it does not require an active decision to stop.

Internal documents that became public. These are not assumptions. These are documented facts from companies’ internal materials that became public. 2021. Frances Haugen — a former Facebook employee — handed over thousands of internal documents to the US Congress and to journalists. This became known as the Facebook Papers. What the documents showed. Facebook knew that the feed algorithm amplifies polarisation. An internal study in 2018 showed that a change to the algorithm that reduced polarisation also reduced user engagement. The company chose engagement. Facebook knew that content provoking an «angry» reaction spreads five times faster than neutral content. The algorithm gave such posts priority — because they generated more interactions. Interactions are a metric. Facebook knew that Instagram harms the mental health of teenage girls. An internal study showed that 32% of girls who felt bad about their bodies said Instagram made them feel worse. The company concealed this study from the public. 2023. Lawsuits against TikTok in several US states opened up part of the internal documentation. It emerged that TikTok had an internal name for the state the algorithm tried to create in users — «flow state». A state of flow. When a person is fully absorbed and does not notice time. This is not a random side effect. This is the goal.

How the algorithm builds the loop. The mechanics of retention work through a loop that the algorithm builds individually for each user. Step one — probing. A new user is shown diverse content. The algorithm watches the reactions. What is watched to the end. What is lingered on. What is rewatched. What provokes a reaction — a like, a comment, a repost, or simply a long pause before scrolling. Step two — clustering. The algorithm finds which cluster of users you belong to. Not demographically — behaviourally. People who react the same way you do — what did they react to next? What held them longer? Step three — exploitation. The algorithm begins to show content that statistically holds people of your behavioural type the longest. Not what you asked for. Not what you liked. What holds you. Step four — deepening. If you react to a certain topic — the algorithm goes deeper into that topic. Watched one video about diets — you’ll get more about diets. Then about extreme diets. Then about eating disorders. Not because the algorithm wants to harm you — because each next step generates slightly more engagement than the previous one. This is called a rabbit hole. A user comes in to watch a video about healthy eating — and through a series of recommendations ends up on content about anorexia. The algorithm led them there step by step. Each step was small. The total distance — enormous.

Radicalisation through recommendations. A documented effect. This is one of the most studied consequences of algorithmic recommendations. A researcher who in 2019 created hundreds of fictitious accounts on YouTube and systematically observed where recommendations lead from different starting points recorded the following. The algorithm consistently led users from moderate content to more radical content on the same topic. Not instantly. Gradually. Each next clip slightly more intense than the previous. The mechanics are simple. Moderate content about politics competes with thousands of similar videos. Radical content provokes a stronger emotional reaction. A stronger reaction means more watch time. The algorithm does not distinguish moderate from radical. It sees only the viewing time. YouTube disputed the study’s conclusions. It changed the algorithm in 2019, stating that it had reduced recommendations of «borderline» content. Independent researchers recorded a partial improvement. Not a complete one. The same effect is documented in Facebook regarding political information. In WhatsApp regarding the spread of disinformation. In Instagram regarding content connected with appearance and self-esteem. The algorithm has no political position. It does not want to radicalise people. It does not seek to harm. It simply optimises a number. And the number turns out to be higher for content that provokes stronger emotions. And stronger emotions are more often provoked by content that is further from the moderate. This is not malice. This is mathematics with unforeseen consequences that the companies saw — and preferred not to fix, because it reduced engagement.

Personalisation that creates different realities. There is a consequence of algorithmic recommendations that is less obvious but perhaps more significant. Two people who ask the same question on YouTube or TikTok get different answers. Not because one question sounds different. Because the algorithm knows who each of them is. And shows different content. This means that each person’s information environment is unique. Personalised to their past reactions. To their behavioural profile. To what the algorithm considers holding for that particular person. Two people in the same city, of the same age, with similar education — can get fundamentally different answers to one and the same question. Live in different versions of reality. Not know about it. This is called a filter bubble. But since this term appeared in 2011, the technology has become significantly more powerful. The personalisation of 2024 is incomparable with that of 2011. The consequences for society are being recorded by researchers. People who receive different information realities find it harder to reach a common language. Harder to come to agreement. Harder to understand why another person thinks differently — because they literally saw a different world through their screen. The polarisation we observe in politics in many countries has many causes. Algorithmic recommendations are one of them. A documented one.

Part two — how the algorithm works from the inside

Let us continue. How exactly the algorithm decides what to show you next. What data it uses. How it builds a model. And why this model knows more about you than you think.

Two layers that work together. The recommendation algorithm in large platforms is not one system. It is at least two layers that work in sequence. The first layer — candidate generation. From millions of units of content the system selects several hundred that could potentially interest you. Fast. Rough. Based on general patterns. The second layer — ranking. From these several hundred candidates the system chooses the few that it will show you right now. Slow. Precise. Taking into account everything it knows about you. It is precisely the second layer that is the place where the most interesting things happen.

What the ranking model knows about you. The data used for ranking is significantly broader than just viewing history. Interaction history. Not only what you watched — but how. Whether you watched to the end. Where you stopped. Whether you rewatched specific fragments. Paused. Rewound. Each of these actions carries information about how much the content captured attention. Temporal patterns. At what time of day you are active. What content you watch in the morning — and what in the evening. How preferences change on weekends. The platform knows that the same person wants different things at different times of day. And adapts. Device context. From phone or computer. At home or on the move — this is visible from session patterns. Connected to WiFi or mobile internet. This affects recommendations — on a phone on the road short content is recommended, at home on a big screen — long content. Social graph. What people who are similar to you behaviourally watch. Not your friends — behavioural doubles. Users whose interaction history correlates with yours. This is called collaborative filtering. If thousands of people with a similar profile watched video B after video A — you too will be shown B. Content substance. Video transcripts, tags, descriptions, visual content through computer vision. The algorithm understands what the content is about — and matches it with what you reacted to earlier. Real-time signals. What is popular right now. What is gaining views quickly. What is being discussed right now. This is added to the personal data — the platform balances between personal preferences and topicality.

The user model. How the algorithm represents you. All of this together forms what engineers call a user embedding — a vector representation of the user. It is a mathematical object in a multidimensional space. Thousands of numbers that describe you as a pattern of behaviour. Each piece of content is also represented as a vector in the same space. The algorithm’s task is to find content whose vector is close to your vector right now. Right now — the key words. Your vector is not static. It is updated in real time. What you watched five minutes ago affects what you will be shown right now. What you watched half a year ago affects less — but still affects. This means the algorithm works with two versions of you simultaneously. The long-term profile — who you are in general, what your stable interests are. And the short-term context — what state you are in now, what captured you in the last few minutes. This is precisely why the algorithm is able to «catch» you. If you started watching anxious content — the short-term context signals that you are now receptive to the anxious. The algorithm shows more anxious content. You watch. The context intensifies. The loop has closed.

YouTube from the inside. What is known from public sources. In 2016 Google published a research paper on the YouTube recommendation system. This is one of the rare cases when a large platform disclosed details of its architecture. What is written there. The system processes hundreds of millions of videos and billions of users. Every second, 500 hours of new content is uploaded to YouTube. The algorithm’s task is to choose from this volume several dozen videos for a specific user at a specific moment. As a training signal, not only the click on a video is used — but also the viewing time after the click. This is fundamental. A clickbait video gets a click — but does not get viewing time if it deceives expectations. The algorithm penalises this. The system uses data about implicit signals — not only likes and comments, but also behaviour that the user does not perceive as an action. How much time was spent on the video’s page before pressing play. How quickly it was closed after starting. Whether they returned to the previous video. What was not disclosed in this paper — and what became known from other sources. The algorithm has a mechanism that engineers call «exploitation vs exploration». Exploitation — showing what you will definitely react to based on past behaviour. Exploration — occasionally showing something new to broaden the profile and find new interests. The balance between these two modes is one of the key decisions in the design of an algorithm. Too much exploitation — the user gets stuck in a narrow bubble. Too much exploration — engagement is lost. Platforms do not disclose how exactly they balance this. But the effect is visible — the algorithm periodically shows something unexpected. This is not chance. This is exploration.

TikTok. An algorithm that works differently. TikTok differs fundamentally from YouTube and Facebook in one key aspect. YouTube and Facebook build recommendations predominantly on the basis of the social graph and history. Who you are subscribed to. What you watched before. What people similar to you watch. TikTok builds recommendations predominantly on the basis of the reaction to content — independently of social connections. You do not have to subscribe to anyone. The algorithm looks only at how you react to the videos it shows. This means two things. First — TikTok builds an accurate profile significantly faster. YouTube needs weeks to understand a new user. TikTok — a few hours. Because it does not wait for you to subscribe to channels. It simply shows videos and watches the reaction. Second — TikTok is significantly more effective at promoting new content. On YouTube a new creator with no audience is practically invisible. On TikTok every video passes through a testing mechanism — it is shown to a small group, the reaction is measured, and if it is above a threshold — it is shown to a larger group. Organic virality is built into the architecture. This makes TikTok an especially powerful retention tool. And especially hard to study — because every user sees a unique feed that is not determined by public activity. What is known about TikTok’s internal workings is significantly less than about YouTube. The company has not published research papers. Litigation opened up part of the documents. They show that TikTok tracked significantly more data than it disclosed publicly — including clipboard data, network data, detailed behavioural analytics.

Spotify. An algorithm that creates states. Spotify is a different story. Not video — audio. Not active viewing — background listening. This changes exactly what is optimised. Spotify optimises for listening time and session frequency. How much time you listen per session. How often you return. The key discovery Spotify made on the basis of data — different states require different music. And the transition between states is the moment when users most often leave the platform. Therefore Spotify’s algorithm does not simply recommend music you like. It recommends music that holds you in your current state long enough that you do not switch. Hence — the phenomenon of playlists for work, for sleep, for workouts. This is not merely convenient categorisation. It is algorithmically optimised sequences that hold a state long enough that you do not close the app. Spotify also uses audio analysis — the tempo, key, energy of each track — and matches it with the time of day, activity, listening history. This allows it to predict what you need now — not on the basis of what you asked for, but on the basis of patterns the algorithm saw in you and in similar users.

Why algorithms are not disclosed. Three reasons. Companies explain the opacity of algorithms by competitive considerations. This is true — but not the whole truth. Reason one — competition. A recommendation algorithm really is a competitive advantage. To disclose its details is to help competitors. Reason two — manipulation. If content creators know exactly how the algorithm works — they optimise for it. Not for the quality of the content, but for the algorithm’s metrics. This is already happening — but full knowledge of the algorithm would make this even more systematic. Reason three — and this is the one they do not name publicly — liability. If the algorithm is disclosed and documented as optimising for engagement through anxiety and anger — this becomes grounds for legal liability. For psychological harm. For radicalisation. For influence on elections. Opacity is not only a business decision. It is a legal defence.

The AI Act and algorithmic transparency. The European AI Act — a law that took effect in 2024 — contains transparency requirements for recommender systems. Platforms with more than 45 million users in the EU — Very Large Online Platforms — are obliged to provide users with at least one recommendation option that is not based on profiling. That is, a chronological feed or a random selection — without personalisation based on behavioural data. This is the world’s first legislative requirement for recommendation algorithms. How effectively it will be enforced — practice will show. Regulators have also received the right to request data about how algorithms work in order to assess systemic risks. This means that for the first time state bodies can — theoretically — look inside the black box. The practice is only beginning. But the direction is set.

Part three — how this affects one’s picture of the world. And what to do about it

What happens to a person and to society when algorithms decide what we see. And what can be done — really, not in theory.

The picture of the world you did not choose. There is a simple fact that is hard to accept. Most of the information about the world you receive not through your own choice. Through algorithmic choice. What to show. What to push forward. What to hide. What to amplify. This is not news — media have always influenced what people know about the world. Editors chose what to print. Television producers decided what to show. But algorithmic recommendations differ from traditional media in two fundamental aspects. The first — the scale of personalisation. A newspaper showed the same content to millions of readers. The algorithm shows unique content to each of a billion users. There is no common picture that everyone sees. There are a billion individual pictures. The second — the criterion of selection. An editor selected content by criteria that were, at least declaratively, connected with the public interest. The algorithm selects by a criterion that has nothing to do with the public interest — maximum retention of a specific user. These are not the same thing. And the difference has consequences.

What research says about the consequences. Concretely. These are not speculative reflections. These are documented effects that have been measured on real people. Effect one — narrowing of information diversity. A study published in the journal Nature Human Behaviour in 2023 analysed the news consumption of 6,000 Facebook users. Users who saw an algorithmic feed consumed significantly less diverse content than those who saw a chronological feed. The algorithm deepened existing interests — it did not broaden them. This means that over time a person’s information diet narrows. They learn more about what they already know. And less about what they did not know. Effect two — reinforcement of existing beliefs. The algorithm shows content that provokes a reaction. Content that contradicts a person’s beliefs provokes a reaction — but an unpleasant one. Content that confirms beliefs provokes a pleasant reaction. Over time the algorithm shifts towards confirming content — because it generates more positive reactions that correlate with continued use. This is called confirmation bias amplification — the algorithmic amplification of confirmation bias. Beliefs become stronger and less flexible. Not because a person found more evidence — but because the algorithm showed more content that confirmed those beliefs. Effect three — degradation of the capacity for boredom. This sounds strange — but it is a real, documented effect. Boredom is a functionally important state. It is precisely in a state of boredom that the brain switches into default mode network mode — the wandering mind, which processes experience, builds connections, generates ideas. This is necessary for long-term cognitive health. Algorithmic recommendations are designed so that boredom never sets in. The next video begins before the previous one has ended. The feed is infinite. The moment when the brain could switch into wandering mode is filled with the next piece of content. Research shows that the chronic avoidance of boredom through constant stimulation correlates with a decline in the capacity for concentration, with difficulty in independently structuring time, with heightened anxiety when external stimulation is absent. The phone has become a prosthesis that takes on the function of managing one’s state. This changes how people endure the absence of external stimulation. Effect four — influence on the collective perception of reality. This is the largest-scale and hardest-to-measure effect. When a significant part of the population receives information about the world through algorithms that optimise for engagement — the collective picture of reality shifts towards emotionally charged, conflict-laden, anxious content. Not because the world has become this way. But because such content holds better. Researchers have recorded this shift in several countries through analysis of the content that gained the greatest reach on social networks compared with traditional media. Algorithmically amplified content is systematically shifted towards negativity, conflict and anxiety. The world we live in is becoming less frightening by many objective measures. Violence is declining. Poverty is declining. Life expectancy is rising. But the perceived world through the screen of an algorithmic feed feels ever more dangerous and conflict-ridden. This is the gap between reality and its algorithmic reflection. And it has consequences for political decisions, for personal well-being, for social trust.

What platforms do in response to criticism. This is important to understand — because platforms do react. But the reaction is often the management of reputational risk, not a real change. YouTube announced in 2019 a reduction in recommendations of «borderline content». Independent researchers recorded a partial improvement in the US — and practically no change in other countries. The algorithm was changed for English-language content, which was under the greatest media attention. For the rest of the world — practically without change. Facebook after the Facebook Papers launched several initiatives to reduce «toxic content». Internal documents showed that these initiatives were consistently cancelled or weakened when engagement metrics began to decline. TikTok after the lawsuits in the US introduced Screen Time Management — tools for limiting usage time. They exist. They are not enabled by default. The user has to actively configure them. Most do not. The pattern is the same. A public reaction to criticism. Minimal changes that do not touch the base optimisation metric. Because the base metric is the business model. And the business model does not change under the pressure of criticism. It changes under the pressure of regulators or under the pressure of a change in market conditions.

What really works. Not how to avoid the algorithm — how to live with it consciously. To fully avoid algorithmic recommendations in the modern digital world is unrealistic. That is not the goal. The goal is conscious interaction. To understand what is happening. To make decisions deliberately where everything used to happen automatically. First — separating passive and active consumption. There is a difference between what you look for yourself and what you are shown. When you type a query into search — that is an active choice. When you watch the next video from an automatic recommendation — that is passive consumption. A conscious practice is to periodically stop and ask. Am I watching this now because I chose to — or because the next video started automatically? This one question changes one’s relationship to content consumption. Second — deliberate breaks. The algorithm is optimised for continuity. The next video, the next post, the next track. A deliberate break is a resistance to this mechanic. Not to switch off the phone forever. Just to close the app after a specific video. Before the next one began. This sounds trivial. In practice it is harder than it seems, precisely because the interface is designed against it. Third — deliberately diversifying sources of information. The algorithm narrows the information diet. The counter is deliberate broadening. To read what you do not agree with. To subscribe to sources that look at the world differently. This does not mean accepting others’ views. It means having a real picture of what different people think — rather than an algorithmically filtered version. Fourth — a chronological feed where it is available. Under the AI Act, large platforms in the EU are obliged to provide a non-recommendation feed option. This is already available in a number of services. Use it — at least periodically. Look at the difference between what the algorithm shows and what the people you follow publish in chronological order. The difference is often striking. Fifth — regular clearing of history. The algorithm builds a profile on the basis of past behaviour. Clearing the viewing history, the search history, the interaction data partially resets this profile. Not fully. But substantially. On YouTube — account settings, watch history, delete all history. On TikTok — settings, privacy, clear cache and history. On Spotify — fewer options, but you can hide listened tracks from the history.

Finally. About the responsibility that is blurred. There is a question that rarely comes up in this conversation. Who bears responsibility for the consequences of algorithmic recommendations? The platforms say — we provide a tool. What people watch is their choice. We do not create content. We only recommend. But the algorithm is not a neutral tool. It is a system that actively shapes the information environment of billions of people. That makes decisions about what to amplify and what to suppress. That is optimised for a specific metric — and this metric does not coincide with the interests of users or of society. This is a new type of power. Without a name. Without regulation that matches its scale. Without transparency that would allow its consequences to be assessed. The AI Act is a first step. But a first step on a very long path. While this path is being travelled — the only defence that works right now is understanding. Knowledge of how the system works. Conscious interaction with it. The algorithm makes thousands of decisions about you every day. Most of them are unnoticed — but now you know that they are happening. That is already a different position.