There is a particular kind of scrolling in which everything looks familiar even when none of it has been seen before.
The same style of caption.
The same enormous subtitles over another short video.
The same dramatic pause before somebody explains something simple.
The same conflict, edited around a different person.
The same product attached to another personality.
The same joke recreated by another account.
None of this proves the internet has become fake.
Billions of real people are still using it to speak, argue, learn, make friends, build careers, waste time and create genuinely strange things.
But the machinery deciding what reaches them has changed.
The feed replaced the destination
Earlier versions of the web often required people to decide where they wanted to go.
A forum had to be visited.
A blog had to be bookmarked.
A website had to be found and then remembered.
A user subscribed to particular creators or joined a particular community and returned because they intended to.
Modern platforms increasingly begin from the opposite direction.
Open the app and the material is already waiting.
The platform predicts what should come next.
That recommendation can be extraordinarily useful.
It allows an unknown musician, filmmaker, comedian or specialist to reach millions of people who would never have typed their name into a search box.
It also changes the relationship between creator and audience.
The creator is no longer speaking only to people who chose to arrive.
They are increasingly producing material for a system deciding whether the audience should be delivered.
Content learns what distribution rewards
A recommendation system does not need to send creators a secret instruction manual.
The scoreboard is visible.
One video reaches four thousand people.
Another reaches four million.
A particular opening keeps viewers watching.
A particular title attracts clicks.
A particular conflict generates comments.
Creators learn.
Agencies learn.
Brands learn.
Competitors copy what worked.
When millions of creators learn from the same scoreboard, different people can begin producing strangely similar work.
This does not mean platforms only recommend outrage
Recommendation systems are complicated and constantly changing.
Platforms incorporate many signals and publicly describe goals beyond raw engagement, including relevance, originality, safety and user satisfaction.
Their own business results nevertheless show how important engagement remains.
Meta said ranking improvements introduced during the fourth quarter of 2025 increased views of organic Facebook feed and video posts by 7 per cent, while video time spent in the United States grew by double digits year-on-year.
The point is not that watching something for longer is automatically bad.
It is that time, clicks, replies, shares and other measurable behaviour inevitably become part of the economic environment in which online culture develops.
Europe now treats recommender systems as something users deserve to understand
The European Union's Digital Services Act requires online platforms to explain the main parameters used by their recommender systems.
Very large platforms must also provide at least one recommendation option that is not based on user profiling.
In practice, major services including TikTok, Facebook and Instagram offer European users ways to move away from fully personalised recommendation feeds.
The regulation does not declare personalised feeds bad.
It recognises that there is a meaningful difference between choosing what to follow and continuously allowing a platform to choose what comes next.
Regulation / 2026
The design of the feed has become a regulatory issue.
In July 2026, the European Commission preliminarily found Meta in breach of the Digital Services Act over the addictive design of Instagram and Facebook.
The investigation focuses on features including infinite scroll, autoplay, push notifications and highly personalised recommendation systems. The finding is preliminary, not a final judgment.
Professional creators are not evidence that the internet became less authentic
Online creation was commercial long before generative AI.
Websites sold advertising.
Bloggers ran businesses.
YouTube turned people filming in bedrooms into production companies.
Podcasting created networks.
Social platforms created entire industries of editors, managers, agencies, brand strategists and analytics specialists.
Professionalisation produced genuinely better work.
Independent creators can now make documentaries, investigative videos, music and entertainment that once required access to conventional media organisations.
A person earning money from creativity does not make the creativity fake.
The more interesting change is what happens when the same analytics and commercial incentives reach almost every layer of cultural production.
An ordinary post can become a miniature performance report
A creator can now know exactly how many people stopped watching after the first three seconds.
They can compare thumbnails.
They can see which subjects produced new followers.
They can measure which upload converted attention into money.
That information is useful.
It also means creative decisions take place under unusually intense measurement.
What begins as personal expression can slowly become optimisation.
Not because the creator became dishonest, but because the creator learned what survival on the platform requires.
Then generative AI made production dramatically cheaper
Artificial intelligence did not invent repetitive content.
It changed the cost of producing it.
Text can be drafted in seconds.
Images can be generated in large batches.
Voices can be synthesised.
Music can be created without conventional recording.
Video generation has improved quickly enough that platforms now maintain specific transparency policies for synthetic footage.
Those tools can give a small creator capabilities previously available only to a larger team.
They can translate material, remove technical barriers and make experimentation affordable.
They can also make another hundred disposable posts almost free.
We finally have evidence that AI-written webpages are becoming common
In August 2026, Pew Research Center published an analysis of almost half a million English-language webpages collected through the Common Crawl archive.
Pew used an AI-detection model to estimate which pages showed significant signs of having been written or substantially edited by artificial intelligence.
In its July 2026 snapshot, 10 per cent of all sampled webpages showed significant signs of AI authorship.
The entire sample includes older webpages that could not have been produced with modern generative systems.
When Pew looked only at pages published after the release of ChatGPT, more than one third showed signs of AI authorship.
English-language webpages examined across Pew's five-year Common Crawl sample.
Pages in the July 2026 snapshot showing significant signs of AI authorship.
Approximate share among pages published after ChatGPT's public release.
Detection is not a census
That research needs a qualification.
Pew did not possess the private writing history of every webpage.
It used a machine-learning detector.
AI-detection systems can produce false positives and false negatives.
The study is therefore evidence of a major trend, not a perfect accounting system capable of assigning every sentence on the internet to either a human or machine.
The direction is nevertheless difficult to dismiss.
The amount of machine-assisted writing on the open web has increased rapidly since late 2022.
“AI slop” describes economics as much as aesthetics
The term "AI slop" is deliberately insulting.
It usually refers to low-effort generated material produced in volume.
The problem is not simply that AI touched the work.
A novelist can use software to organise notes.
A filmmaker can use generative tools for an effect.
A journalist can use transcription.
An illustrator can incorporate generated material intentionally into a larger composition.
The cultural complaint is normally about material whose reason for existing appears to be that producing another item became nearly free.
Quantity is the strategy.
The platforms clearly know this problem exists
In July 2025, YouTube renamed its "repetitious content" monetisation rule to "inauthentic content."
The company clarified that repetitive or mass-produced videos had long been ineligible for monetisation.
Its current rules specifically identify AI-generated material using generic or unoriginal templates in a way that gives the impression of mass production as an example of content that may not qualify.
YouTube's separate spam rules also prohibit automated or synthetic mass-production used to flood the platform with highly similar material.
The policy does not ban AI creation.
It distinguishes between creation and industrialised repetition.
Meta has moved towards rewarding original material too
Meta said in March 2026 that it had been changing Facebook Feed and Reels to increase distribution of original content while reducing the reach of unoriginal material.
According to the company's own reporting, views and time spent watching original Facebook Reels approximately doubled in the second half of 2025 compared with the same period in 2024.
Meta also said 75 per cent of Instagram recommendations in the United States were coming from original posts by the fourth quarter of 2025.
Those are Meta's own measurements, not an independent audit.
They are still revealing.
A platform does not need a major programme to promote original creators unless duplication, repackaging and low-value reproduction have become important enough to affect what users see.
TikTok is now testing against AI spam directly
TikTok has gone further in acknowledging generative volume.
In July 2026, the company said it was testing improved detection aimed at accounts dedicated to posting AI-generated spam that crowds out original creators.
It had already begun testing an option allowing users to influence how much AI-generated content appears in their For You feed.
Again, this is not a ban on generated culture.
TikTok itself promotes generative tools.
It is recognition that abundance creates a ranking problem.
When creation becomes cheap enough, the platform has to distinguish somebody experimenting with a tool from an account using automation as a content factory.
What major platforms are doing
The response is increasingly about originality and transparency.
- YouTube restricts monetisation for repetitive or mass-produced "inauthentic content."
- Meta says original Facebook and Instagram material receives greater recommendation priority.
- TikTok is testing detection against accounts posting AI-generated spam at scale.
- YouTube and TikTok require disclosure of certain realistic AI-generated or materially altered media.
- Major platforms are adopting or supporting provenance technology such as C2PA Content Credentials.
Search is changing from a map of the web into an answer
Social media is only one part of the change.
Search has historically been one of the most important ways users left one website and discovered another.
AI-generated search summaries alter that journey.
Pew analysed the web-browsing behaviour of 900 US adults during March 2025.
Fifty-eight per cent encountered at least one search page containing an AI-generated summary during the month.
When a Google search included an AI summary, users clicked a conventional search-result link in 8 per cent of visits.
When no summary appeared, they clicked a result in 15 per cent.
Links cited inside the AI summaries themselves received a click in only around 1 per cent of visits to pages containing a summary.
Search / Pew Research Center
The answer can now arrive before the website.
Pew's 2025 browsing study does not prove AI summaries caused every difference in clicking behaviour. Queries that trigger summaries may differ from those that do not.
The result still demonstrates a structural change: a growing share of users can receive a synthesised answer without visiting the pages from which online information originally came.
That changes the economics of publishing
A website can produce the information from which an answer is derived without necessarily receiving the visit.
That matters because visits fund much of the open web.
Publishers sell advertising.
Businesses sell products.
Independent writers sell subscriptions.
Communities gain new members.
Creators build reputations.
If discovery increasingly ends inside the layer performing the recommendation or summary, the destination has to find another reason for users to arrive.
The web can become more mediated while becoming more convenient
Convenience is real.
A good recommendation saves time.
A good search summary answers a simple question immediately.
Automatic translation opens communities across languages.
AI tools allow people with limited technical skills to create things that previously required specialised software.
The cost is that another layer increasingly sits between the user and the person who originally created something.
The feed selects.
The search engine summarises.
The platform labels.
The recommendation model predicts.
The user sees the output of that chain.
Synthetic personalities make the question stranger
Online identity itself is becoming easier to manufacture.
A social-media account can use generated photographs, synthetic video, generated captions and an invented biography.
There is nothing inherently wrong with a fictional digital character.
Entertainment has always involved characters.
The distinction is whether the audience understands what it is seeing.
A virtual personality openly presented as artificial is different from an account deliberately encouraging followers to believe that a non-existent person is documenting a real life.
That becomes especially important when the persona recommends products, discusses politics or claims direct experience.
Platforms increasingly require synthetic-media disclosure
YouTube requires creators to disclose realistic content that has been meaningfully altered or generated when it makes a real person appear to do something they did not do, alters a real event or creates a realistic scene that never occurred.
TikTok similarly requires labels for realistic AI-generated content and uses both creator disclosures and technical detection.
These systems cannot guarantee every synthetic post will be labelled correctly.
They establish an increasingly important cultural expectation.
Artificial creation is not necessarily the problem.
Hidden artificial creation can be.
Provenance may matter more than trying to spot weird fingers
For several years, advice about generated imagery often focused on visual mistakes.
Count the fingers.
Look at the teeth.
Inspect background text.
Generation systems keep improving.
The more durable answer may be information about where the media came from.
C2PA Content Credentials provide a technical standard through which software can attach cryptographically signed information describing the creation and editing history of digital media.
TikTok joined the C2PA Steering Committee in 2026, and other major technology companies already support the standard.
Content Credentials are not a universal truth machine.
Legitimate content can lack them.
Screenshots and re-uploads can remove useful provenance.
But the direction is important.
The internet is beginning to treat origin as information that may need to travel with the file.
Bots talking to bots are not science fiction
Automation does not only create posts.
It can create interaction around them.
Systems can draft replies, operate accounts, respond to messages and schedule publication.
That does not mean every suspicious comment section is artificial.
It does mean that visible activity no longer guarantees an equivalent quantity of human intention behind it.
A hundred comments can still represent a hundred people.
They can also represent far fewer people using automation.
Without access to platform data, a user often cannot know which.
Advertising became part of ordinary online speech
The commercial internet has always contained advertising.
Modern creator culture can make the boundary between advertising and ordinary speech less visually obvious.
A sponsored product appears inside a creator's normal video.
An affiliate recommendation looks similar to an unpaid recommendation.
Shopping results sit beside informational results.
A platform inserts advertisements directly into the same feed as friends, journalists, musicians and strangers.
Disclosure rules help.
The deeper cultural change is that commercial persuasion no longer always arrives in a separate advertising break.
It inhabits the same visual language as everything else.
Originality itself has become something platforms need to measure
There is something historically strange about that sentence.
For most of human culture, originality was argued about by critics, audiences and other artists.
It is now also a ranking variable.
Platforms have to decide whether an account created a video, copied it, minimally modified it, reacted to it or added enough value for the new version to deserve distribution.
That decision can affect millions of views and real income.
Culture has not only become measurable.
Some of its most subjective qualities have become inputs to automated distribution systems.
Smaller communities can feel different because they optimise for something else
One response has been a return to smaller online spaces.
Private group chats.
Specialist forums.
Small Discord servers.
Subscriber communities.
Group conversations where people recognise one another's names.
These spaces are not automatically good.
They can become hostile, insular, badly moderated or confidently wrong.
What changes is the incentive structure.
A person speaking to thirty familiar members does not necessarily need to maximise distribution.
Reputation inside the community can matter more than reach outside it.
A post can be successful because five relevant people found it useful.
The direct relationship starts to matter again
Newsletters, subscriptions, podcasts and private communities all partly recreate an older internet relationship.
The audience chooses the destination.
The creator does not need every piece to win a recommendation lottery because some people intentionally return.
Recommendation platforms can still be how the audience discovers that creator in the first place.
The two systems are not enemies.
One is powerful for discovery.
The other can create continuity.
Nostalgia edits the old internet
It is easy to remember the earlier web as a beautiful collection of weird personal websites, forums and people making things because they cared.
Those things existed.
So did extraordinary amounts of spam.
Pop-up advertising.
Chain emails.
Scams.
Malware.
Harassment.
Terrible search manipulation.
Stolen content.
Low-quality websites created purely to capture advertising traffic.
A less sophisticated internet was not automatically a more sincere one.
Nostalgia becomes useful only when it identifies what structurally changed.
The biggest structural change may be abundance
Publishing used to require more friction.
A website required setup.
A video required editing.
Distribution required an audience or a publisher.
That friction excluded enormous numbers of talented people too.
Removing it has been liberating.
Generative AI removes another layer.
Producing a technically competent piece of text, imagery or audio can now take less time than deciding whether it should exist.
The scarce resource therefore moves.
It is no longer only production capacity.
It is attention.
The internet's old problem was not enough information. Its new problem is deciding what deserves a human being's finite attention.
Cheap content encourages proven formats
Originality is risky.
A creator can spend a week making something unusual and watch it fail.
A format that has already produced millions of views is easier to justify financially.
This logic existed long before generative AI.
AI makes variation cheaper.
The result does not have to be identical posts.
It can be an unlimited number of posts that share the same skeleton.
Change the celebrity.
Change the game.
Change the city.
Change the product.
Keep the structure.
Human-made may become a meaningful label
Mass production once increased the cultural value attached to things described as handmade.
Something similar may happen online.
People may increasingly care that an essay was actually written by somebody.
That a photograph documents a real moment.
That a song was performed by a particular musician.
That a restaurant review came from somebody who ate there.
That the person describing grief, migration, love, illness or embarrassment actually experienced what they are talking about.
This will not make generated culture worthless.
It may make origin part of the value.
Human intention and human labour are not the same thing
There is another distinction worth protecting.
A piece can involve extensive automation and still be deeply intentional.
A photographer can use software throughout the editing process.
A musician can use synthesised sounds.
A filmmaker can use generative imagery.
A writer can use translation or transcription tools.
The more meaningful question is sometimes not "Was AI used?"
It is "Did somebody care what the final thing became?"
That is why AI disclosure alone cannot solve the cultural problem
A label can tell users that a video was generated.
It cannot tell them whether it is good.
A human can make manipulative garbage without any AI.
An artist can use AI inside thoughtful, original work.
Authenticity is partly about provenance.
It is also about motive, responsibility and context.
No metadata standard can measure all of that.
Users still have more control than the infinite feed suggests
Algorithmic distribution can feel inevitable because it is the default on many services.
It is not the only way to use the internet.
People can intentionally follow smaller creators.
They can subscribe directly.
They can use non-personalised feeds where platforms provide them.
They can bookmark sites.
They can join communities where reach is not the main measure of success.
They can click through to the original reporting instead of relying only on summaries.
None of those actions dismantles the economics of the modern internet.
They change the small part of it an individual user inhabits.
A more intentional internet
Ways to make the web feel less like one endless recommendation system
- Follow people directly rather than relying entirely on suggested content.
- Use chronological or non-profiled feed options where available.
- Subscribe directly to writers, creators and publications you want to return to.
- Visit specialist forums and smaller communities built around recurring participants.
- Click through to original sources instead of treating summaries and reposts as the destination.
- Look for synthetic-media disclosure and provenance information when authenticity matters.
- Remember that popularity measures distribution, not necessarily originality or truth.
The internet may split into two cultural layers
The future does not have to move in one direction.
Large public feeds may become more automated, personalised and synthetic.
Smaller spaces may become more intentionally human precisely because the larger environment is not.
People can move between both.
A person might spend half an hour consuming a personalised stream selected by machine learning and then spend the evening talking to ten people they have known online for six years.
Both experiences are the internet.
The internet is not dead
The phrase "dead internet" is tempting because it turns a complicated cultural shift into a dramatic conclusion.
It is also wrong as a literal description.
Real people remain everywhere online.
They are making extraordinary things.
They are discovering communities that would have been impossible without recommendation systems.
They are using AI to make work that would otherwise have been beyond their technical or financial reach.
They are also operating inside systems containing much more automation, optimisation and commercially motivated content than before.
Both realities can exist at once.
What people may actually be missing
When somebody says the internet feels less human, they probably do not mean they want dial-up connections, broken websites and videos that take twenty minutes to download.
They may be describing the disappearance of visible intention.
Somebody had to care enough to build the strange little website.
Somebody had to return to the same forum.
Somebody had to write the post because they wanted that specific post to exist.
Modern technology removes friction because friction is inconvenient.
That is usually progress.
But some friction also acted as evidence that a person had made a choice.
The challenge is not rebuilding the old web
The older internet cannot simply be recreated.
Nor should every part of it be.
The modern web is more accessible, more capable and capable of connecting creators with audiences at a scale the early web could never have managed.
Generative technology will not disappear.
Recommendation systems will not disappear.
Advertising will not disappear.
The more realistic challenge is building signals of human intention into a system where production and distribution can increasingly happen automatically.
That may mean provenance.
It may mean stronger originality rules.
It may mean users having more control over recommendation systems.
It may mean direct subscriptions and smaller communities becoming more valuable.
And it may simply mean learning to ask a question that used to feel unnecessary.
Who wanted this to exist?
Reporting note
Sources used for this article
This article draws on Pew Research Center's 2025 and August 2026 analyses of AI-generated web content and search behaviour; current YouTube policies on inauthentic, repetitive and synthetically generated material; Meta's 2026 reporting on original-content recommendations and ranking changes; TikTok's 2025 and 2026 transparency measures for AI-generated content; the European Commission's Digital Services Act guidance and 2026 enforcement work concerning recommender systems and addictive platform design; and current C2PA guidance on Content Credentials. Platform performance figures are identified as company-reported where applicable, and AI-authorship estimates are treated as detector-based estimates rather than perfect identification.
If you believe this article contains a factual error, visit our corrections page .
