The Firewall Works. Good Luck Leaving.

We can block malicious websites, secure our networks, and reroute traffic when the internet fails. But we can't tell a social platform what to recommend next. And leaving isn't so simple when the people we came to see are still there.

In 2024, researchers at Dublin City University set up ten smartphones.

Five for TikTok. Five for YouTube Shorts.

They created accounts representing teenage boys and young men. Some would explore interests such as football, gaming, fitness, and comedy. Others would deliberately seek out Andrew Tate, red-pill philosophy, and the broader collection of influencers promising to explain everything that's wrong with modern masculinity.

Two accounts would remain neutral controls.

The researchers wanted to know what the platforms would recommend.

On YouTube Shorts, an account configured to represent an ordinary 18-year-old male encountered manosphere content after two minutes and one second.

The interesting part?

The researchers hadn't applied their first search prompt.

The account hadn't been deliberately steered toward football, gym tips, Andrew Tate, or anything else on its assigned list of interests.

The recommendation appeared during the initial viewing round.

The researchers continued.

They watched hundreds of videos, introduced search terms, and interacted with the results. When problematic material appeared, they watched and liked it.

The feeds responded.

By the final round, after roughly 400 videos per account, the researchers found that an average of 76% of recommendations on TikTok and 78% on YouTube Shorts fell into their broader category of toxic content. That included misogyny, rigid alpha-male ideology, reactionary politics, and conspiracy narratives.

The numbers were striking.

They also need context.

This was a small experiment involving artificial accounts, not a representative sample of actual teenagers. The researchers deliberately reinforced problematic recommendations. Even the generic-interest accounts received search prompts, some of which concerned sexual attraction and men's mental health.

Those final percentages don't describe what would happen to someone passively watching random videos.

And the experiment didn't establish that any person had been persuaded to believe anything.

The Dublin City University study measured what the platforms recommended under specific conditions.

That was interesting enough.

Nobody broke into YouTube. There was no compromised password, malicious code, or previously undiscovered security vulnerability.

The researchers used the applications.

The applications observed their behavior.

And the recommendations changed.

We don't know precisely why that first video appeared. The platform's internal selection process wasn't available for independent inspection.

Which raises a question considerably larger than Andrew Tate, misogyny, or the latest collection of men explaining how to become millionaires from the driver's seat of somebody else's Lamborghini.

We have spent decades developing extraordinary tools for securing digital networks.

We can control which computers communicate, which applications reach the internet, which websites are accessible, and even which internet connection a particular device should use.

But once somebody opens a social-media application, many of the most consequential decisions move inside systems they don't own.

Of course, they can always leave.

Except their friends are there. Their communities are there. Their favorite musicians, local organizations, and the people they haven't seen in twenty years are there.

The service is doing something valuable.

That's part of the problem.

The Internet Didn't Start With a Feed

The World Wide Web began as a way to share information.

Researchers at CERN needed to exchange documents across different computers and institutions. Tim Berners-Lee developed a system built around hyperlinks and common technical standards.

In 1993, CERN released the Web software into the public domain.

Anybody could publish a website. Somebody else could build a browser. A third person could create a directory to help people find things.

These activities didn't have to belong to the same company.

The early internet wasn't a particularly civilized place.

It had scams, pornography, misinformation, obnoxious advertising, and people who discovered that arguing with strangers was considerably more entertaining than doing their jobs.

But its architecture left room for different ways of participating.

Then the internet got crowded.

Search engines helped organize websites. Social networks helped people maintain relationships. Recommendation systems helped them discover things they might otherwise miss.

Useful developments.

Facebook introduced News Feed in 2006. Instead of visiting individual profiles to see what friends were doing, users could receive updates in one place.

A sensible improvement.

As networks grew, presenting everything chronologically became increasingly difficult.

Someone with hundreds of connections could never reasonably read every update, and most updates weren't equally interesting.

So platforms began deciding what deserved attention.

Which posts were relevant?

Which people mattered most to a particular user?

What might somebody enjoy that they hadn't thought to search for?

The answers could be genuinely useful.

People discovered musicians, educational material, communities, businesses, and opportunities that would otherwise have remained invisible.

In 2012, YouTube explained that it was changing how videos were recommended.

The service had been rewarding clicks.

That created an obvious problem.

A misleading thumbnail could persuade somebody to open a video, disappoint them almost immediately, and still receive credit for attracting the viewer.

YouTube wanted something better.

It began emphasizing watch time.

If people stayed to watch a video, that was a potentially useful indication that they valued it.

Fair enough.

Better recommendations improved discovery. Creators who made engaging videos could reach more people. YouTube benefited when viewers continued watching.

Everyone had a reasonable incentive.

Then creators began organizing their work around the measurements.

Titles changed.

Thumbnails changed.

Videos changed.

People experimented with pacing, emotional intensity, controversy, and the moment a viewer might lose interest.

The platform measured the response.

Creators adapted.

The platform adapted again.

By 2016, Google's researchers were publishing details of YouTube's large-scale neural recommendation architecture, capable of identifying and ranking videos from an enormous collection of possibilities.

Machine learning made this process increasingly sophisticated.

Short-form video removed even more of the work involved in choosing what to watch.

Open the application.

Watch.

Swipe.

Repeat.

The platform solved a real problem. There was too much information for people to sort through themselves.

And it did more than solve that problem.

It made enormous libraries of content accessible through remarkably simple interfaces. Someone with an inexpensive phone could find a university lecture, learn to repair an engine, or discover a musician on the other side of the planet.

Centralization helped make those experiences possible.

Large platforms could invest in infrastructure, search, moderation, security, and discovery at scales that smaller services often couldn't match.

People didn't adopt these systems because they were foolish.

They adopted them because they worked.

But as the platforms became central to discovery, creators became increasingly dependent on the systems deciding what deserved to be discovered.

The recommendation system didn't need an ideology.

The people competing for its attention had plenty.

And the more successfully a platform connected people, the more costly it could become to leave.

A social network's greatest asset isn't necessarily its software.

It's everyone who uses it.

A Business That Understands You

The manosphere is an unusually revealing example of what happens when social needs, commercial incentives, and recommendation systems begin reinforcing one another.

It's also widely misunderstood.

There isn't one manosphere organization with a headquarters and a coordinated mission.

The term describes overlapping communities involving anti-feminist and male-supremacist ideologies, along with adjacent audiences for fitness, dating advice, entrepreneurship, appearance optimization, and political commentary.

Those adjacent interests are not inherently part of an extremist movement.

A man learning to lift weights isn't on a political trajectory merely because a misogynistic influencer also recommends exercise.

That would be an impressive misunderstanding of both exercise and politics.

Young men have legitimate reasons to seek advice.

They want relationships. They want jobs that pay enough to build a life. They want confidence, friendship, financial security, and some indication that they're heading in the right direction.

Some are lonely.

Some are angry.

Some are simply curious.

And some online creators provide genuinely useful information.

A fitness instructor can explain exercise better than an uninformed friend. A skilled tradesperson can demonstrate an unfamiliar technique. A thoughtful adviser can help somebody navigate relationships.

The technology makes those connections possible.

But imagine a teenager searching for advice because dating has become confusing.

He might find someone explaining communication and emotional maturity.

He might find a comedian making jokes about rejection.

Or he might find somebody explaining that his frustration is evidence of a society organized against men, that women are fundamentally untrustworthy, and that recovering his rightful position requires adopting a particular worldview.

Possibly available through a monthly subscription.

The last explanation has certain advantages.

It gives a complicated problem a simple cause.

It identifies an enemy.

It offers membership in a community that appears to understand the frustration.

And it promises that following the right instructions will produce a different life.

Being understood feels good.

Especially when the rest of life feels difficult to interpret.

Hours of intimate conversation with an influencer can create the feeling of knowing that person. The relationship is one-sided, but the familiarity can be meaningful to the viewer.

The influencer becomes part adviser, part entertainer, part companion.

Eventually, the distinction between following somebody and belonging to their community becomes less obvious.

And a community is considerably more valuable than an isolated video.

It gives people a reason to return.

The commercial opportunities follow.

Courses, memberships, sponsorships, gambling promotions, supplements, and other products can all generate revenue.

Some are ordinary businesses. Others have attracted allegations of deception or exploitation.

The research doesn't establish a universal sequence in which commercial motives create ideological beliefs, or ideological beliefs inevitably produce commercial schemes.

Different people follow different paths.

But the feedback loop is easy to recognize.

A creator offers an explanation.

An audience responds.

The platform measures that response.

The creator discovers which explanations attract attention.

Revenue makes it possible to produce more.

The audience becomes more familiar with the creator.

And the creator learns what that audience wants to hear.

Nobody has to design the entire ecosystem.

Each participant can make decisions that seem reasonable from where they're standing.

The result can still reward increasingly distorted explanations of the world.

The Crowd May Have Been Hired

There is another part of this economy that deserves more attention.

Paid clipping.

A creator produces a long livestream. Other people cut it into short videos and distribute them across TikTok, YouTube Shorts, Instagram, and similar platforms.

Some do it because they're fans.

Others participate in commercial campaigns.

Content Rewards, which developed through the Whop ecosystem and now operates its own campaign platform while using Whop for payment processing, provides one example.

Businesses can establish campaigns in which participants receive compensation for approved promotional videos, including models that pay according to verified views.

There are legitimate reasons to do this.

A band could encourage people to distribute concert clips. A filmmaker could promote a trailer. A small business could pay creators to introduce a product.

Now consider the same model applied to someone whose livelihood depends on controversy.

One confrontational livestream becomes dozens of clips.

Those clips appear on different accounts.

Some distributors are independent fans. Others may be compensated. Algorithms can amplify both.

A viewer opens an application and sees the same personality repeatedly.

It feels as though everybody is suddenly talking about him.

Maybe they are.

Or maybe some of that apparent momentum was purchased.

The two aren't mutually exclusive.

What matters is the difference between organic enthusiasm and commercial distribution that looks organic.

Paid promotion isn't inherently deceptive. But financial relationships can become material information when a promotional message appears to be an independent recommendation.

The Federal Trade Commission's endorsement guidance addresses precisely that problem.

Clipping marketplaces demonstrate a distribution mechanism. Their existence does not prove that any particular influencer used one, or that a campaign deliberately promoted misogynistic content.

Those claims require evidence about specific campaigns.

Still, the arrangement raises useful questions.

How much apparent popularity comes from independent sharing?

How much comes from compensated distribution?

Can a viewer identify the difference?

And how do recommendation systems respond when many accounts distribute substantially similar material?

Those are empirical questions.

The data needed to investigate them isn't equally available to everyone.

Then there is the question of bots and coordinated influence operations.

Google's Threat Analysis Group bulletin for late 2025 documents terminated YouTube networks associated with coordinated influence campaigns linked to Russia, China, and other countries. Some operations were connected to identified organizations, including consulting firms.

That establishes the reality of organized influence activity on major platforms.

It does not establish that foreign governments secretly control the manosphere.

A campaign linked to a country isn't automatically directed by that country's government. Paid human distributors aren't necessarily bots. Coordinated commercial promotion isn't the same thing as covert political influence.

And a sudden increase in popularity isn't proof that anybody manipulated it.

An unexplained pattern appears.

Somebody offers a compelling theory.

The theory spreads faster than the evidence.

We've seen that movie before.

The useful questions concern identifiable behavior: financial relationships, account coordination, amplification patterns, undisclosed sponsorship, and the circumstances under which a platform chooses to recommend material.

Platforms possess considerable information about these activities.

Outside researchers and users generally have less.

Which means a person can encounter the appearance of widespread enthusiasm without reliably knowing how it was produced.

It Might Not Be the Algorithm

There is a popular story about online radicalization.

A young person watches something harmless.

An algorithm recommends something questionable.

A few recommendations later, the person has become an extremist.

It's a compelling narrative.

The evidence is considerably less cooperative.

In a 2024 study published in PNAS, researchers examined YouTube recommendations using experimental bots and real viewing histories.

On average, bots following recommendations consumed less partisan content than the real users whose behavior they were modeled on.

Other research has found that viewing of extremist material is concentrated among people already inclined toward those ideas, often reached through subscriptions or links from outside the platform.

That complicates the claim that recommendations universally push unsuspecting people toward extremism.

It also doesn't invalidate the Dublin City University findings.

The studies examined different formats, populations, content categories, and outcomes.

One studied political partisanship in long-form YouTube viewing.

The other examined short-form recommendations and content associated with misogyny, rigid masculinity, reactionary politics, and conspiracy narratives.

They aren't interchangeable experiments.

And encountering a video is not the same thing as accepting its claims.

People have agency.

They seek out information. They reject arguments. They change their minds. They deliberately pursue ideas that interest them.

Sometimes their beliefs were established long before they opened the application.

Technology doesn't erase any of that.

But it can change the environment in which those choices happen.

A recommendation system can influence what becomes visible, what recurs, and how easily somebody encounters competing explanations.

Those effects should be measured rather than assumed.

There is a difference between exposure, persuasion, and behavior.

The research needs to distinguish them.

The fact that we have competing studies is not a reason to abandon the question.

It's a reason to investigate it more carefully.

And perhaps to reconsider why so much of the information needed to study these systems remains concentrated within the companies operating them.

The Government's Favorite Setting

Governments have begun intervening.

Understandably.

Parents want children protected. Regulators want enforceable rules. Politicians want to demonstrate that they're responding to problems people can see.

Unfortunately, complex technical systems don't become simple because a legislature needs something it can put into a statute.

Australia began enforcing social-media account restrictions for children under sixteen in December 2025.

The objective was to reduce exposure to potentially harmful online environments.

In July 2026, Australia's eSafety Commissioner released an early evaluation.

The proportion of children aged ten to fifteen reporting ownership of social-media accounts fell from 52.4% to 42.1%.

A statistically significant reduction.

But the proportion reporting any social-media use declined from 85.9% to 81.5%.

Most were still using social media.

The evaluation identified some encouraging developments, including fewer children reporting that they felt left out because they weren't using it.

It also found problems with age-assurance enforcement and changes in parental awareness.

Three months isn't enough to determine the law's long-term effects on wellbeing or exposure to harm.

But the early results reveal an important limitation.

An account is not an experience.

Reducing account ownership doesn't necessarily produce an equivalent reduction in use.

And reducing use doesn't automatically establish which harms have been prevented.

Governments prefer requirements they can measure.

Platforms prefer compliance they can demonstrate.

Parents want children to be safe.

Each objective makes sense.

The problem arises when the measurement becomes a substitute for the goal.

The European Union has taken a more technically ambitious approach.

Its Digital Services Act addresses platform transparency, systemic risks, and researcher access.

Under Article 38, very large platforms and search engines using recommendation systems must provide at least one option that isn't based on profiling.

The European Commission has also investigated features such as infinite scrolling, autoplay, and highly personalized recommendations.

In 2026, it issued preliminary findings concerning TikTok and, separately, Facebook and Instagram.

These are preliminary regulatory findings, not final judgments.

But the approach is significant.

Instead of focusing exclusively on which accounts should be restricted or which individual posts should be removed, regulators are examining how the products operate.

That's closer to the actual machinery.

It also creates risks.

Regulation can become an expensive compliance exercise. Large companies may absorb those costs more easily than smaller competitors.

Governments can acquire powers over communication that extend well beyond the problems those powers were introduced to address.

And an unpopular opinion isn't a computer virus.

Disagreement about politics, religion, or gender roles isn't evidence that a network has been compromised.

Fraud, unlawful threats, impersonation, and deceptive commercial activity raise different questions.

Treating every undesirable idea as a form of digital danger would make governing the internet easier.

It would also give someone far too much authority.

The challenge isn't simply deciding which material should disappear.

It's determining which mechanisms produce identifiable harms, which interventions actually address those harms, and who gets to control the interventions.

The people making today's rules won't necessarily be the people administering them tomorrow.

The Network Is Secure. The Feed Isn't Yours.

Consider an ordinary home.

There are laptops, phones, televisions, tablets, game consoles, smart appliances, and possibly a camera keeping watch over a driveway that has remained largely unchanged since yesterday.

Somebody is working.

Somebody is watching a movie.

A child is doing homework, or at least has the homework open in another tab.

Everything connects to the internet.

Businesses spend considerable money establishing policies around environments like these.

Fortinet offers firewall and network-security systems. Arista provides enterprise networking and related security capabilities. Dell Technologies supplies other parts of enterprise computing, including endpoints and infrastructure.

Open-source software such as OPNsense can provide sophisticated routing and firewall functions without requiring a conventional enterprise security subscription.

These technologies serve different purposes.

But they operate within a familiar engineering discipline.

Decide what the system should allow.

Establish where a rule can be enforced.

Verify that it works.

You can separate untrusted devices from sensitive computers.

Restrict access to known malicious domains.

Apply different policies to different network segments.

Monitor suspicious connections.

You can even connect two internet providers and configure traffic to move when one connection fails.

Cable internet stops working?

With the appropriate equipment and configuration, traffic can move to a Starlink backup.

We've developed remarkably sophisticated ways of managing the movement of information.

Now imagine trying to establish a different rule.

A teenager wants to learn about weightlifting.

The family has no objection to training advice, exercise videos, or the enormous amount of educational material available on YouTube.

But they don't want every visit to the platform to become an increasingly repetitive collection of hostile explanations about women, relationships, and society.

They don't want to block YouTube.

They want to control the recommendations.

That turns out to be a different problem.

A firewall can restrict access to YouTube.

It generally cannot inspect the meaning of each individual video inside YouTube's encrypted service, determine why it was selected, and replace the next recommendation with something else.

DNS filtering can identify domains.

Application controls can recognize and restrict services.

Device-management systems can impose schedules and access policies.

But none of those functions provides a household with independent authority over YouTube's internal recommendation engine.

The relevant decisions happen inside the platform.

And the same applies to many other services.

Of course, there are existing controls.

YouTube allows people to manage watch history, provide recommendation feedback, avoid certain channels, and use supervised experiences. In 2026, it expanded parental controls over teenagers' Shorts viewing.

These are useful features.

A parent can establish boundaries around time and access.

But limiting how long somebody spends inside an application is different from controlling how the application selects information during that time.

Installing a more expensive firewall won't automatically solve that problem.

The firewall may be doing its job extremely well.

The recommendations aren't being made there.

This is the technical gap beneath much of the cultural debate.

We have mature tools for governing connections, devices, identities, and access to digital resources.

We have fewer independent tools for governing how information is selected within the applications those resources support.

The technology industry knows how to enforce policies.

But user-controlled recommendation policy hasn't become an equally mature feature of the consumer internet.

And even if somebody dislikes the recommendations enough to leave, there's another problem waiting.

You Can Leave. Your Friends Are Still There.

Suppose you decide that a particular social platform no longer deserves your attention.

You don't like the recommendations.

You don't trust its leadership.

You think another service would provide a better experience.

So you leave.

Technically, this is straightforward.

Socially, it can be expensive.

Your friends are still there.

Your family is still there.

The venue announcing concerts is there.

The musicians you follow, local businesses, neighborhood organizations, and the person you haven't spoken to since high school are all there.

Leaving the application may mean leaving part of the community.

This is why telling people to use another platform doesn't solve the underlying problem.

A social network isn't simply a collection of software features.

It's a collection of relationships.

And relationships aren't as portable as software.

Centralized platforms have considerable advantages here.

A large service can offer sophisticated search and discovery, reliable infrastructure, integrated messaging, spam prevention, moderation, and tools for creators and advertisers.

Many people don't want to configure recommendation providers, maintain subscriptions, or investigate the administrative structure of the services they use.

They want to see their friends.

Preferably without having to become part-time systems administrators.

That's reasonable.

And it's one reason decentralized alternatives face such a difficult challenge.

Being technically superior at one task doesn't guarantee that a social network will attract the people needed to make it valuable.

There is also an important economic consequence.

A company whose users can easily move their identities, relationships, and audiences to competitors faces a different competitive environment from a company whose users must rebuild those connections from scratch.

If switching becomes less costly, service quality may matter more.

If switching means abandoning a community, people may tolerate experiences they would otherwise reject.

That doesn't prove any particular platform deliberately makes its recommendations unpleasant.

It explains why the ability to leave matters even when a service is useful.

And it is where alternative network architectures become interesting.

What If the Network Didn't Belong to One Company?

Consider email.

Someone using Gmail can communicate with someone using Proton Mail.

They don't need to agree on one provider.

Their providers operate different systems using compatible standards.

Social networking doesn't have to work exactly like email, but the principle is useful.

The relationships shouldn't necessarily depend on everyone accepting one company's application.

The Fediverse attempts something similar.

Services such as Mastodon, Pixelfed, and PeerTube use federated technologies that allow independently operated systems to communicate.

One important standard is ActivityPub, published by the World Wide Web Consortium in 2018.

A community can operate its own server. Another can use a different provider. Participating systems can exchange information without requiring one corporation to own every account.

There are limitations.

Moderation requires labor. Servers cost money. Abuse doesn't disappear because the software is open source.

Communities can refuse to communicate with one another. Moving accounts and preserving relationships isn't always seamless.

Decentralization creates more places where decisions can be made.

It doesn't guarantee good decisions.

Bluesky offers a different approach.

Its AT Protocol separates important elements of social networking, including identity, personal data, feed generation, and moderation services.

Users can choose from feeds created by different providers.

Instead of accepting one company's universal ranking system, they can select alternatives that organize information differently.

Someone could follow a chronological feed.

Another person could choose a feed devoted to a particular community or subject.

An independent developer could offer a different recommendation method.

The architecture also supports independent labeling services that applications can use for moderation.

These capabilities don't make Bluesky completely decentralized in practice. Shared infrastructure still matters, and some services may become concentrated.

Nor are AT Protocol and ActivityPub interchangeable.

But both demonstrate that hosting a social relationship and controlling the information presented around it don't necessarily have to belong to the same company.

A person can choose a feed without necessarily changing every other part of their social experience.

That's a meaningful shift.

There is also a much older technology worth remembering.

RSS.

You subscribe to a website.

It publishes something.

Your reader retrieves the update.

No elaborate prediction system is needed to decide that you might find the material interesting.

You already did.

RSS has limitations. Discovering reliable sources requires effort. Managing subscriptions can become tedious. And people can deliberately subscribe to misinformation just as easily as excellent reporting.

But it illustrates a principle that has become strangely unfashionable.

Sometimes people want to see information because they asked to see it.

These alternatives aren't going to replace every major platform tomorrow.

And they won't automatically solve the recommendation problems inside services that remain closed.

The more useful long-term objective is making platform choice less consequential to maintaining a social identity, audience, or community.

A musician shouldn't have to rebuild a following from nothing every time a platform changes its policies.

A neighborhood organization shouldn't have to depend entirely on one proprietary feed to reach its members.

A person shouldn't lose their social relationships simply because they prefer a different application.

The most important improvement may not be persuading everyone to move to one better platform.

It may be reducing the cost of moving at all.

Who Gets to Write the Rules?

Imagine that choosing how information reaches you became as ordinary as choosing which email provider to use.

You might want posts from people you follow in chronological order.

Someone else might prefer recommendations organized around music, local events, or subjects they want to learn about.

A third person might want a service that clearly identifies commercial promotions or prioritizes reporting from particular sources.

These are different preferences.

They don't require one universally correct feed.

The objective would be to make recommendation choices understandable, meaningful, and portable wherever the underlying technology permits.

Some of these capabilities already exist.

Others would require new standards, platform interfaces, commercial agreements, or regulatory obligations.

A home firewall can't simply impose its own recommendations inside TikTok.

An independent service can't automatically rewrite every closed application's feed.

And moving social relationships between networks raises difficult questions about identity, privacy, moderation, and consent.

But those difficulties are engineering and governance problems worth addressing.

We already know that independently operated recommendation services are possible.

We know that digital identities can be separated from certain hosting arrangements.

We know that commercial relationships can be disclosed.

And we know that researchers can investigate recommendation behavior when they have appropriate access.

A more user-directed internet would combine those capabilities without pretending that every information problem is a security problem.

There is also a useful role for AI.

An assistant could help someone evaluate an unfamiliar claim, identify a disclosed sponsorship, compare competing explanations, or understand which settings affect their recommendations.

Some of that work could happen locally, reducing the need to transmit private activity to another service.

But an AI assistant shouldn't quietly become an authority over which opinions people are allowed to encounter.

Models can make mistakes.

They can misunderstand satire, misclassify unusual opinions, and reflect the assumptions of the organizations that develop them.

A tool that helps someone evaluate an argument is different from one that prevents them from hearing it.

That distinction becomes especially important inside families.

A young child and a teenager approaching adulthood need different boundaries.

Protecting children shouldn't require preserving permanent surveillance over their developing judgment.

And the same monitoring capabilities that could help a parent protect a child could be abused by an employer, government, or controlling partner.

Any serious system of digital policy needs to address the possibility that the person setting the rules is the problem.

Who has authority?

Who can inspect what happened?

What information is collected?

Can somebody challenge a mistake?

Can they leave?

A technically effective system can still be a terrible system to live under.

These questions aren't reasons to reject digital protections.

They're reasons to design them around the people affected rather than merely the people operating them.

The Part Nobody Designed

The strange thing about the modern internet is how understandable most of the decisions behind it were.

Engineers wanted better recommendations.

Creators wanted audiences.

Advertisers wanted customers.

People wanted entertainment, information, and community.

Parents wanted children protected.

Governments wanted rules they could enforce.

None of those objectives is inherently unreasonable.

Together, they helped produce an environment in which enormous influence over information discovery is concentrated inside a relatively small number of private systems.

The companies operating those systems can observe user behavior in extraordinary detail.

The people using them have much less independent access to the processes deciding what they see.

And leaving can mean abandoning something genuinely valuable.

That isn't proof of malicious intent.

It's what happens when technology, incentives, and human relationships become intertwined.

The manosphere makes some of these mechanisms easier to notice because its commercial promises can be so conspicuous and its most notorious personalities so provocative.

But the same questions appear elsewhere.

Someone researching investments can encounter financial misinformation.

A person looking for medical advice can find excellent education alongside unsupported treatments and people selling certainty by the bottle.

Political communities can reward increasingly dramatic accounts of their opponents.

The mechanisms aren't identical, and the outcomes can't be assumed.

But the larger problem extends well beyond any one ideology.

It concerns who determines what becomes visible, how that selection can be examined, and what choices remain available to the people receiving it.

Removing one influencer doesn't resolve those questions.

Neither does restricting one category of account.

Nor does installing a better firewall.

There isn't a single product or policy that fixes the entire system.

But there are meaningful changes worth pursuing.

Give people more independent choices over how information is ranked.

Make social relationships more portable across compatible services.

Allow researchers to test the claims platforms make about their own behavior.

Enforce rules against identifiable deception without turning unpopular opinions into security threats.

And preserve the ability to reject the recommendations of both corporations and governments.

The objective isn't to build an internet where nobody encounters bad ideas.

It's to build one where people have more meaningful control over the systems selecting which ideas reach them.

That requires something more demanding than another safety warning.

It requires changing who gets to make certain decisions.

And it requires acknowledging why people accepted the current arrangement in the first place.

The services worked.

They helped us find one another.

The difficulty is that finding one another made us dependent on the places where we met.

Good Luck Leaving

Return to the Dublin City University experiment.

Ten phones.

A collection of experimental accounts.

Recommendation systems observing behavior and selecting what appeared next.

The study couldn't tell us what actual teenagers would come to believe.

It could show what those platforms delivered under the conditions the researchers created.

And it exposed something that remains difficult for ordinary users to examine.

The company operating a recommendation system can study the machinery from the inside.

Outside researchers often have to reconstruct its behavior through experiments.

The person watching receives the finished result.

Now think about the devices in your house.

The router is working.

The firewall is enforcing its policies.

The internet connection is stable.

Maybe there's even a backup ready to take over if the first connection fails.

The network is functioning exactly as intended.

Someone opens an application to watch a video.

They enjoy it.

Another appears.

They watch that one, too.

Then another.

At some point, the original reason for opening the application may be long forgotten.

Nothing has to be wrong with any individual video.

Nobody has to be conspiring against the viewer.

The recommendation system is simply selecting what it predicts will hold that person's interest.

The router delivers every recommendation perfectly.

It doesn't know why those videos were selected.

And the person watching has little independent way to find out.

They could close the application.

They could choose another service.

They know that.

But the conversation they came for is still happening inside.

And so are the people they came to see.