Are we setting our children up to fail?
A child of around twelve is often beginning to be trusted to walk home from school alone. To cross roads, judge traffic, manage a key, handle the small daily negotiations of being unsupervised in the world. There is no single age at which we declare a child ready; we leave it to the parent, the child's maturity, the route. We extend that trust not because twelve-year-olds are infallible but because we understand — or used to — that judgment is built by graduated exposure to manageable risk. You do not learn to handle the world by being kept out of it until a birthday declares you ready.
That same child, under the ban the UK government announced on 15 June, will not be permitted a social media account until sixteen, with the first regulations expected in force around spring 2027. Old enough to walk home alone; too young to log on. Set the two beside each other and something incoherent falls out — not about screens, but about how we think competence is made.
It is worth holding both ends of a young person's early life in view at once, because the same logic is failing at both.
The professional on-ramp is closing
Start at the far end, when that child finishes their education and looks for a first job. The entry-level rung is being pulled up behind them.
A Stanford Digital Economy Lab study, Canaries in the Coal Mine?, built on payroll records for millions of US workers, finds that since generative AI went mainstream, early-career workers aged 22–25 in the most AI-exposed occupations have seen a 16% relative decline in employment, even after controlling for firm-level shocks — while older workers in the same jobs held steady or grew. For 22–25-year-old software developers specifically, headcount is down around 20% from its late-2022 peak. And the declines cluster where AI automates rather than augments the work — the codified, boilerplate tasks that used to be a junior's whole first year.
The IMF gets to the same finding by another route: young workers are concentrated in precisely the occupations most exposed to automation and least complemented by AI — what its economists call the "steppingstones" of a career ladder — with early US evidence that entry-level hiring is already thinning where the work is automatable rather than complementary.
That is the mechanism, and it is the part that matters here. The tasks that gave a junior their first five years of experience are increasingly done by a tool in the hands of someone who needs no junior. The on-ramp is not closing because the work vanished. It is closing because the work that taught you got handed to a machine.
The honest version is more tangled than "AI took the jobs," and worth saying plainly. The UK government's own snapshot of entry-level hiring finds that, in aggregate, junior hiring is falling in step with the whole market rather than collapsing uniquely — and it is careful to note that the steepest declines sit in roles where AI has advanced, but that this "is not yet causal evidence." Fair. Yet beneath the average, the pattern sharpens exactly where you'd expect: entry-level Software Engineer postings down 27%, Graphic Designer 28%, Accountant 29%, while customer-facing roles grow — and in engineering a gap has opened between junior hiring (–19%) and senior (–10%) that only appeared in early 2026. That is the Stanford finding, surfacing in British data. Meanwhile UK graduate vacancies have fallen 42% year on year, the steepest on record, surveyed UK employers now say AI handles around 39% of the work once expected of graduates, and the number of 16–24s not in education, employment or training has passed a million for the first time.
Even the Stanford authors are disciplined about causation — in a follow-up note they concede the cleanest AI signal only emerges from 2024, while adding that interest-rate swings don't explain the entry-level-specific gap either. Grant all the caveats. The thing none of them dispute is the shape: the first rung is narrowing, fastest where the machine can do the junior's job. The economist Kenneth Arrow named what that costs us back in 1962 — learning by doing. Competence isn't transferred in a lecture hall and then deployed; it accrues through the act of working. Remove the entry-level role and you sever the mechanism by which a person becomes good at anything. A cliff where there used to be a climb.
The developmental on-ramp is closing too
Now come back to the twelve-year-old, and correct the question — because the ban's champions and its lazy critics share a false premise: that it produces an offline child. It doesn't. Australia ran the experiment, and the kids didn't go to the library. They scattered — onto fringe apps no parent has heard of, onto the platforms the law itself exempted, and back onto the mainstream through a VPN. The ban didn't remove them from the internet; it removed them from view.
So the real question is not how a child learns to be safe online if she is never allowed there. It is how she learns to be safe — or tells anyone when she isn't — once we have dispersed her somewhere less watched than where she started, and made her being there a secret she has to keep from us.
None of this is a libertarian objection. The harms are real and the design is often predatory — the UN's human rights chief, Volker Türk, is right that the damage children take online flows from deliberate commercial choices: addictive infinite scroll, autoplay, persistent notifications. The question was never whether to act. It is whether a blanket age-gate is the instrument that answers those harms — or one that mostly moves them out of sight.
Australia is the one country to have actually run this, and the early evidence is not kind. Months in, more than 60% of children were still using social media, and downloads of fringe apps like Yope and Lemon8 had surged by up to 251% as under-16s decamped to platforms almost no parent could name. Worse, the ban's own exemption list built the exits: Discord, Roblox and Steam were waved through, and have quietly become new teenage "town squares," carrying the very unmoderated risks the law was written to fix. The boundary the legislation drew pointed children straight at its least-supervised side.
Ministers cite 4.7 million disabled accounts as proof it is working — a figure that says nothing about where those children went, and it was not to the library. The UK has already drawn the same lines: messaging apps excluded, gaming platforms kept with their features "locked down." The escape routes are being published alongside the rule.
And many never left the mainstream at all — they simply route back through a VPN, sitting on the same platforms as before, invisible to the age-check rather than absent from the app. Britain has noticed that loophole and is already moving to close it — but a VPN cannot tell how old its user is, so shutting children out means age-checking everyone who uses one, which is identity verification for the whole adult population in a child-safety coat, and points down the road toward a national digital ID. That is a rabbit hole for another day; the point here is narrower.
The ban's real effect is not an offline generation but a fragmented one, dispersed across fringe apps, exempted platforms and VPN'd-back mainstream accounts — every fragment harder to see, moderate, research or reach than the single visible population it replaced.
Each layer compounds the last. A less-moderated destination; no parental familiarity to draw on, because the app is one the adults never used; none of the safety scaffolding the large platforms were forced to build; and, because the child is now somewhere she is not meant to be, a reason not to tell anyone when something goes wrong. Secrecy between a child and the adults who protect her is not a side effect of online abuse — it is the mechanism. A rule that turns a child's presence into something she must hide from her own parents does not sit neutrally beside that risk. It manufactures the conditions for it.
To their credit, ministers saw the cliff-edge coming — and their answer is telling. For 16- and 17-year-olds, livestreaming and stranger-contact will stay switched off by default "to prevent a cliff-edge at 16". But that is restriction deferred, not practice granted. You cannot learn to handle a feature that is off by default at seventeen any more than at fifteen; the core social-media gate still swings open in a single motion at sixteen.
Campaigners including the Molly Rose Foundation, no libertarians, call it a blunt response — one that risks a dangerous jump from an abrupt zero to the full adult experience the moment a birthday clears the gate.
The weight of expert opinion lands there too. UNICEF's position is blunt: age restrictions alone will not keep children safe online. The skills that actually protect a young person — self-regulation, spotting manipulation, recognising misinformation, handling conflict, guarding their own privacy — are, as a body of research summarised here puts it, things "no amount of restrictions will help kids develop." A child kept off the platforms until sixteen still arrives needing every one of them — only now with no practice.
And the children being protected mostly say it won't work. In Girlguiding's research, 56% thought a ban would be ineffective, and only 15% of 10–16-year-olds believed it would make them feel safer online. On the Australian evidence, they have the mechanics right.
Abstinence-only does not have a good record as a pedagogy. We learned that with sex education. It is strange to be relearning it with the internet.
What we actually know about building judgment
Here is where the walking-home comparison earns its place. We do not keep children off roads until they are sixteen and then release them into traffic. We teach kerb drill at four, walk beside them at seven, let them go alone when they and we are ready. The risk is real at every stage; the exposure is calibrated to the developing capacity to handle it. That is how judgment grows — in physical space, in social life, and, if we were sensible about it, online.
There is a form of regulation that respects this, and it is not "do nothing." Make the platforms safe by design rather than safe by exclusion — strip the addictive mechanics, not the access. Support parents instead of conscripting them: as UNICEF notes, mums and dads are currently asked to do the impossible — to police platforms they didn't design and algorithms they can't see, around the clock and alone. A seatbelt law works by changing the design of the car and sharing the responsibility for safety — the manufacturer builds the belt in, the driver buckles up, the law requires both. The blunt age-gate does the opposite. It treats every platform as identically toxic, removes the graduated exposure entirely, and leaves the underlying machine untouched for the day the child is finally let in.
There is a sharp edge to that conscription the debate has mostly missed. The ban's teeth are not a fine posted through the family letterbox — the penalties fall on the platforms, and enforcement runs through age verification: an ID or a face scan at the door. So a parent who has read the evidence and concluded that compliance is the worse option for a child in their care cannot simply, quietly disagree. To keep a fourteen-year-old in a supervised online life rather than a hidden one, they have to help the child defeat the age check — disagreement has to become action. And look who that makes an offender: not the disengaged parent the policy pictures, handing over the tablet to let the algorithm do the raising, but the most involved one, in breach precisely because they took child safety seriously enough to learn what the ban does. A rule sold as protecting children ends up making quiet objectors of the parents most committed to protecting them. When a law can only be obeyed by acting against the interests of the child in your care, it has mistaken the gesture of safety for the substance.
Britain didn't start this
It would be comforting to read all this as one government's overreach — a British peculiarity you could vote out at the next election. It isn't. The UK is the latest country to move, not the first. Australia lit the fuse in December 2025, and by this May 23 of the 27 EU member states were drafting or weighing some form of under-age restriction. France has an under-15 ban awaiting Brussels' sign-off for a September start; Greece's arrives in January 2027; Denmark has agreed one; Portugal has already passed a parental-consent version. Britain joins at the back of the queue, holding a near-copy of the Australian model — which means it had the rare luxury of watching the experiment run first, saw the roughly 60% of Australian children still online months later and the drift to fringe apps, and reached for the same lever regardless.
Two things follow, and the argument needs both. The breadth is real evidence that the harm is real: two dozen democracies do not independently panic about nothing, and the design features they are all naming — infinite scroll, autoplay, engagement-maximised feeds — genuinely do damage. But converging on the same instrument is not the same as converging on the right one. What a continent is demonstrating, in loose formation, is a shared failure of imagination: reach for the age-gate, because it is legible and fast and satisfying, and defer the harder work of redesigning the machine. Even the EU has declined to make it a bloc-wide ban, leaving each parliament to build its own version while Brussels quietly lays the age-verification plumbing beneath. Two dozen variations on the same blunt tool will produce two dozen variations on the same outcome — the dispersal, the secrecy, the on-ramp pulled up — only now at continental scale.
Both cliffs, one society
Put the two halves together and the shape is hard to miss. The same society automating away the jobs where young adults learn to work is also sealing off the spaces where children learn to navigate the digital world — and doing the second in the name of protecting them from the first's tools.
Because notice what one of the genuinely future-proof capacities actually is. Not "being online" in the shallow sense of a TikTok account, but digital judgment: the ability to recognise an algorithmic incentive, manage your own attention, spot manipulation, protect your privacy, work alongside AI systems, and know when a machine is quietly shaping the room around you. That is precisely the fluency the labour market is coming to demand — and precisely the fluency a locked door cannot produce. It is grown the only way judgment is ever grown: through practice, guidance, mistakes, correction, and steadily increasing responsibility. We are removing the on-ramp to work, and the on-ramp to the judgment that might let someone survive its removal, in the same breath.
I have a daughter — and nieces, and a nephew: eleven children I love, the youngest three, the eldest twenty-two. My own family already runs the length of this thing. Most of them are under sixteen — the ban's exact target, only not the abstract child of a policy paper: specific kids, some already online, some nearly there, all of them about to be walled off from where their friends are, or quietly pushed somewhere I can't watch. The eldest is at the other end, out looking for a foothold in a labour market that is lifting the first rung out of reach. The thing I find hard is not any single one of these measures — several have a genuine case, and the harms they answer are not imaginary. It is that they compose. Across a childhood and into the first years of work, the gradual climbs by which a person becomes capable are being replaced, one after another, with sudden drops — each defensible on its own, and together a coherent project of pulling up the ladder while insisting it is for their protection.
The honest question is not whether children face real risk online, or whether AI is reshaping work. Both are plainly true. It is whether you can produce a competent adult by removing, at every stage, the graduated exposure through which competence has always been made. We are about to run that experiment on a generation. I don't think we will like the result — and I don't think a birthday will fix it.