What You Might Also Like

Open almost any app that offers something to watch, read, listen to or buy, and the first thing you see is a choice someone else has made for you. Netflix shows rows of films “because you watched” something else. YouTube offers the next video before the current one has ended. Spotify assembles a playlist each Monday. Amazon suggests what customers like you also bought. TikTok shows a single full-screen video and, when you swipe, another, chosen from billions, without your having asked for anything at all.

These choices are made by recommendation systems: software that predicts what each person is most likely to want, or at least to watch, click or buy, next. They are among the most consequential technologies of the century. Netflix engineers wrote in 2015 that recommendations influenced about 80 percent of the hours its members streamed.¹ YouTube said in 2018 that more than 70 percent of the time people spent watching it came from its recommendations.² For billions of people, the recommender has become the main way they meet films, songs, ideas and news.

This essay follows how recommendation systems came to occupy that position. Its story runs from a research project for filtering email to a feed that requires nothing of the user but attention. Its deeper subject is choice: what it means to choose when the options are chosen for you, what a machine should be asked to optimise, and how a prediction of what we want can change what we want. To understand it, we need to start with the problem recommendation was invented to solve.

Before the algorithm: critics, charts and clerks

People have always relied on others to choose. Before recommendation systems, the job was done by critics, editors, librarians, disc jockeys, bestseller lists, friends and shop assistants. A good bookseller knew the stock and the customer and could say, if you liked this, try that. Critics argued about what was worth attention. Charts and bestseller lists showed what others were buying, and created a kind of social proof. Each of these was limited: a critic had particular tastes, a bookseller knew only so many customers, a chart showed only what was already popular.

The problem these intermediaries solved grew acute with digital abundance. A video shop might stock a few thousand films; Netflix’s catalogue, a music streaming service or YouTube offers millions of items, far more than anyone could browse. Herbert Simon had observed in 1971 that a wealth of information creates a poverty of attention. Recommendation systems are the most ambitious attempt to allocate that attention automatically.

Collaborative filtering: other people as a guide

In 1992 researchers at the Xerox Palo Alto Research Center, David Goldberg, David Nichols, Brian Oki and Douglas Terry, described a system called Tapestry for filtering the growing volume of email and electronic news. Tapestry let users annotate messages, marking them as interesting or not, and let others filter their own mail using those annotations, such as showing only messages a particular colleague had found useful. The authors called this collaborative filtering: people helping each other filter by recording their reactions.³

The idea was soon automated. In 1994 a team at the University of Minnesota and MIT, led by Paul Resnick and John Riedl, built GroupLens, which predicted how much a user would like a Usenet news article by finding other users who had rated past articles similarly and averaging their ratings. In 1995 Upendra Shardanand and Pattie Maes at the MIT Media Lab built Ringo, which recommended music the same way, and which became a company, Firefly, bought by Microsoft in 1998. The principle was simple and powerful: people who agreed in the past are likely to agree in the future. You did not need to know anything about the music or the news, only how people reacted to it.

Collaborative filtering had a distinctive philosophy. It did not try to understand content or to judge quality. It treated taste as a pattern of agreement among people, discoverable from their behaviour. A recommendation was not a claim that something was good, but a prediction that someone like you would like it.

Amazon: customers who bought this

Amazon applied collaborative filtering to commerce. Matching each customer to similar customers, as GroupLens did, was too slow for millions of customers and products. Amazon’s engineers, Greg Linden, Brent Smith and Jeremy York, developed an approach they called item-to-item collaborative filtering: rather than finding similar customers, the system computed in advance which products were bought together unusually often, and then, when a customer viewed or bought a product, showed the related items. The method scaled independently of the number of customers and could produce recommendations instantly. They described it in a 2003 paper that the journal’s editors later chose as the one from its first twenty years that had best stood the test of time.⁴

“Customers who bought this item also bought” became one of the best-known features on the web. Recommendations appeared on product pages, in emails and on the home page. They expanded what customers found beyond what they searched for, and they turned the behaviour of millions of shoppers into a guide for each one.

The Netflix Prize

In October 2006 Netflix, then a DVD-by-mail company, made an unusual offer. It released a dataset of about 100 million ratings that its customers had given to films, with names removed, and offered $1 million to any team that could improve the accuracy of its own recommendation system, Cinematch, by 10 percent.⁵ Thousands of teams entered. Progress was rapid at first and then slow; the last percentage points took years. In September 2009 the prize went to BellKor’s Pragmatic Chaos, a combination of teams that had merged their methods, which submitted its winning entry minutes before a rival team with an equally good result.

The competition was a landmark for the field. It popularised techniques such as matrix factorisation, which represent users and films as points in a space of hidden factors, and showed the power of combining many models. But its outcome for Netflix was mixed. Netflix used parts of the winning methods and later explained that the full ensemble was too complex to deploy for the gain it offered. More importantly, the business had changed. By 2009 Netflix was moving from mailing DVDs to streaming, and with streaming it could observe what people actually watched, how long they watched and when they stopped. What people did turned out to be a better guide than how they rated. A star rating recorded what people thought they should like; viewing recorded what they chose at the end of a long day.

The competition also exposed a danger. In 2008 two researchers at the University of Texas, Arvind Narayanan and Vitaly Shmatikov, showed that individuals in the supposedly anonymous dataset could be identified by comparing their ratings with public reviews on the Internet Movie Database, and suggested that ratings could hint at sensitive matters such as political views. A planned second competition drew a lawsuit, which argued among other things that viewing histories could reveal a person’s sexual orientation, and questions from the US Federal Trade Commission, and Netflix cancelled it in March 2010.⁵ Data about taste turned out to be data about people.

In 2015 two Netflix executives, Carlos Gomez-Uribe and Neil Hunt, described how recommendations shaped the streaming service. They wrote that recommendations influenced about 80 percent of hours streamed, with the rest coming from search, and estimated that personalisation and recommendations saved the company more than $1 billion a year by reducing the number of people who cancelled.¹ Netflix’s rows of personalised titles, and even the artwork chosen for each title for each viewer, were designed to help members find something to watch within a minute or two, before they gave up.

YouTube: from clicks to watch time

YouTube faced the problem at a different scale. By the early 2010s hundreds of hours of video were uploaded every minute. YouTube initially ranked and recommended videos largely by clicks and views. That rewarded misleading titles and thumbnails: videos that people clicked and abandoned after a few seconds. In 2012 YouTube changed its systems to favour watch time, the total time people spent watching, on the reasoning that if viewers kept watching, they were getting value.⁶ Clickbait lost ground; viewing sessions grew longer. Creators adapted, some by making videos more engaging, some by making them longer, and some by making them more provocative, because outrage and suspense held attention.

In 2016 Google engineers described YouTube’s recommendation system as two deep neural networks: one to select a few hundred candidates from millions of videos, and another to rank them, trained to predict expected watch time.⁷ At the Consumer Electronics Show in January 2018, YouTube’s chief product officer, Neal Mohan, said that more than 70 percent of the time people spent watching YouTube was driven by its recommendations.²

As YouTube’s recommendations became central to how people found video, critics argued that they led viewers towards ever more extreme content, a “rabbit hole” of conspiracy theories and radical politics. In 2019 YouTube announced it would reduce recommendations of “borderline” content, videos that came close to violating its rules without crossing the line, and of harmful misinformation. By the end of the year it reported a 70 percent average drop in US watch time of such content coming from recommendations to non-subscribers.⁸

The research that followed complicated the rabbit-hole story. A 2020 audit by Manoel Horta Ribeiro and colleagues found that users who commented on extreme channels had often commented on milder ones before, consistent with a pathway of radicalisation, though it could not show that recommendations caused the movement. A 2021 study by Homa Hosseinmardi and colleagues, using the browsing histories of a large panel of American users, found that consumption of radical content on YouTube closely matched people’s preferences elsewhere on the web, and concluded that it was driven largely by users’ own tastes, without ruling out a role for the algorithm.⁹ The consensus that emerged was that recommendation systems can amplify what people already seek, and can matter a great deal for small groups, but are not the main cause of radicalisation for most users. The evidence for the period before 2019, when YouTube changed its systems, is harder to establish.

TikTok: the feed without friends

The most radical change came from China. In September 2016 ByteDance, a Beijing company whose news app Toutiao was built around algorithmic recommendations, launched a short-video app called Douyin. Its international version, TikTok, followed in 2017 and merged with the lip-syncing app Musical.ly in 2018.

TikTok’s main screen, the “For You” feed, showed a single full-screen video at a time, chosen by the algorithm from all the videos on the service, not just from accounts the user followed. Every swipe, pause, replay and share was a signal. Because each video was short, the system collected dozens of signals in minutes and could learn a new user’s interests almost immediately. A newcomer with no friends on the app saw a feed tuned to them within a session. The social graph, the network of friends and followers on which Facebook, Instagram and Twitter had been built, became optional. The essay on social media describes this as the end of the friend list.

The model was extraordinarily successful, and rivals copied it. Instagram launched Reels in 2020 and YouTube launched Shorts. In July 2022 Instagram tested a full-screen, video-first feed filled with recommended content from accounts users did not follow. Users and prominent creators protested, with a widely shared petition asking the company to “make Instagram Instagram again,” and Instagram rolled back some of the changes while continuing to increase recommended content more gradually. The direction was clear across the industry: feeds would be chosen less by whom people followed and more by what machines predicted would hold their attention.

TikTok’s success also made its recommendation system a matter of national security. US officials argued that a Chinese company’s control over what a large share of Americans saw created risks of data access and influence. In April 2024 Congress passed a law requiring ByteDance to sell TikTok’s US business or see the app banned, and in January 2025 the Supreme Court upheld it. After repeated extensions, a joint venture majority-owned by American investors, including Oracle, took over TikTok’s US operations in January 2026, with a requirement to retrain the recommendation algorithm on US user data and keep it free from outside manipulation.¹⁰ The recommendation algorithm had become something governments negotiated over.

Music and the shape of taste

Recommendation changed music as well. Spotify launched Discover Weekly in 2015, a personalised playlist of thirty songs delivered every Monday, combining collaborative filtering with analysis of audio and of what people wrote about music online. Playlists, many chosen by algorithm, became a main route by which listeners found songs, and the economics of music adjusted: artists aimed to be added to playlists, songs tended to become shorter, and the first seconds of a track mattered more because listeners skipped quickly.

The filter bubble debate

In 2011 the activist Eli Pariser argued in The Filter Bubble that personalisation would trap people in bubbles of information that confirmed their existing views, invisible to them and chosen by algorithms. The idea was widely adopted. The evidence turned out to be more complicated. A 2015 study by Facebook researchers found that users’ own choices of whom to befriend and what to click limited their exposure to opposing views more than the algorithm did. In 2023 a set of studies by academic researchers working with Meta on the 2020 US election, published in Science and Nature, found that the median Facebook user saw a majority of political content from like-minded sources; but when tens of thousands of users were switched to a chronological feed for three months, they spent less time on the platforms and saw more content from untrustworthy sources, and their political attitudes and polarisation did not change measurably.¹¹ Algorithmic feeds shape what people see; whether they change what people think, at least over short periods, is much less clear.

Regulating the feed

Regulators have increasingly treated recommendation systems as something to govern. The European Union’s Digital Services Act, whose obligations for the largest platforms began to apply in 2023, requires very large online platforms to explain the main parameters of their recommender systems and to offer at least one option that is not based on profiling, such as a chronological feed.¹² It also requires them to assess and mitigate systemic risks, including risks to public debate and to minors, that their systems may create.

In the United States, the question of responsibility reached the Supreme Court in Gonzalez v. Google (2023), in which the family of a victim of a terrorist attack argued that YouTube’s recommendation of terrorist videos was not protected by the law that shields platforms from liability for content posted by users. The Court declined to decide the question, sending the case back in light of a related ruling.¹³ Whether recommending is a form of publishing, with the responsibilities that implies, remains unsettled in American law.

What recommendation systems are today

Today recommendation systems are everywhere: on streaming services, shops, news sites, social networks, dating apps, job sites and app stores. Most combine several methods: collaborative filtering based on behaviour, analysis of content, and deep neural networks trained to predict particular actions, such as watching, clicking, buying or sharing. They are optimised for measurable objectives chosen by the companies that run them, usually some combination of engagement and satisfaction measured through surveys and behaviour, and constrained by policies on what not to amplify. They learn continuously from what people do, which means that what they show shapes the data they learn from. Generative AI is beginning to add a new layer, describing and even creating the items it recommends.

How recommendation systems changed the way we choose

Several shifts follow from the rise of recommendation systems.

From searching to being shown

Search requires a person to know what they want. Recommendation requires only that they show up. The default way of meeting new things moved from asking to receiving.

From stated to revealed preference

Early systems asked people to rate things. Later systems watched what they did. What people choose in the moment came to count for more than what they say they value, and the systems became very good at serving the first.

From the critic to the crowd and the model

Recommendation once came from people with known tastes and reputations. Now it comes from patterns in the behaviour of millions, distilled by a model no one fully understands. The authority of taste moved from individuals to aggregates.

From the social graph to the interest graph

Feeds used to show what friends and followed accounts posted. TikTok showed that a feed could be built entirely from predictions of interest. Social connections became one signal among many.

From a fixed catalogue to a loop

A shop’s shelves did not change because of what one customer looked at. A recommender’s offerings change with every interaction, and the content creators who supply it change what they make in response to what is recommended. Taste, catalogue and system shape each other continually.

What the history suggests

Recommendation systems repeat patterns found across this site. Each new measure, ratings, clicks, watch time, was adopted because it was better than the last, and each became a target that creators and systems optimised directly, diverging in new ways from what people valued. Users decided what the products were for: Netflix members revealed through their viewing what their ratings concealed. Incumbents absorbed the challenger’s best idea when Instagram and YouTube copied TikTok’s feed.

But recommendation also shows something more particular. It is the clearest case of a technology whose central design choice is a choice of objective. Every recommender is told to maximise something. The history is a sequence of discoveries about what happens when that something is ratings, or clicks, or minutes, and a growing awareness that none of them is quite the same as what people would want for themselves on reflection. That gap is where the philosophy begins.

Understanding the philosophy

The rest of this essay examines the ideas beneath recommendation systems, one at a time. Each is explained in plain terms, connected to the tradition of thought that has examined it, traced in the history above, tested against its strongest objection and related to how people and organisations decide today. The connections to philosophers, economists and psychologists are interpretations offered to make the ideas clearer. The engineers who built recommenders were solving practical problems and did not, as a rule, set out to answer these thinkers. The essay on social media examines recognition, mimetic desire, social comparison and the economics of attention; this essay concentrates on the act of choosing and the system that chooses.

1. What is taste?

The idea in plain terms

A recommender predicts what you will like. But what is liking? Is taste a fixed fact about a person, waiting to be discovered, or is it learned, social and changeable? Is there such a thing as good taste, or only different tastes? A recommender has to take a position on these questions, usually without saying so.

The tradition

In his essay “Of the Standard of Taste” (1757), the Scottish philosopher David Hume faced the puzzle directly. Everyone agrees that tastes differ, and the proverb says there is no disputing them. Yet everyone also agrees that some judgements are absurd: no one would rank a minor poet with Milton. Hume’s answer was that there is a standard of taste, found in the joint verdict of “true judges,” critics with delicacy of perception, practice, the habit of comparison and freedom from prejudice. Taste can be educated, and the judgements of experienced critics converge over time.

Immanuel Kant, in the Critique of Judgement (1790), argued that when we call something beautiful we speak as if everyone ought to agree, even though no rule can prove it. Judgements of taste, for Kant, claim a kind of universal validity, unlike mere statements of what we find agreeable, such as a preference for one wine over another.

The French sociologist Pierre Bourdieu, in Distinction (1979), drew on large surveys of French tastes to argue that taste is socially produced. Preferences in music, food, art and furniture correlated closely with education and class, and people used taste to mark their position and distinguish themselves from others. What seemed natural personal preference was, in his account, learned through upbringing and used in social competition.

Where it shows up

Collaborative filtering takes an implicit position close to Bourdieu’s and far from Hume’s. It assumes that taste is a pattern shared among similar people, discoverable from behaviour, and it makes no judgement of quality. A recommender does not ask whether a film is good, only whether people like you watched it. Netflix’s move from star ratings to viewing behaviour pushed further: it stopped asking people what they judged worthwhile and watched what they chose. The “true judges” of Hume’s standard were replaced by the aggregate behaviour of millions.

The strongest objection

The objection is that recommenders do not need a theory of taste, because they serve people what people demonstrably choose, and that the alternative, experts deciding what others should like, is elitist. Bourdieu’s analysis itself suggests that expert taste is often a mark of class. This has force. Recommenders have opened up many niches that critics ignored and given audiences to work that gatekeepers dismissed. But serving only what people choose in the moment treats taste as fixed, when Hume and Bourdieu agreed that it is formed by exposure. A system that shows people only what they already like gives taste no room to grow.

What it means today

The question for anyone designing or relying on a recommender is what kind of taste it is serving: the taste people have, or the taste they could develop. For organisations, the same applies to recommendations about strategy, hiring or investment drawn from patterns of past behaviour: they will reproduce existing preferences well and suggest new ones poorly. Some exposure to what experts, or simply other kinds of people, value is worth building in deliberately.

2. Nudges and the architecture of choice

The idea in plain terms

Every way of presenting options influences what people choose. The first item on a list is chosen more often; the default is rarely changed; what is easy to reach is reached. There is no neutral way to arrange choices. A recommender is a choice architect of extraordinary power, arranging the options for billions of people individually.

The tradition

In Nudge (2008), the economist Richard Thaler and the legal scholar Cass Sunstein argued that because every presentation of choices influences behaviour, those who design them, whom they called choice architects, cannot avoid influencing people, and should do so in ways that help people make choices they would endorse, while keeping all options open. They called this libertarian paternalism. Their examples included placing healthy food at eye level in cafeterias and making enrolment in pension schemes the default.

Critics raised two concerns. One was that nudges can be used for the architect’s benefit rather than the chooser’s, as in designs that exploit inattention to sell more. The philosophers Daniel Susser, Beate Roessler and Helen Nissenbaum argued in 2019 that online manipulation, hidden influence that exploits people’s decision-making vulnerabilities, threatens autonomy because it bypasses their capacity to reason about what they want. The other concern was that even well-meant nudges treat people as objects to be steered rather than agents to be persuaded.

Where it shows up

Recommendation systems are choice architecture at scale. Netflix’s rows, the artwork chosen for each viewer, YouTube’s autoplay, which starts the next video unless the viewer intervenes, and TikTok’s single full-screen video all shape choice by what they put first and what they make effortless. Autoplay turns the default from stopping to continuing. The infinite feed removes natural stopping points. These designs follow Thaler and Sunstein’s insight that defaults and ease determine behaviour, but they apply it to the platform’s objectives, such as time spent, not necessarily to the choices users would endorse. The EU’s requirement of a non-profiled feed is itself a nudge debate: offering an alternative matters little if the profiled feed remains the default.

The strongest objection

The objection is that people use these products voluntarily and could stop at any time; calling recommendations manipulation insults their agency. Recommendations also help: they save time and surface things people are glad to find. And all media, from newspapers to television schedules, have always arranged choices to hold attention. These points are fair. The difference with recommenders is precision and scale: each person’s choices are arranged individually, using detailed knowledge of their behaviour, to maximise an objective the person did not choose and may not know.

What it means today

For those designing products, the useful test is the one Thaler and Sunstein proposed: would people, on reflection, endorse the choices the design steers them towards? For users, noticing defaults, such as autoplay and infinite scroll, and changing them where possible restores some control. For organisations, every dashboard, form and default in internal systems is choice architecture too, and the same test applies.

3. Wanting what we want to want

The idea in plain terms

People often want things they wish they did not want. Someone may want to keep watching videos at midnight while also wanting to be someone who goes to sleep. Which desire is really theirs? A recommender learns from what people do, which reflects their immediate desires. It cannot easily see the desires they have about their desires, which may matter more to who they are.

The tradition

In 1971 the American philosopher Harry Frankfurt published “Freedom of the Will and the Concept of a Person,” in which he distinguished first-order desires, desires for things, from second-order desires, desires about which first-order desires to have and act on. A person who wants a cigarette but wants not to want it has a conflict between orders. Frankfurt argued that what makes someone a person, rather than a creature driven by impulse, is the capacity to have second-order desires and to care which desires move them. Freedom of the will, on his account, is the ability to act on the desires one endorses.

John Stuart Mill, in On Liberty (1859), defended individual freedom partly because people can develop themselves through choice and experiment, which he called “experiments of living.” For Mill, the value of choosing lies not only in getting what one wants but in exercising and developing one’s faculties. Behavioural economists have documented the practical form of Frankfurt’s distinction: people discount the future steeply and inconsistently, preferring a small reward now to a larger one later even when they would choose the opposite in advance.

Where it shows up

Recommenders trained on behaviour serve first-order desires very efficiently. Watch time, clicks and swipes record what people do in the moment, including at midnight. Netflix’s discovery that viewing was a better predictor than ratings was, in Frankfurt’s terms, a discovery that first-order behaviour predicted choices better than second-order judgement. That made recommendations more accurate and less aligned with what people might endorse on reflection. Some platforms have added tools that address second-order desires, such as reminders to take a break and controls on usage time, and some ask users in surveys how satisfied they were with what they watched, which brings reflection back into the signal.

The strongest objection

The objection is that privileging second-order desires is paternalistic. Who is to say that the person who wants to watch more videos is less real than the one who wishes they did not? People’s reflective judgements can be pretentious or guilty; their behaviour shows what they actually value. A system that served only reflective preferences would be a scold. This is a serious point; there is no simple way to decide which desires count. But it does not follow that a system should ignore second-order desires altogether, especially when it is optimised to exploit the gap between them.

What it means today

For individuals, the most effective defence is to act at the second-order level in advance: setting limits, turning off autoplay, choosing what to watch before opening an app. For designers, giving users ways to tell the system what they want to want, and taking those signals seriously, is both more respectful and, over time, more likely to keep users who do not regret their time. For leaders, the same distinction applies to organisations: what people do under pressure may not be what they would choose on reflection, and systems should help them act on the latter.

4. Exploration, exploitation and serendipity

The idea in plain terms

A recommender that shows only what it knows you like will never discover what else you might like. To learn, it must sometimes show you something uncertain. That trade-off, between exploiting what is known and exploring what is not, runs through every recommender. It is also the trade-off that decides whether you ever stumble on something unexpected.

The tradition

Mathematicians studied this trade-off as the multi-armed bandit problem, named after a gambler facing a row of slot machines with unknown payouts. Each pull either exploits the best-known machine or explores another. The statistician Herbert Robbins formalised it in 1952, and in 1979 John Gittins found an optimal solution for an important class of cases. The general lesson is that some exploration is always worth its cost when the future matters, and that systems that only exploit get stuck.

The word serendipity was coined by the English writer Horace Walpole in 1754, after a Persian fairy tale, The Three Princes of Serendip, whose heroes were always making discoveries, by accident and sagacity, of things they were not looking for. The sociologist Robert K. Merton studied the role of serendipity in science and showed that many discoveries came from noticing unexpected findings. Serendipity is not pure chance: it requires both an encounter with the unexpected and the readiness to recognise its value.

Where it shows up

Recommenders deliberately explore. TikTok shows new users a range of videos to learn their interests quickly and continues to insert some unfamiliar content. Spotify’s Discover Weekly was built for exploration, recommending songs the listener had not heard. But the commercial pressure is usually towards exploitation: showing what is most likely to be watched now. A system optimised for immediate engagement explores only as much as it needs to learn, not as much as would serve the user’s growth. And people who rely entirely on recommendations lose the kind of serendipity that came from browsing shelves, reading a whole newspaper or hearing a friend’s odd suggestion.

The strongest objection

The objection is that recommenders produce more serendipity, not less. They regularly surface things people would never have found: a song from another country, a documentary on an obscure subject, a creator with a few hundred followers. TikTok in particular is known for showing users content far outside their networks. This is true, and recommenders can be engines of discovery. The question is what kind of discovery: the system explores within the space of what is likely to engage, and so its surprises tend to be surprises of the same kind.

What it means today

For individuals, it is worth seeking out sources of the unexpected that are not tuned to engagement: other people’s recommendations, libraries, curated publications, the parts of a catalogue no algorithm shows first. For product designers, explicitly valuing exploration and diversity in the objective, not only as a means to learn, can produce systems people trust more. For organisations, the bandit problem applies to strategy: a company that only exploits its current business will not discover its next one.

5. Predictions that make themselves true

The idea in plain terms

A weather forecast does not change the weather. A recommendation does change what people watch. If a system predicts you will like a video and shows it to you, you are more likely to watch it, which the system then takes as evidence that its prediction was right. Over time, a recommender can create the tastes it claims to discover, and the content creators who supply it make more of what it recommends.

The tradition

In 1948 the sociologist Robert K. Merton described the self-fulfilling prophecy: a false definition of a situation that evokes behaviour which makes the originally false conception come true. His example was a sound bank that collapses because a rumour that it is insolvent leads depositors to withdraw their money. The philosopher Ian Hacking described looping effects in the classification of people, in which the people classified change in response to the classification. The sociologist Donald MacKenzie, studying financial economics, argued in An Engine, Not a Camera (2006) that economic models do not merely describe markets but help create them, as traders acted on the models’ prices.

Where it shows up

Recommendation systems are self-reinforcing in exactly these ways. Popular items are recommended more, which makes them more popular. Content that the system recommends gets more views, which creators notice, so they make more of it, which the system then recommends. YouTube’s move to watch time changed the length and style of videos; Spotify’s playlists changed the structure of songs. Researchers in the field call the technical form of this problem feedback loops or algorithmic confounding: data generated under one recommender is used to train the next, so the system learns from the world it created. The site’s idea When a Measure Becomes a Target describes the related problem from the side of those being measured.

The strongest objection

The objection is that recommendations only accelerate what would have happened anyway. Popular things have always become more popular, through charts, word of mouth and advertising. Creators have always adapted to their medium, as pop songs fit the three minutes of a single. Recommenders are not uniquely self-fulfilling. This is partly right. What is distinctive is the speed and the closure of the loop: a recommender sees the response to its own recommendations instantly and adjusts, for billions of people, without anyone deciding the direction.

What it means today

For anyone using recommendation data to understand what people want, the data reflects what was shown as much as what was wanted; experiments and deliberate exploration are needed to see past it. For creators and businesses that depend on recommenders, the system’s preferences become a market force, and adapting to them is rational, but it ties their work to an objective they do not control. For leaders, any metric that drives what an organisation produces will, over time, reshape both the product and the people producing it.

6. The public sphere and the personal feed

The idea in plain terms

Democracies depend on citizens sharing some common information and encountering views they disagree with. Newspapers and broadcasters, whatever their faults, offered everyone roughly the same front page. Personalised feeds give everyone a different one. The worry is that a society in which no two people see the same news cannot deliberate together.

The tradition

The German philosopher Jürgen Habermas, in The Structural Transformation of the Public Sphere (1962), described the rise in eighteenth-century Europe of a public sphere, coffee houses, salons and newspapers where private individuals met to discuss public matters, and argued that it was later degraded by commercial mass media. A healthy public sphere, for Habermas, requires open access, rational argument and attention to the common good.

In Republic.com (2001) Cass Sunstein worried that the internet would let people construct what the technologist Nicholas Negroponte had called the “Daily Me,” a news diet tailored to their preferences, reducing shared experiences and encounters with opposing views, and pushing groups towards more extreme positions. Pariser’s filter bubble extended the worry to algorithmic personalisation, which, unlike the Daily Me, people did not choose consciously.

Where it shows up

The history above shows the worry and the evidence pulling in different directions. Recommenders do personalise what people see, and in the Meta studies the median Facebook user saw mostly politically like-minded content. But users’ own choices often narrow their exposure more than algorithms do, and replacing the algorithmic feed with a chronological one for three months did not measurably reduce polarisation. Studies of YouTube found that consumption of extreme content largely reflected users’ prior preferences. The public sphere has fragmented, but recommendation algorithms are one cause among many, alongside partisan media, social sorting and political choices.

The strongest objection

The objection is that the public sphere of the broadcast era was never as shared or as open as nostalgia suggests. Mass media excluded many voices, and the “common” front page reflected the views of a narrow class of editors. Recommenders have let minorities, dissidents and niche communities find audiences. Moreover, people have always chosen partisan newspapers. These points are important. The answer is not to restore a lost consensus but to ask whether the systems that now mediate public information are accountable for their effects on it.

What it means today

For citizens, it is worth deliberately reading sources one would not be shown, and noticing that the feed is not the news. For platforms, the DSA’s requirement to assess systemic risks to public debate is a recognition that recommenders are now part of democratic infrastructure. For organisations, the internal version of the problem is real: teams that receive personalised dashboards and feeds may lose the shared picture needed to coordinate.

7. Optimising the wrong thing

The idea in plain terms

A recommender is told to maximise a number: clicks, watch time, sessions, purchases. It will do so with relentless efficiency, including in ways its designers did not intend. If the number is not exactly what people value, the system will drift towards giving people what increases the number rather than what is good for them. Recommenders are the most widely deployed example of a problem that applies to all powerful optimising systems.

The tradition

Aristotle distinguished in the Nicomachean Ethics between pleasure and eudaimonia, usually translated as flourishing or living well. Pleasure accompanies good activities, but a life aimed at pleasure alone is not a good life; flourishing consists in the exercise of virtue and reason over a complete life. John Stuart Mill, though a utilitarian, distinguished higher and lower pleasures and wrote that it is better to be a human being dissatisfied than a pig satisfied.

The computer scientist Stuart Russell, in Human Compatible (2019), used social media recommendation algorithms as an example of the danger of giving machines fixed objectives. A system told to maximise clicks, he argued, can learn to change people so that they become more predictable and easier to engage, rather than serving what they actually want. His proposed remedy was for machines to be uncertain about human preferences and to learn them, deferring to people, rather than pursuing a fixed objective. The essay on large language models discusses the related problem of systems trained to please.

Where it shows up

The history of recommendation is a sequence of objectives, each adopted to fix the last. Netflix moved from predicting ratings to predicting viewing. YouTube moved from clicks to watch time, then added surveys of satisfaction and limits on borderline content. Each step brought the measure closer to something valuable and each left a gap. Watch time rewarded videos that held attention, including through outrage; engagement on social feeds rewarded content that provoked reactions. Platforms’ own research, as revealed in leaks and investigations, at times found that what maximised engagement was not what users said they valued.

The strongest objection

The objection is that platforms know this and do not optimise for one crude number; they combine many signals, including surveys of satisfaction and long-term retention, because users who feel manipulated leave. Aligning with users’ long-term interest is in the platform’s commercial interest too. There is truth in this, and the evolution of objectives shows companies trying to close the gap. But commercial and user interests align only partly, and the measures of long-term well-being are much weaker than the measures of engagement, so the optimisation still leans towards what can be counted.

What it means today

For anyone building systems that optimise, the most important decision is the objective, and it deserves as much scrutiny as the algorithm. For leaders, every target set for a team is an objective for an optimiser, and teams, like algorithms, will find ways to hit it that were not intended. Asking regularly what the measure misses, and supplementing it with judgements of whether people are actually better off, is the human equivalent of Russell’s uncertainty about objectives.

8. Who is responsible for what the feed shows?

The idea in plain terms

When a recommender suggests a harmful video to a vulnerable teenager, or amplifies a lie to millions, who is responsible? The person who posted it, the person who watched it, or the company whose system chose to show it? A recommender is not a passive host of other people’s content, but it is not an editor in the traditional sense either. Its decisions are made by a model trained on behaviour, at a scale no human could supervise.

The tradition

Philosophers have long debated responsibility for outcomes produced through complex systems. The German philosopher Hans Jonas argued in The Imperative of Responsibility (1979) that modern technology, by extending the reach and consequences of human action far beyond what traditional ethics imagined, demands a new ethics of responsibility for long-term and distant effects. The philosopher Dennis Thompson described the “problem of many hands”: when outcomes result from the contributions of many people, it becomes hard to hold anyone responsible, and responsibility can vanish.

In law, the distinction between publishers, who are responsible for what they print, and distributors or carriers, who generally are not, shaped the rules for online platforms. The US Communications Decency Act of 1996, in Section 230, protected platforms from being treated as publishers of content provided by users. Whether recommending content is a form of publishing was the question the Supreme Court declined to answer in Gonzalez v. Google.

Where it shows up

Recommendation sits uneasily between the categories. YouTube’s decision to reduce borderline content was an editorial choice about amplification, distinct from its decisions about removal. The EU’s Digital Services Act treats very large platforms as responsible for assessing and mitigating the systemic risks their recommenders create, without making them liable for each item. The US forced change in TikTok’s ownership partly because of who would control its recommendation algorithm. In each case, the choice of what to amplify was recognised as a decision with consequences, made by someone.

The strongest objection

The objection is that holding platforms responsible for recommendations would lead them to recommend only safe, bland content, or to stop personalising altogether, which would harm the many users who benefit. Recommendation is not endorsement; a system that shows millions of items cannot vouch for each. And users bear responsibility for what they choose to watch. These are serious concerns, and regulators have generally avoided making platforms liable for each item. The response of the responsibility tradition is that systems which shape what billions of people see must be accountable for their overall effects, through assessment, transparency and design, even if not for every individual case.

What it means today

For companies, the decision of what to amplify is a decision for which they are increasingly held accountable, legally and publicly, and it deserves governance comparable to decisions about what to publish. For leaders in any organisation that uses algorithms to allocate attention, opportunities or resources, the problem of many hands is a warning: if no one owns the objective and its effects, no one will be responsible when they go wrong.

Bringing the threads together

Recommendation systems made a choice about taste: that it is a pattern in behaviour, discoverable without judging quality. They became choice architects for billions, arranging options to serve objectives their users did not choose. They served first-order desires efficiently and left second-order desires, what people want to want, largely unseen. They balanced exploration and exploitation in favour of the predictable. They made their own predictions true, reshaping taste and the content made to satisfy it. They fragmented the shared feed, though less than feared. They optimised measures that were never quite the good they stood for. And they made decisions about amplification for which someone must be responsible.

What you might also like

The phrase “you might also like” is modest. It suggests an offer, not a command, and that is how recommendation began: a way to help people find things in an abundance too large to browse. It has done that brilliantly. People discover music from other continents, documentaries on subjects they never knew existed and creators they would never have met. Most recommendations, most of the time, are welcome.

But the history also shows what happens when the offer becomes the environment. When the feed chooses everything, the objective it pursues becomes the most important design decision on the platform, and the people it serves can see neither the objective nor the alternatives. The working principles that follow are not complicated. Know what a recommender is optimising before you trust it. Keep some sources of the unexpected that are not tuned to your past. Act on what you want to want, by setting defaults in advance. If you build such systems, treat the objective as a moral choice, measure what it misses and give people real control. And if you lead an organisation, remember that every target you set is a recommender for your people’s attention. The question of what you might also like is worth asking. It is worth asking, too, who decided what you should be shown.

Selected bibliography

• Aristotle. Nicomachean Ethics. Fourth century BCE.
• Walpole, Horace. Letter to Horace Mann, 28 January 1754.
• Hume, David. “Of the Standard of Taste.” 1757.
• Kant, Immanuel. Critique of Judgement. 1790.
• Mill, John Stuart. On Liberty. 1859.
• Mill, John Stuart. Utilitarianism. 1863.
• Merton, Robert K. “The Self-Fulfilling Prophecy.” Antioch Review 8, no. 2 (1948): 193–210.
• Robbins, Herbert. “Some Aspects of the Sequential Design of Experiments.” Bulletin of the American Mathematical Society 58 (1952): 527–535.
• Habermas, Jürgen. The Structural Transformation of the Public Sphere. 1962.
• Frankfurt, Harry G. “Freedom of the Will and the Concept of a Person.” Journal of Philosophy 68, no. 1 (1971): 5–20.
• Bourdieu, Pierre. Distinction: A Social Critique of the Judgement of Taste. 1979.
• Gittins, J. C. “Bandit Processes and Dynamic Allocation Indices.” Journal of the Royal Statistical Society, Series B 41, no. 2 (1979): 148–177.
• Jonas, Hans. The Imperative of Responsibility. 1979.
• Thompson, Dennis F. “Moral Responsibility of Public Officials: The Problem of Many Hands.” American Political Science Review 74, no. 4 (1980): 905–916.
• Goldberg, David, David Nichols, Brian M. Oki and Douglas Terry. “Using Collaborative Filtering to Weave an Information Tapestry.” Communications of the ACM 35, no. 12 (1992): 61–70.
• Resnick, Paul, et al. “GroupLens: An Open Architecture for Collaborative Filtering of Netnews.” Proceedings of CSCW (1994).
• Shardanand, Upendra, and Pattie Maes. “Social Information Filtering: Algorithms for Automating ‘Word of Mouth’.” Proceedings of CHI (1995).
• Sunstein, Cass R. Republic.com. Princeton University Press, 2001.
• Linden, Greg, Brent Smith and Jeremy York. “Amazon.com Recommendations: Item-to-Item Collaborative Filtering.” IEEE Internet Computing 7, no. 1 (2003): 76–80.
• MacKenzie, Donald. An Engine, Not a Camera: How Financial Models Shape Markets. MIT Press, 2006.
• Narayanan, Arvind, and Vitaly Shmatikov. “Robust De-anonymization of Large Sparse Datasets.” IEEE Symposium on Security and Privacy (2008).
• Thaler, Richard H., and Cass R. Sunstein. Nudge. Yale University Press, 2008.
• Pariser, Eli. The Filter Bubble. Penguin, 2011.
• Bakshy, Eytan, Solomon Messing and Lada A. Adamic. “Exposure to Ideologically Diverse News and Opinion on Facebook.” Science 348, no. 6239 (2015): 1130–1132.
• Gomez-Uribe, Carlos A., and Neil Hunt. “The Netflix Recommender System: Algorithms, Business Value, and Innovation.” ACM Transactions on Management Information Systems 6, no. 4 (2015): article 13.
• Covington, Paul, Jay Adams and Emre Sargin. “Deep Neural Networks for YouTube Recommendations.” Proceedings of RecSys (2016).
• Susser, Daniel, Beate Roessler and Helen Nissenbaum. “Online Manipulation: Hidden Influences in a Digital World.” Georgetown Law Technology Review 4 (2019): 1–45.
• Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. Viking, 2019.
• Ribeiro, Manoel Horta, et al. “Auditing Radicalization Pathways on YouTube.” Proceedings of FAT* (2020).
• Hosseinmardi, Homa, et al. “Examining the Consumption of Radical Content on YouTube.” PNAS 118, no. 32 (2021).
• Guess, Andrew M., et al. “How Do Social Media Feed Algorithms Affect Attitudes and Behavior in an Election Campaign?” Science 381, no. 6656 (2023): 398–404.
• Regulation (EU) 2022/2065 (Digital Services Act), Articles 27, 34, 35 and 38.

Research limitations

The descriptions of how Netflix, YouTube, TikTok and Instagram rank content rely on the companies’ own publications and statements and on outside research; the systems change often and are not fully public. The figures of 80 percent of Netflix viewing and more than 70 percent of YouTube watch time are the companies’ own, from 2015 and 2018 respectively. The date and details of YouTube’s 2012 change to watch time come from trade press and company statements as reported secondarily. YouTube’s 70 percent reduction in borderline viewing is its own measure, for US recommendations to non-subscribers over a specific period. Research on filter bubbles and radicalisation is contested and evolving; the studies cited differ in method and period, and the Meta studies were conducted in collaboration with Meta, with independence safeguards described by the researchers. The account of the TikTok joint venture relies on reports of its announced terms. The philosophical connections to Hume, Kant, Bourdieu, Thaler and Sunstein, Susser, Roessler and Nissenbaum, Frankfurt, Mill, Robbins, Gittins, Walpole, Merton, Hacking, MacKenzie, Habermas, Negroponte, Pariser, Aristotle, Russell, Jonas and Thompson are the author’s interpretation; their works are paraphrased apart from short, attributed phrases. The decisions in the decisions box are judged by the author, separating what could be known at the time (process) from what happened (call).

Footnotes

1. Carlos A. Gomez-Uribe and Neil Hunt, “The Netflix Recommender System: Algorithms, Business Value, and Innovation,” ACM Transactions on Management Information Systems 6, no. 4 (2015), doi:10.1145/2843948. ↩ ↩²

2. “YouTube says 70% of all watch time is driven by its own recommendations,” Tubefilter, 11 January 2018. ↩ ↩²

3. David Goldberg, David Nichols, Brian M. Oki and Douglas Terry, “Using Collaborative Filtering to Weave an Information Tapestry,” Communications of the ACM 35, no. 12 (1992). ↩

4. Greg Linden, Brent Smith and Jeremy York, “Amazon.com Recommendations: Item-to-Item Collaborative Filtering,” IEEE Internet Computing (2003); Amazon Science, “The history of Amazon’s recommendation algorithm”. ↩

5. “Netflix cancels contest plans and settles suit,” New York Times Bits blog, 12 March 2010, archived copy; Netflix Technology Blog, “Netflix Recommendations: Beyond the 5 stars”, April 2012. ↩ ↩²

6. Hootsuite, “How the YouTube algorithm works”; Hypebot, August 2017, on the 2012 shift to watch time. ↩

7. Paul Covington, Jay Adams and Emre Sargin, “Deep Neural Networks for YouTube Recommendations,” RecSys 2016. ↩

8. YouTube, “The Four Rs of Responsibility, Part 2: Raising authoritative content and reducing borderline content,” December 2019. ↩

9. Manoel Horta Ribeiro et al., “Auditing Radicalization Pathways on YouTube” (2020); Homa Hosseinmardi et al., “Examining the consumption of radical content on YouTube,” PNAS (2021). ↩

10. “ByteDance signs deal to put TikTok US in control of Oracle-led consortium,” DatacenterDynamics; “TikTok U.S. joint venture deal set to close in January,” Variety; Tom’s Guide. ↩

11. “A primer on the Meta 2020 US election research studies,” Tech Policy Press; “Meta studies, data and polarization,” Columbia Journalism Review. ↩

12. Regulation (EU) 2022/2065, Article 38. ↩

13. Wikipedia, “Gonzalez v. Google LLC”; Reporters Committee for Freedom of the Press, “Gonzalez v. Google”. ↩