The human brain is as reward-hackable as any other system.

On the (F)utility of Social Media Bans

Something happened to the generation which grew up inside the feed. The cohort that hit adolescence after roughly 2017 is more anxious, more depressed, lonelier, and worse at sleeping and sitting still than any generation we have measured, and the lines might bend hardest for teenage girls. You can argue about the size of the effect, and careful people do; the causal evidence is thinner than the headlines pretend, and there is a long tradition of blaming whatever medium happens to be new for whatever is wrong with the young, running from the novel through the radio to the video game. But the trend is real enough, and it arrived on a tight enough schedule, that governments have stopped waiting for the statisticians to settle it.

The schedule is the part worth holding onto. This is not the slow forty-year drift of television. The break is sharp and recent. It falls in the years around 2017 to 2019, and then again across the long indoor stretch of the pandemic, when systems that had been merely popular became the default setting of an entire generation's social life. Whatever did this did it quickly, and did it then.

Any explanation that cannot account for the timing is not an explanation. It is a mood.

The explanation on offer does not clear that bar. It is the intuitive one: phones, screens, social media. The kid is on the app too much, so take away the app. This is the reasoning behind nearly every bill now moving through many Western legislatures, and it is not so much wrong as imprecise, in the particular way that imprecision hardens into a policy that cannot work.

Australia has gone first. Since late 2025 it has been illegal for the major platforms to let anyone under sixteen keep an account. The law puts the burden on the companies rather than on parents or children, backs it with fines in the tens of millions, and within weeks of taking effect it had stripped several million accounts off the books. Other governments are drafting their own versions and watching this one for cracks: the United Kingdom, a row of European capitals, others further afield. The polling runs lopsidedly in favor. The logic is bipartisan and easy to say to a frightened parent in a single sentence. The thing hurts children, so children may not have the thing.

I want to be fair to this, because it is neither stupid nor cynical (it however does ignore how easy platform bans are circumvented by even children). And there is something almost bracing about a government willing to move at all against products this profitable and this woven into ordinary life. But it has aimed at the wrong target, and the miss is not a near miss. It has banned a place, when the thing worth regulating is a process. A platform is a place. The mechanism that runs inside it is a process, and the process does not care what the place is called or whether it is still standing.

Here is the distinction the legislation never draws, and it is the whole argument.

A decade ago, what you saw was mostly a function of whom you chose to follow. You picked your friends, your YouTubers, a few influencers, and the system showed you what they posted, more or less in the order they posted it. The feed was downstream of your choices. That feed is gone. Follows and subscriptions still exist, but they have been demoted to a garnish at the edge of a screen whose center is now decided by something else. On the surface that something is called a For You page. Underneath it is an optimiser. The system is handed a number to maximise. Call it engagement; in practice it is some weighted proxy for the time and attention you surrender. Then it is pointed at a firehose of candidate content and left to learn, from your every flick and pause and rewatch, which items raise that number for someone like you, and increasingly for you specifically. Nobody sits in a room choosing to exploit your insecurities. They do not have to.

Gradient descent does not require a motive.

The model simply discovers, across billions of sessions, a fact that no one needed to write down: your weaknesses hold your attention more reliably than your wants. Envy keeps you scrolling longer than contentment. Outrage outperforms calm. Self-doubt converts. So that is what gets served, not out of malice but out of arithmetic, because the objective was attention and that is where the attention turned out to be.

What makes this different in kind from every editor and programmer who came before is the loop. A newspaper editor also chose what you saw and chose it partly to keep you reading, but he was guessing at a crowd, printing once, and waiting weeks for sales figures to tell him, coarsely, whether he had guessed right. The modern system is not guessing at a crowd and it is not waiting. It is tuned to you, it is corrected in milliseconds, and the signal correcting it is not your considered judgment but your involuntary behavior, the half-second your thumb hovered, the autoplay you failed to stop. It optimises against the part of you that reacts before you decide. You are not the user of this system in any sense your grandparents would recognise. You are the environment it is searching. You do not scroll the feed so much as the feed runs experiments on you.

Once you see the mechanism clearly, the famous harms stop looking like separate evils and start looking like its direct outputs. Take comparison, the one everyone names. You could always measure yourself against the prettier neighbor or the face on the magazine, and people did, and survived it. But there was one neighbor and one magazine, and they did not adapt. The optimiser sifts millions of lives to find the one most efficient at making you feel insufficient, watches you linger on the wound, and brings you more, refined each pass. The comparison is no longer something you stumble into; it is manufactured to your measurements and restocked on demand. Take sleep. The old feed had a last page: you reached the end of what the people you knew had done that day, and the simple exhaustion of the material let you go. The new feed has no last page, because it draws not from a bounded set of people but from an endless library reordered forever, always one more, always a degree better calibrated than the last. A medium that ends releases you; a medium that cannot end keeps you. Take attention, the inability to sit with anything slow or long. That is what thousands of hours of variable, perfectly-timed reward does to anyone: the most effective schedule for compulsion we have ever found, wired to a generation's reflexes and called a feature. This is the thesis, and it is narrow on purpose. The harm is not screens, which are innocent glass. It is not social media in the abstract, a category broad enough to include a group chat and a recipe forum.

It is this specific mechanism: personalised, engagement-maximising, closed-loop recommendation, the version that took over around 2017 and swallowed the rest of the decade.

If the harm were the platform, you could ban the platform and be done. But the harm is the loop, and the loop is portable. It is already migrating into every product with a feed and a watch-time target. Ban the place and the process simply opens an account somewhere new.

Which is why the bans misfire, and misfire twice. They evict the mechanism from one address while it walks to the next, and they charge the public a strange price for the eviction. To keep a sixteen-year-old out, you have to establish that everyone else is not sixteen, and a wall that admits only adults is a wall that asks every adult to prove it. Build enough of those walls and you have quietly assembled a standing apparatus of age-checking laid across ordinary online life, which is a real cost, paid by everyone, in exchange for a barrier the mechanism routes around. This is the joke buried in the title. The bans are close to futile where it counts, because the thing that does the damage survives them, and they are effective only where it does not count, because closing accounts and erecting checkpoints is precisely the kind of action that can be performed, and photographed, and announced. That is the whole of their utility.

So aim at the process. The remedy follows the diagnosis instead of the headline, and it is less dramatic than a ban, which is usually the sign of a better idea. Do not outlaw the platform. Break the loop. Require that the sequence of items shown to any one person be shared by some large number of others, tens or hundreds of thousands within a country, so that the model is permitted to know your genre but never your name, permitted to learn that you like football but never to learn the exact shape of your particular hunger and feed it back to you. This does not abolish recommendation; you would still get sport if you wanted sport. It abolishes the closed loop, the tightening spiral in which your own live reactions are used to sharpen the next thing aimed at you. The genre survives. The thing hunting you, individually, does not.

Now the obvious objection, the one I think the essay has to meet head-on rather than dodge. If this mechanism harms everyone, and I believe it harms everyone, why protect only children? Why not switch it off for all of us?

Because we do not ban what a competent adult chooses to do to himself, and we are right not to. People may drink, may smoke, may lose a whole evening to a machine engineered to take it, and we permit this not because the activities are good for them but because the alternative, a state with the power to forbid each citizen the pleasures that diminish him, is more dangerous than the diminishment. Adult liberty has never depended on adults choosing wisely. It depends on the conviction that the authority to overrule them is a worse thing to unleash than their mistakes.

But that whole defense rests on a self that is already there. To run a preference-capturing engine on an adult is to act on that self, for better or worse. To run it on a child is to compose a self that does not yet exist, and to decide, for profit and on the child's behalf, who that self will turn out to be worth becoming. The first is a vice. The second is an authorship, and it is not ours to sell. To run a preference-capturing engine on an adult is to act on a self that already exists, for better or worse. To run it on a child is to compose a self that does not yet exist, and to decide, for profit and on the child's behalf, who that self will turn out to be worth becoming.

The first is a vice. The second is an authorship, and it is not ours to sell.

So let the adult who wants the feed that knows him have it, but make him take it on purpose: once, by his own hand, with proof of age, the way one is let into any other vice. Not watched into it, but admitted through a door he chose to open. The privacy-minded will never go near the door and will lose nothing they valued. This is the inverse of the bans, which surveil everyone to find the few. The safe feed becomes the default and demands nothing of anyone; the dangerous one becomes the exception and asks only of the people who insist on it, once, less even than the bottle they show their license for every time they buy it.

Yes, this makes the product worse. The shared feed is duller than the one tuned to you, holds you less, gives back less of the sensation people have been trained to expect. That is not a flaw in the proposal. That is the proposal. The entire claim is that the sensation and the harm are one property and not two, that what makes the feed impossible to put down is the same thing that makes it corrosive, and that asking for the engagement without the damage is asking for a hunter that declines to hunt. A speed limit makes the car slower. That is the point of it.

We have built a machine that studies what a person is and sells the finding back to him as his own desire. Among adults that is a transaction, and adults are allowed transactions that cost them something. We should hold our nerve and let them. The bans get even this backwards: they padlock the building and leave the machine running, then ask the whole country for identification at the door. Tear out the loop instead, and leave the door for the grown.

Only keep the child out of it. Not because childhood is sacred, but because it is unfinished, and a thing designed to decide what you want should never get to a person before he has had the chance to decide it himself.

I did not arrive at this view by argument. I arrived at it by symptom.

For several years I built these systems and used them more than most, and for most of that time I noticed nothing. Then the work began to curdle. There was no single morning of revolt; it was subtler and worse, a steady fog, a sense that everything had become overhead, that working now resembled switching between tasks without ever arriving at one. I suspected the ordinary causes. I was getting older. I was a founder, and founding corrodes a person in its own familiar ways. I assumed the trouble lay somewhere in my life and looked for it there.

It was not there. I set the tools aside, and the fog lifted, and it did not lift gradually. The mind returned. I distrusted the result; in fact I assumed it was false. It felt backwards, since when was being more productive the cause of being more confused? So I tested it against myself, withdrawing the tool and restoring it and withdrawing it again, and each time the effect held. The cause was not age. It was not the season. It was the instrument I had helped to build.

What follows, then, is not a prophecy. It is an explanation assembled after the fact, for something that had already happened to me.

The argument for artificial intelligence is usually made in the language of output. More software will be written. More companies will be built. More research will be produced. More wealth will be created. Tasks that once required days will require hours, and tasks that once required hours will require minutes.

This is probably true. It is also a narrow way of thinking about human life.

A person is not merely a unit of production, and the purpose of thought is not exhausted by the artifacts thought produces. We care about solving problems, but we also care about what solving problems does to us. We care about becoming sharper, more capable, more patient, and more exact. We want to experience our own intelligence as an active force in the world, rather than as a supervisory layer placed on top of machinery.

The issue is not that the technology is useless. The issue is that it is useful in a way that is corrosive to the activity of thinking itself.

Take programming. Before AI, the structure of the work was simple. You began with a system in one state and wanted to move it into another, perhaps to add a feature, fix a bug, or change an architecture. The destination was often reasonably clear. The difficult part was finding the path.

That path was the work.

You decomposed the problem. You formed a model. You discovered that the model was wrong. You noticed constraints you had not understood at the beginning. You revised your approach. You became stuck. You held the structure in your mind for hours, sometimes days, until something yielded. At the end there was code, but the code was not the only product. The mind had also been altered by the attempt. It had met resistance and acquired form.

AI changes the basic structure of this process. You state the destination, and the machine proposes a path. You inspect it. You reject part of it. You ask for a revision, and another path appears. The loop repeats: specify, generate, inspect, nudge, regenerate.

This is not the same activity. It is not even a weakened version of the same activity. It is a different relation to the problem. The human is no longer primarily constructing a solution; the human is selecting among solutions supplied from elsewhere.

There is a profound difference between generation and selection. A person can recognise a beautiful proof without being able to discover it. A person can inspect a piece of code and understand that it works without having produced the reasoning that made it necessary. A person can appreciate a strong argument without being capable of generating the conceptual structure from which the argument arose. Recognition is not creation. Verification is not understanding. Selection is not thought in the same sense as construction.

It is possible to invent cases in which selecting is difficult like to imagine a field in which the main intellectual bottleneck is choosing the right problem, identifying the right direction, or distinguishing the one viable solution from a thousand plausible failures. But this does not describe the ordinary use of AI. In the ordinary case, selection is easier than generation. The machine performs the difficult traversal through the space of possible solutions, while the user decides whether the result is close enough to what was intended. The user may need expertise. The user may need judgment. But the density of thought is lower. The sustained internal construction has been replaced by intermittent evaluation.

The standard defense is that tools have always done this. Calculators replaced arithmetic. Compilers replaced assembly. Search engines replaced memory. Each time, people moved to a higher level of abstraction. There is some truth in this, but not enough.

The relevant distinction is not between old tools and new tools, or between low-level and high-level work. There is nothing sacred about assembly language, and nothing ennobling about performing arithmetic by hand. Some difficulty is merely waste. The relevant distinction is between difficulty that consumes time and difficulty that forms the mind.

A calculator removes a bounded operation from a larger act of reasoning. A compiler removes a layer of symbolic translation. These tools can eliminate forms of labor while leaving the central generative act intact. AI is more general: it increasingly acts on the generative act itself. It does not merely execute a calculation after the problem has been formulated. It proposes the formulation. It supplies the abstraction. It writes the implementation. It diagnoses the error. It explains the result. It revises the answer when challenged.

The difference is not mystical. It is a difference of scope but sufficiently broad scope becomes a difference of function. A tool that removes one rung from the ladder leaves the climb intact. A tool that follows you upward each time you ascend changes the nature of climbing. Each time a person attempts to retreat to a higher level of abstraction, the machine follows him there.

The usual response is that humans will simply move higher still. We will no longer solve problems; we will choose which problems matter. We will set goals, exercise taste, decide what is worth building. This is usually described as a promotion.

In practice, it is mostly not.

Problem selection matters in science, mathematics, and engineering. Taste matters. Judgment matters. Some people are much better than others at recognising where a field is fertile and where it is exhausted. But problem selection is not a replacement activity of equal cognitive depth. Researchers do not sit in silence until the uniquely correct problem reveals itself. They choose problems through a mixture of intuition, accident, tractability, aesthetic preference, institutional inheritance, competition, and boredom. Then they begin the serious work and the serious work is usually the attempt to solve the problem. To automate that work and describe the remainder as more elevated is a form of consolation. It is the story a technological culture tells itself whenever it wants to believe that every subtraction is an advance.

There is also a psychological problem, and it may be the more important one. AI interaction has the structure of a slot machine. You ask. You wait. Something appears. It is plausible but not quite right, so you adjust the prompt and ask again. Another answer appears. Perhaps it is better; perhaps it is worse. You keep pulling the lever.

The reward is not understanding. The reward is movement. Something changed, something appeared, something feels closer. The mind receives repeated signals of progress without having performed the sustained work from which genuine understanding is formed. The process is active enough to be absorbing and passive enough to be degrading. It fills attention while weakening it. This is why the technology is so difficult to use well. It does not merely offer assistance; it encourages a particular rhythm of cognition that is fragmented, reactive, impatient, and dependent on external generation.

Error can provoke attention.

Plausibility sedates it.

The danger is not simply that the machine is wrong. A wrong answer can be useful; it can force the mind to wake up. The greater danger is that the machine is often right enough. It produces an answer coherent enough to accept and incomplete enough to prevent the formation of a real internal model. The user acquires the result without acquiring the path. The answer is possessed externally before it is understood internally.

When I derive something myself, I know where the uncertainty lives. I know which steps were obvious, which were difficult, which were fragile, and which remain unresolved. The structure of the solution is connected to the structure of my effort. When a machine supplies the path, I can inspect it, test it, verify parts of it but the epistemic relation is weaker. I possess the answer more easily than I possess the route by which the answer became necessary.

Over time, this matters. The mind is shaped by what it repeatedly does. A person who repeatedly confronts difficult problems develops patience, precision, independence, and tolerance for uncertainty. A person who repeatedly delegates the construction of solutions and evaluates the results develops a different set of habits: impatience, dependence, fragmented attention, and the expectation that difficulty should be discharged into an external mechanism as quickly as possible.

There are disciplined ways to use AI. One can use it as a critic, a verifier, a search tool, or a hostile reviewer. One can impose rules and preserve the difficult part for oneself. But this does not answer the objection. The existence of a controlled use does not change the character of the ordinary use. A slot machine can also be approached with discipline; that does not make it a good environment for cultivating patience.

The natural form of AI use is delegation. Its natural incentive is convenience. Its natural effect is to remove resistance. And resistance is not always the enemy. Some forms of resistance are the means by which the mind becomes capable of respecting itself.

For me, this is enough. I do not want to build elaborate rituals for ensuring that a machine designed to remove cognitive effort leaves me with enough cognitive effort to remain intact. I do not want to negotiate continuously with a tool whose central promise is that it will spare me the work I most value doing.

The issue is not purity. It is not nostalgia. It is not fear of progress. It is a decision about what kind of life is worth living.

A society can become more productive while its members become less capable of sustained thought. It can create more artifacts while producing weaker authors. It can generate more intelligence externally while cultivating less intelligence internally. This is not a contradiction. It is what happens when a civilisation confuses output with flourishing.

There are difficulties that should be removed. There are also difficulties that constitute the person who confronts them. I do not want to delegate those difficulties.

I do not want to think through machines.