THE GREAT ATTENTION EXPERIMENT: 6 - THE MACHINES ARE LEARNING FASTER THAN WE ARE
Watch one video about running.
Nothing dramatic happens.
The system records that the video appeared.
It records whether you watched.
Whether you stopped after three seconds.
Whether you reached the end.
Whether you replayed part of it.
Whether you opened the comments.
Whether you followed the creator.
Whether you moved on immediately.
One interaction tells the system very little.
Then you watch another running video.
You pause on a review of a particular trainer.
You scroll past a marathon clip.
You finish a video about improving your pace.
You send a joke about runners to a friend.
The feed begins changing.
More running appears.
Then nutrition.
Recovery.
Fitness watches.
Weight loss.
Motivation.
Running holidays.
Products.
People whose lives seem to revolve around the activity you expressed a passing interest in twenty minutes ago.
Perhaps you genuinely want that.
Perhaps you have discovered something useful.
Perhaps the recommendations encourage you to become healthier.
Perhaps you were merely researching a present for someone else.
The system does not need to know.
It only needs to predict which piece of content you are most likely to respond to next.
That is what makes modern recommendation systems so powerful.
They do not wait to understand you as a person.
They learn from behaviour.
They make a prediction.
They observe the result.
Then they make another prediction.
You may take days or weeks to notice that your media environment has changed.
The machine can begin adapting after the next swipe.
A Feed Is Not a Library
A library contains more books than one person could ever read.
It does not decide which book should be placed directly into your hands.
You search.
Browse.
Ask a librarian.
Read a review.
Follow a reference from somewhere else.
The collection is large.
The act of selection remains visible.
A personalised feed is different.
The collection may be far larger than any library in human history.
But the user rarely encounters it as a collection.
They encounter a sequence.
One item.
Then another.
Then another.
The system decides which possibilities become visible and which remain hidden.
This solves a genuine problem.
Digital platforms contain more material than any person could navigate manually.
YouTube receives an enormous and constantly changing supply of videos.
Social platforms contain posts from friends, creators, publishers, businesses and strangers.
Music services contain millions of tracks.
Retailers offer vast catalogues.
Without ranking, search and recommendation, much of this abundance would be unusable.
Recommendation systems reduce information overload.
They help people discover creators with no existing audience.
They surface obscure music, specialist communities, useful tutorials and ideas someone would never have known to search for.
The machine does not merely compete for attention.
It can make attention more valuable by directing it towards something genuinely relevant.
That benefit is real.
It is also the source of the power.
The system stands between the person and the available world.
Chronology Could Not Survive Abundance
Early social feeds were often arranged chronologically.
The newest post appeared first.
Scroll down and time moved backwards.
The system decided relatively little beyond which accounts someone had chosen to follow.
That model became increasingly difficult as platforms grew.
A person might follow hundreds of accounts.
Thousands of posts could become available between visits.
Important material could disappear beneath whatever happened to be published most recently.
Chronology treats each post equally according to time.
It does not know that someone cares deeply about a photograph from a close relative but has little interest in another routine company announcement.
Ranking promised a better experience.
Instead of showing everything in order, the platform could estimate what mattered most.
Meta describes its Facebook feed-ranking systems as using machine learning to predict which content will be most relevant and valuable to each person from a vast collection of possible posts. The process gathers eligible content, applies models that predict different actions or outcomes, combines those predictions into scores and ranks the results.
Instagram uses different ranking systems for Feed, Stories, Explore, Reels and Search because people use each part of the application differently. Its published explanation says these systems draw upon information about the content, the person who posted it, the viewer's previous activity and their history of interacting with other accounts.
The machine does not simply ask what was posted last.
It asks what this particular person is most likely to value, watch, open, share or respond to now.
That is a far more useful question.
It is also far more commercially consequential.
The Social Graph Became the Interest Graph
Traditional social networks were organised primarily around people someone already knew.
Friends.
Family.
Colleagues.
Classmates.
Accounts the user had consciously chosen to follow.
The network reflected a social graph.
It mapped relationships.
Recommendation-led platforms can operate differently.
They do not need someone to follow a creator before showing their content.
They can infer that the person may be interested.
TikTok's For You feed is built around this principle.
TikTok says recommendations are ranked using factors including user interactions, video information and some device or account settings. Strong indications of interest, such as completing a longer video, can carry more weight than weaker contextual signals. TikTok also says the feed continuously learns from engagement and adjusts according to both positive and negative feedback.
The platform can therefore become useful before someone has built a network.
A new user does not need to locate hundreds of relevant accounts.
The system begins testing possibilities.
Sports.
Comedy.
Animals.
Politics.
Music.
Parenting.
Fitness.
Books.
Beauty.
Science.
It watches what happens.
The social graph asks:
Who have you chosen to connect with?
The interest graph asks:
What appears capable of holding your attention?
The second question can be answered much faster.
A friendship may take years to understand.
A behavioural prediction may begin forming within minutes.
There Is No Single Algorithm
People often talk about the algorithm as though it were one hidden machine making every decision.
Modern recommendation systems are usually collections of models, rules, databases, filters and ranking stages.
One component retrieves possible content.
Another predicts relevance.
Another estimates the likelihood of a click, view, share or other response.
Another checks eligibility and safety.
Another introduces diversity.
Another prevents repetition.
Another balances commercial content against ordinary material.
Another attempts to ensure the system does not become too narrow.
The output the user sees is the result of multiple objectives competing inside a technical system.
YouTube's published recommendation architecture described a two-stage process.
A candidate-generation network reduced an enormous collection to a smaller set of potentially relevant videos.
A separate ranking network then scored those candidates to determine which should appear most prominently.
Instagram Explore uses a similar funnel.
Meta describes retrieval, first-stage ranking, second-stage ranking and final re-ranking, with each stage reducing or reorganising the candidates before the final recommendations appear.
By 2025, Meta said Instagram's recommendation infrastructure had expanded to support more than one thousand machine-learning models serving different products and objectives.
That does not mean one thousand models are all analysing one person simultaneously.
It illustrates the scale and specialisation of the system.
There is no single switch labelled make this person keep scrolling.
There are many components making many predictions under objectives defined by the company.
The Machine Does Not Know You
Recommendation systems are often described as knowing people better than they know themselves.
That is an effective marketing line.
It is also misleading.
The system does not know that you watched a video about grief because it reminded you of someone you lost.
It does not know that you paused on an expensive car because you disliked the design.
It does not know that you finished an upsetting video because you were horrified.
It does not know that you searched for a medical condition on behalf of someone else.
It does not know that your daughter borrowed your phone.
It sees behaviour.
Then it assigns probabilities.
This person may watch another video like that.
This post may receive a response.
This product may attract a click.
This creator may retain the viewer.
The distinction matters.
Prediction is not understanding.
A machine can become extremely effective without possessing the human meaning behind the behaviour.
Weather software can predict rain without experiencing cold.
A chess system can select a move without understanding humiliation or ambition.
A recommendation system can predict that someone will continue watching without knowing whether they feel informed, inspired, anxious, angry or ashamed.
The prediction can still work.
Commercial systems are rewarded for accuracy of outcome, not depth of empathy.
Explicit Feedback Is Rare
The clearest way to learn what someone wants is to ask.
Do you want more of this?
Was this recommendation useful?
Did this improve your experience?
Do you regret the time you spent here?
Would you choose it again?
Digital platforms do ask some of these questions.
People can like or dislike content.
Follow or unfollow.
Rate a film.
Mark something not interested.
Hide a post.
Report a recommendation.
Complete a survey.
These are forms of explicit feedback.
They are valuable because the person intentionally communicates a preference.
The difficulty is that most people do not provide explicit feedback on everything they encounter.
They simply behave.
So systems rely heavily upon implicit signals.
Watch time.
Completion.
Pauses.
Clicks.
Shares.
Searches.
Replays.
Scrolling speed.
Return frequency.
The next item selected.
These behaviours are abundant.
They are also ambiguous.
A person can watch something without liking it.
They can share something to criticise it.
They can read comments because they are appalled.
They can replay a video because it was confusing.
They can stop scrolling because someone entered the room.
The system must convert messy human activity into a prediction.
That conversion is never neutral.
It reflects what the designers decide the behaviour is likely to mean.
Watch Time Is Not the Same as Value
Watch time is attractive because it is measurable.
A person who watches a video for fifty seconds has given more observable attention than someone who leaves after two.
The system can interpret this as evidence that the content was more relevant or enjoyable.
Often that inference will be correct.
TikTok has advised creators that watch time contributes to recommendation and that holding viewers' attention can improve a video's performance.
YouTube's recommendation research similarly describes ranking systems built around expected watch time and other measures intended to predict useful viewing.
But time is a proxy.
It stands in for something harder to measure.
Satisfaction.
Meaning.
Interest.
Regret.
Long-term value.
A person may spend twenty minutes watching an excellent lecture and feel grateful.
They may spend twenty minutes moving through outrage and feel worse.
Both create watch time.
This proxy problem is now acknowledged openly by the companies building the systems.
In January 2026, Meta explained that recommendation models relying mainly upon signals such as watch time, likes and shares can misinterpret what people genuinely care about. It said models trained only on these signals may favour high short-term engagement while failing to capture deeper interests or long-term product value.
That is a significant admission.
The system can become highly effective at predicting measurable behaviour while remaining wrong about what the person actually wants.
The Difference Between What We Do and What We Value
Human beings behave inconsistently.
We say sleep matters and continue watching.
We say family matters and interrupt conversations.
We say we dislike outrage and click it.
We say we want thoughtful information and respond fastest to whatever makes us angry.
This is not because we are dishonest.
Different motives operate at different times.
Immediate curiosity can defeat long-term intention.
Fatigue reduces resistance.
Emotion narrows judgement.
Habit begins before conscious reflection.
A recommendation system trained primarily upon behaviour learns the version of us that acts.
It does not automatically learn the version that later wishes we had acted differently.
That creates a dangerous possibility.
The machine may become excellent at predicting our impulses while being almost blind to our values.
A person might tell a platform:
I care about meaningful time with my family.
The system cannot easily monetise that statement.
But it can observe that short, provocative videos keep the same person watching late at night.
Which signal is more likely to shape the next recommendation?
The one that can be measured repeatedly.
The Machine Learns From Everyone
A recommendation system does not learn only from one individual.
It learns from patterns across populations.
People who watched this often watched that.
People who bought this considered those products.
People who enjoyed these songs often liked that artist.
People with similar histories responded to this recommendation.
This is the basic intuition behind collaborative filtering.
The system can identify connections that no individual user could detect.
Perhaps thousands of people who enjoy two obscure authors also enjoy a third.
Perhaps viewers who complete one specialist tutorial respond well to another creator with a different audience.
The machine can turn collective behaviour into personal discovery.
This is one of the most valuable aspects of recommendation.
It enables niche content to find the people most likely to appreciate it.
A creator does not necessarily need celebrity status.
A product does not need mass appeal.
The system can match a small audience with something unusually relevant.
But collective learning creates an asymmetry.
You learn from your own experience and the people you happen to know.
The machine learns from millions or billions of behavioural traces.
It sees patterns no person could hold in their mind.
This does not make its judgement wiser.
It makes its statistical reach larger.
The Machine Can Test What It Does Not Know
A recommendation system faces a constant tension.
Should it show what it already believes someone likes?
Or should it introduce something new?
Show only familiar material and the feed becomes repetitive.
Explore too widely and the recommendations feel irrelevant.
This is known as the exploration versus exploitation problem.
Exploitation uses what the system already believes.
Exploration tests another possibility.
A person who watches football may receive more football.
That is exploitation.
The system may also test rugby, athletics, sports documentaries or comedy from a football creator.
That is exploration.
If the person responds, the profile changes.
The user experiences discovery.
The system experiences information.
This is not inherently sinister.
Without exploration, new creators would struggle to reach audiences.
People would be trapped inside narrow histories.
The system could never learn that interests had changed.
But every exploratory recommendation is also a small behavioural test.
Will this person watch?
Will they pause?
Will they reject it?
Will they follow the pathway?
The user may believe they are simply browsing.
The system is reducing uncertainty.
Real-Time Learning Changed the Speed
Earlier recommendation systems were often trained in batches.
Data was collected.
Models were updated periodically.
A person's morning behaviour might not influence the system until a later training cycle.
Modern systems increasingly aim to learn much faster.
ByteDance researchers described a recommendation architecture called Monolith designed for online training and time-sensitive feedback.
Their paper explains that short-video and advertising systems require models capable of responding to rapidly changing user behaviour and content. The architecture was built to allow recent feedback to influence recommendations more quickly than traditional systems separating training from live operation.
ByteDance's published Monolith framework says real-time training helps capture emerging trends and enables rapid discovery of new interests.
The important phrase is new interests.
A system does not need to wait for someone to declare a new identity.
It can detect a pattern emerging through behaviour.
Yesterday, no gardening content.
Today, three completed videos.
Tomorrow, seeds, tools, allotments and creators.
The system updates because recent behaviour may predict the next response better than an old profile.
The person may not yet have decided whether gardening is an interest.
The feed has already begun treating it as one.
Learning Faster Does Not Mean Thinking Better
The title of this essay can easily be misunderstood.
Machines are not necessarily learning in the rich human sense.
They do not learn what it feels like to become a parent.
They do not develop wisdom from regret.
They do not understand why one conversation matters for the rest of a life.
They update parameters.
Adjust probabilities.
Recognise statistical patterns.
Improve predictions against defined objectives.
Within that narrow domain, they can learn extraordinarily quickly.
They can process more behavioural examples than any human being could observe.
They can run many models simultaneously.
They can rank thousands or millions of possibilities in fractions of a second.
Meta says its advertising recommendation systems can rank thousands of adverts within a few hundred milliseconds.
Its Andromeda retrieval system reduces tens of millions of possible adverts to a smaller set before more complex ranking models decide which should be shown.
In 2026, Meta reported a new retrieval architecture capable of dramatically higher throughput and greater computing efficiency for user-generated content recommendations.
The machine learns faster because the problem has been narrowed.
Predict the click.
Predict the view.
Predict the share.
Predict the response.
A human being is trying to solve a much larger problem.
What do I actually want?
Was that good for me?
What kind of person am I becoming?
What deserves my limited life?
The system can update before we have even asked those questions.
One Person Sees One Version
Suppose a platform tests two versions of a feed.
One group sees more familiar content.
Another sees more novelty.
The company can compare millions of outcomes.
The individual sees only one version.
They cannot know what would have happened under the alternative.
Perhaps they would have left earlier.
Perhaps they would have discovered something better.
Perhaps their mood would have changed differently.
Perhaps they would have spent the next hour doing something else.
This makes algorithmic influence unusually difficult to perceive.
The machine possesses the comparison.
The user possesses the experience.
The company can see average differences across populations.
The person cannot see the life they would have lived under another ranking system.
That invisible alternative is where much of the argument about agency becomes stuck.
We can observe what someone did.
We cannot easily observe what they would have chosen in an environment arranged differently.
The Recommendation Changes the Evidence
A recommender system learns from what people select.
But people can select only from what the system shows them.
This creates a feedback loop.
The machine recommends.
The user responds.
The response becomes training data.
The updated system recommends again.
Imagine a music service that believes someone prefers one genre.
It recommends more of that genre.
The person listens because it is available.
The additional listening confirms the original belief.
Other genres receive fewer opportunities to be selected.
The profile becomes more confident.
But confidence does not prove that the original preference was fixed.
The recommendation helped create the behaviour later used as evidence.
Researchers describe this as algorithmic confounding.
Simulation research has shown that recommendation systems trained upon behaviour already influenced by previous recommendations can homogenise consumption and reduce the usefulness of recommendations over time.
Further theoretical work has shown how recommender systems and changing user preferences can create feedback loops capable of narrowing exposure or producing degenerative patterns if the system continually reinforces its own predictions.
Other research has found that such loops can amplify popularity bias, reduce aggregate diversity and shift the representation of users' tastes, with stronger effects sometimes falling upon minority-interest users.
These studies do not prove that every real platform traps every user in a filter bubble.
Many use models or simulations because independent researchers cannot reproduce the complete internal systems of major platforms.
The findings identify a structural risk.
The machine may begin treating its own influence as evidence of natural preference.
The Feed Does Not Merely Reflect Taste
People often defend recommendation systems by saying:
It only shows you what you engage with.
There is truth in that.
Someone who repeatedly watches a topic will usually receive more of it.
The system is responding to behaviour.
But the sentence is incomplete.
The user did not generate every possible item and arrange them neutrally before selecting one.
The system selected what was available to engage with.
The process is circular.
The feed reflects behaviour.
The feed shapes the behaviour available to be reflected.
This does not mean preference is manufactured from nothing.
A football fan was probably interested in football before the platform noticed.
A person seeking extremist content may bring those beliefs with them.
A child struggling with body image may already feel insecure before receiving recommendations.
Algorithms are not the sole authors of human desire.
But amplification matters.
Frequency matters.
Sequence matters.
A passing curiosity exposed repeatedly can become more psychologically prominent.
A fear encountered occasionally can begin to feel universal.
A fringe belief shown continuously can appear mainstream.
A person may mistake the contents of a personalised feed for a reliable picture of the wider world.
Behaviour Can Reflect and Shape Mental State
Online behaviour does not emerge from nowhere.
People search according to how they feel.
Someone feeling anxious may seek threatening information.
Someone feeling low may consume darker material.
Someone feeling lonely may search for connection.
The content then affects the person who receives it.
A 2025 study of web-browsing behaviour found that the emotional quality of information people sought reflected their current mood and could also influence subsequent mood, creating a potential feedback loop between psychological state and information consumption.
The study concerned web browsing rather than one particular social-media recommender.
It does not establish that platforms cause mental-health conditions.
It illustrates the wider principle.
People influence the information they consume.
Information influences people.
An adaptive system sits inside that loop.
If a person's behaviour changes because they are vulnerable, the recommender may respond to the changed behaviour without understanding the vulnerability beneath it.
It sees increased engagement with a category.
It may supply more.
AI Can Affect the Human Judgement That Trains It
The feedback-loop problem extends beyond social media.
Experiments involving more than 1,400 participants found that biased AI judgements could alter subsequent human perceptual, emotional and social judgements.
Those altered human responses then risked reinforcing the biased system in a continuing human-AI loop.
This research did not test ordinary entertainment feeds.
Its significance is broader.
AI does not merely receive human information.
Its outputs can change the humans who later provide the next information.
The system and user do not remain independent.
They adapt to one another.
When this process occurs across billions of daily interactions, distinguishing between preference discovery and preference formation becomes extraordinarily difficult.
Creators Learn the Machine Too
The recommendation system does not influence only viewers.
It influences creators.
A creator learns which openings retain attention.
Which video length performs.
Which topic receives distribution.
Which thumbnail attracts clicks.
Which emotional tone produces comments.
Which controversy generates reach.
The creator adjusts.
The system then receives content better adapted to its existing metrics.
This creates another feedback loop.
The algorithm rewards certain characteristics.
Creators produce more of them.
Users encounter more of the resulting content.
Their behaviour confirms that the content performs.
The platform distributes it further.
This does not mean creators lose all originality.
Recommendation systems can help genuinely original work find an audience.
They can bypass traditional gatekeepers.
But the system becomes an invisible editor operating across an entire culture.
It does not need to tell creators what to make.
It distributes rewards unevenly until people learn what travels.
The Machine Learns Our Short-Term Selves
A recommendation model often receives feedback quickly.
Did the person click?
Did they watch?
Did they return?
Long-term consequences are harder to connect.
Did the person sleep well?
Did they feel the session was worthwhile?
Did the content improve their understanding?
Did they become more anxious over six months?
Did they lose interest in slower activities?
Did a family relationship change?
These outcomes may occur far away from the platform.
They may be difficult to measure.
They may appear years later.
They may have many causes.
The system therefore learns most easily from what happens immediately.
That produces a structural bias towards short-term signals.
An outrage-inducing post generates an immediate comment.
Its effect upon trust in society is delayed and diffuse.
A short video generates a measurable completion.
Its effect upon tolerance for slower material is uncertain.
A recommendation generates another session.
The lost sleep happens later.
The machine learns rapidly where the evidence is dense.
Human consequences often emerge where the evidence is slow.
More Intelligence Does Not Automatically Produce Better Objectives
Machine-learning systems are becoming more capable.
They can process richer sequences.
Understand text, audio and video.
Model recent behaviour alongside long-term interests.
Predict several possible outcomes simultaneously.
Use direct survey feedback.
Balance relevance with novelty and diversity.
These improvements can create better experiences.
Meta's recent work on Facebook Reels attempts to move beyond watch time and ordinary engagement signals by incorporating feedback about whether content genuinely matches people's interests, feels novel or contributes to long-term value.
That development is encouraging.
It also reveals an important truth.
A more powerful model does not solve the problem of choosing the correct objective.
It predicts whatever it has been asked to predict.
If the objective is short-term watch time, better AI may optimise short-term watch time more effectively.
If the objective includes satisfaction, diversity, safety and long-term value, it can attempt to balance those outcomes instead.
The intelligence sits in the method.
The values sit in the objective.
Machines do not decide what deserves optimisation.
People do.
Safety Systems Are Learning Too
Recommendation technology is not used only to increase engagement.
The same machine-learning capability can identify spam.
Reduce repetitive recommendations.
Detect content that violates platform rules.
Limit material judged unsuitable for broad audiences.
Diversify feeds.
Protect children.
TikTok says it deliberately interrupts some recommendation patterns, limits repeated exposure to certain categories and introduces diverse content to avoid overly narrow feeds.
Meta says Instagram Explore incorporates additional ranking and re-ranking stages intended to improve diversity, quality and relevance rather than simply repeating the most obvious prediction.
The existence of these safeguards matters.
It demonstrates that recommendation systems are not inevitably harmful.
They can be configured differently.
It also demonstrates that undesirable outcomes are not external to the machine.
They are design problems requiring deliberate counterweights.
If a platform must engineer diversity into the system, pure prediction did not automatically create it.
If it must restrict repetitive or harmful pathways, engagement signals alone were insufficient.
Safety does not emerge automatically from optimisation.
It must become an objective.
Regulators Now Treat Recommendations as Infrastructure
For years, public arguments focused mainly upon whether individual pieces of content should be removed.
That remains important.
But regulators increasingly recognise that the distribution system matters as much as the content itself.
A harmful post seen by ten people and the same post recommended to ten million people are not socially equivalent.
The UK's communications regulator has described personalised recommender systems as a principal route through which children encounter harmful material online.
Under the Online Safety regime, services judged to present medium or high risks to children must configure recommendation systems to filter or reduce harmful material in children's feeds.
In March 2026, Ofcom said it was issuing legally binding information requests to major platforms as part of its assessment of their recommendation systems.
It also called for significant product changes to be risk-assessed before being deployed to children.
By May 2026, Ofcom said it remained concerned about the responses from some major platforms, including TikTok and YouTube, concerning how they would make recommender feeds safer for children.
The European Union's Digital Services Act also requires greater transparency about the main parameters used by recommender systems.
Very large platforms using recommender systems must provide at least one option that is not based upon profiling.
These measures do not eliminate algorithmic influence.
They change the assumption that recommendation is merely a private product feature.
Ranking systems have become part of the information infrastructure of society.
Users Can Teach the Feed Deliberately
The machine's speed does not remove human agency.
People can influence recommendation systems intentionally.
TikTok allows users to mark content as unwanted, adjust preferences and refresh the For You feed so that recommendations begin rebuilding from new interactions.
Instagram offers tools to mark content as not interested, filter topics and reset suggested content across Feed, Explore and Reels. After a reset, recommendations personalise again according to new behaviour.
YouTube allows users to remove videos from watch history, pause history, reject channels and indicate that recommendations are not relevant.
When watch history is removed or switched off, future recommendations no longer use that activity in the same way.
These controls are useful.
They also expose the imbalance.
The system learns through almost every ordinary interaction.
Correcting it often requires deliberate effort.
The user must notice that the feed has drifted.
Understand that their behaviour trained it.
Locate the appropriate control.
Provide clearer feedback.
Perhaps erase part of the history.
Then resist reproducing the same pattern.
The machine learns by default.
The human being must often unlearn deliberately.
A Reset Does Not Remove the Business Model
Recommendation resets and user controls are positive developments.
They allow people to interrupt a profile that no longer reflects them.
But the underlying system remains.
After the reset, personalisation begins again.
The platform still needs to select content.
The person still produces behavioural signals.
The business still benefits when recommendations create valuable activity.
The machine still updates faster than the user can fully understand its effects.
Controls can return some agency.
They do not resolve the economic incentive determining what the system is primarily built to optimise.
That is why digital literacy cannot consist only of teaching people which buttons to press.
People need to understand the loop.
You are not seeing the internet.
You are seeing a selection from it.
The selection is based partly upon predictions drawn from previous behaviour.
Your response becomes evidence shaping the next selection.
The feed is not simply reflecting who you are.
It is participating in what receives enough repetition to become familiar, important or normal.
The Child Cannot Compete With the Model
This asymmetry becomes most serious for children.
An adult may understand that the feed is personalised.
They may recognise clickbait.
Notice emotional manipulation.
Question why a topic keeps appearing.
Remember that another person sees a completely different world.
A child is still developing those capabilities.
They may interpret frequency as importance.
Popularity as truth.
Repetition as normality.
Personalisation as coincidence.
They do not know how many alternatives the system rejected before selecting what appeared.
They cannot easily distinguish between something chosen because it is beneficial and something chosen because similar children watched it for longer.
The child is learning about themselves.
The system is learning how to retain them.
Those are not equal processes.
The child requires years to develop judgement.
The model can update through the next interaction.
We Are Still Adapting to Yesterday's System
Public understanding often trails technology.
By the time society learned to recognise banner advertising, targeted advertising had developed.
By the time people understood tracking cookies, platforms were shifting towards first-party data, device signals and logged-in ecosystems.
By the time parents understood chronological social feeds, recommendation-led short-form video had become dominant.
By the time schools began teaching social-media safety, generative artificial intelligence had started creating personalised text, images, audio and video.
The machines are learning faster than we are partly because they are designed to.
Continuous data.
Continuous testing.
Continuous deployment.
Human understanding moves through research, public discussion, education, legislation and cultural change.
Each takes time.
The system updates tonight.
The school curriculum may update in several years.
The regulator investigates after evidence accumulates.
The parent notices after behaviour changes.
The child may understand only when the habit has already formed.
This Is Not a Contest We Can Win Through Willpower
It is easy to respond by telling people to become more disciplined.
Pay more attention.
Scroll less.
Think critically.
Set better boundaries.
All of those actions can help.
But the contest is structurally uneven.
One side is a human being who is tired, distracted, emotional and attempting to live a full life.
The other is an infrastructure capable of observing enormous populations, evaluating countless possibilities and updating predictions at machine speed.
The answer cannot be for every person to maintain perfect vigilance throughout every interaction.
That would place the entire moral burden upon the least informed participant.
Individuals need agency.
Parents need to teach.
Schools need to prepare children.
But platforms must also accept responsibility for what their optimisation systems repeatedly produce.
Regulators must understand ranking, not merely content.
Designers must test long-term value, not only immediate behaviour.
Companies must be prepared to accept that the most engaging recommendation is not always the one that deserves to appear.
What Would It Mean for Humans to Learn Faster?
We will never process behavioural data at machine speed.
That should not be the objective.
Human beings possess forms of understanding recommendation systems do not.
We can question the goal.
Recognise regret.
Value a difficult experience that produced no immediate pleasure.
Choose something because it matters rather than because it is easy.
Decide that an accurate prediction should not be acted upon.
The machine may correctly predict that outrage will hold attention.
A human being can decide not to optimise for outrage.
The machine may identify that a vulnerable child will continue watching a harmful category.
A human being can decide that the category should not be recommended.
The machine may discover that removing a stopping point increases use.
A human being can restore the stopping point.
Learning faster does not mean matching calculation with calculation.
It means recognising sooner that prediction is not wisdom.
Engagement is not wellbeing.
Preference is not destiny.
Behaviour is not consent.
What holds attention is not automatically worthy of it.
Final Thought
The machines do not know us.
They predict us.
They watch what we do.
Compare it with what others have done.
Choose something likely to produce a response.
Observe whether the prediction was correct.
Then update.
They do this across billions of interactions.
At a speed no individual human being can match.
We experience one feed.
They compare millions.
We see one recommendation.
They evaluate countless alternatives.
We may need weeks to recognise that our interests, habits or mood have shifted.
The system can respond after the next action.
That does not make the machine conscious.
It does not make influence absolute.
It does not remove human agency.
It creates an extraordinary imbalance in learning speed.
The recommendation affects the behaviour.
The behaviour trains the recommendation.
The loop continues.
And somewhere inside that loop, the difference between what we already wanted and what we learned to want becomes increasingly difficult to see.
The machines are learning faster than we are.
But speed has never been the same thing as wisdom.
The future of human attention will depend upon whether we learn that distinction before the systems become even better at predicting the parts of us we have not yet learned to control.
Sources & Research Gaps
Principal Sources
Google Research, Deep Neural Networks for YouTube Recommendations
Google's research paper describes YouTube's large-scale recommendation architecture, including separate candidate-generation and ranking stages.
The paper demonstrates how deep learning is used to reduce an enormous catalogue to a smaller collection of personalised recommendations.
Google Research, Recommending What Video to Watch Next
This research describes a large-scale multi-objective ranking system for predicting which video should be recommended next.
It highlights the challenge of balancing several objectives and interpreting user feedback that may already contain selection bias.
TikTok, How the For You Feed Works
TikTok's official explanations identify several recommendation signals, including user interactions, video information and contextual account or device settings.
The company states that watch completion and other strong behavioural signals can carry greater weight than weaker contextual indicators.
TikTok also describes its recommendation system as continuously learning from user engagement.
Meta Engineering, Facebook Feed Ranking
Meta describes a multi-stage process through which eligible posts are gathered, scored and ranked according to machine-learning predictions about relevance and value to an individual user.
Meta, Instagram Ranking Explained
Instagram's official explanation states that Feed, Stories, Explore, Reels and Search use different ranking systems based upon how people use each surface.
Signals include information about content, creators, user activity and interaction history.
Meta Engineering, Instagram Explore
Meta's engineering publications describe the retrieval, ranking and re-ranking stages used to reduce large collections of possible content to final personalised recommendations.
Meta Engineering, Journey to 1,000 Models
Meta reported in 2025 that Instagram's recommendation infrastructure supported more than one thousand machine-learning models across different products, objectives and performance requirements.
Meta Engineering, Reels Recommendation and User Feedback
Meta acknowledged in 2026 that traditional signals such as likes, shares and watch time can be noisy and may overemphasise short-term engagement rather than people's deeper interests or long-term product value.
ByteDance, Monolith
ByteDance researchers describe Monolith as a recommendation architecture designed for online training and time-sensitive feedback.
The work demonstrates the industry's movement from periodic model updates towards systems capable of adapting more rapidly to changing interests and emerging content.
Chaney, Stewart and Engelhardt, Algorithmic Confounding
This research uses simulations to demonstrate how systems trained upon behaviour already influenced by recommendations can homogenise consumption and reduce recommendation utility.
Jiang and Colleagues, Degenerate Feedback Loops
The authors provide a theoretical analysis of the interaction between recommendation systems and changing user preferences.
Their work identifies mechanisms through which repeated recommendation and response can produce narrowing or degenerative feedback loops.
Mansoury and Colleagues, Bias Amplification
This simulation research found that recommendation feedback loops can amplify popularity bias, reduce aggregate diversity and shift representations of user taste.
The effects were not uniform across all user groups.
Glickman and Colleagues, Human-AI Feedback Loops
A series of experiments involving more than 1,400 participants found that biased AI outputs could influence later human perceptual, emotional and social judgements.
Those altered human judgements could then reinforce the AI system's bias.
Kelly and Colleagues, Browsing and Mood
The researchers found evidence that the emotional quality of online information people sought both reflected and influenced their mood, creating a possible feedback loop between psychological state and information consumption.
Ofcom, Online Safety Regulation
Ofcom has identified personalised recommendation feeds as an important pathway through which children encounter harmful material.
Its regulatory work requires relevant services to configure recommendation systems to reduce children's exposure to harmful content and to assess risks from major product changes.
European Union Digital Services Act
The Digital Services Act requires platforms to provide information about the principal parameters used by recommender systems.
Very large platforms using recommender systems must also provide at least one recommendation option not based upon profiling.
Platform Recommendation Controls
TikTok, Instagram and YouTube provide tools allowing users to reject recommendations, reset or refresh feeds, remove behavioural history and influence future personalisation.
Research Gaps and Limitations
The phrase the machines are learning is used metaphorically.
Machine-learning systems update statistical models and predictions.
They do not necessarily possess awareness, understanding, intention or human-like knowledge.
Public explanations of recommendation systems describe broad principles.
They do not reveal every signal, weight, rule, model or commercial objective used by major platforms.
Recommendation systems change continuously.
A technical paper may accurately describe one architecture without representing the platform's complete current system.
The ByteDance Monolith paper describes an industrial recommendation architecture and its deployment within BytePlus Recommend.
It should not be treated as a complete public description of every system used by TikTok.
A behavioural signal does not have one fixed meaning.
Long watch time may indicate enjoyment, confusion, horror, distraction or simple inactivity.
Companies use multiple signals and models rather than interpreting every action identically.
Feedback-loop studies often rely upon simulations, theoretical models or limited audits because researchers lack access to complete platform data and production systems.
These studies identify plausible and important risks.
They do not prove that every user experiences a filter bubble or that every platform inevitably homogenises behaviour.
Personalisation may broaden exposure as well as narrow it.
Recommendation systems can help people discover unfamiliar topics, minority creators and content outside their existing social networks.
The relationship between recommendation and preference is bidirectional.
People bring existing interests, beliefs and vulnerabilities to platforms.
Algorithms respond to those characteristics while also influencing what becomes visible and repeated.
It is usually difficult to isolate the proportion of a later preference caused by recommendation.
Engagement does not automatically indicate harm.
Many highly engaging experiences are valuable, educational or socially meaningful.
Low engagement does not automatically indicate quality.
A difficult book, important warning or challenging piece of education may receive less immediate interaction while remaining highly valuable.
User controls can influence recommendations, but their practical effectiveness depends upon whether people understand, locate and use them.
Regulatory requirements are evolving.
Implementation, enforcement and platform compliance vary across jurisdictions.
Further independent research is needed into:
- how quickly major recommendation systems adapt to individual behaviour;
- which implicit signals carry the greatest weight on different platforms;
- how recommendation objectives balance watch time, satisfaction, diversity, safety and revenue;
- the extent to which users' stated preferences conflict with observed behaviour;
- how often recommendation systems create new interests rather than detecting existing ones;
- how ranking affects children's beliefs about popularity, normality and importance;
- the long-term effects of recommendation feedback loops upon identity and preference;
- whether feed resets produce lasting changes or merely restart the same behavioural cycle;
- how creators adapt content to algorithmic distribution incentives;
- how generative AI will change personalised recommendation;
- whether regulators can audit complex ranking systems effectively without exposing legitimate security or trade-secret information;
- which non-profiled or chronological alternatives meaningfully restore user agency.