The AI Jobpocalypse, Your Kids, and Chicken Impossible

AI is shaving 16% off entry-level hiring and adding 1.3 percentage points of US productivity. The real shift is who your kids will be competing against.

Cover Image for The AI Jobpocalypse, Your Kids, and Chicken Impossible

A few weeks ago, my eight-year-old came home from his after-school art club (Crafty Hands) with a concept design for a computer game. Watercolour and marker on A4: spawn strip on the left, forest mazes, two roads with cars going in opposite directions, more forest, finish line on the right with the chicken who's made it across. He wanted to call it Chicken Impossible.

Robin's after-school-club concept for Chicken Impossible. Spawn strip on the left, forest bands and roads in the middle, finish on the right with the chicken who's made the crossing.

My contribution that afternoon was about five minutes of plumbing: spin up a GitHub repo, open a PyCharm project, and show him how to dictate into the Claude Code prompt with whisper-flow-linux. From that point on, it was him. He explained his concept to the model, watched the code appear, ran it, found bugs ("the chicken keeps walking right through the cars, that's not fair"), asked for changes, ran it again. Forty-five minutes later, there was a fullscreen Pygame build of his drawing on the screen. He cannot type at any meaningful speed. He cannot write Python. He didn't need to.

I tell you this because the noise about AI and jobs is at full volume right now, and most of it focuses on the wrong people. The real story is not whether your accountant or your associate solicitor keeps their seat through 2027. It is who your eight-year-old is going to be competing with by the time they finish school. The honest answer is: not other eight-year-olds. Other eight-year-olds with AI tools.

The jobpocalypse, by the numbers

The jobpocalypse narrative goes like this: AI eats white-collar tasks, white-collar workers lose their jobs, the knowledge economy collapses. It is a satisfying story because it has clear villains, clear victims, and a clear arc. It is also wrong in roughly the same way the original automation panics of the 1970s and 1990s were wrong, and wrong in some new ways too.

The first piece of evidence everyone reaches for is the Brynjolfsson productivity number. Stanford's Erik Brynjolfsson, writing in the Financial Times in February 2026, worked through the latest Bureau of Labor Statistics benchmark revisions and found something striking. US payrolls were revised down by approximately 403,000 jobs, while real GDP held up, including 3.7% growth in Q4. Decoupling output from labour input is the textbook signature of productivity growth. His updated estimate puts US productivity growth for 2025 at roughly 2.7%, almost double the 1.4% annual average of the previous decade.

That is not nothing. That is the productivity J-curve that economists have been waiting on since the first ChatGPT release, finally turning the corner from "investment phase" to "harvest phase". The same effect powered the late-1990s computing boom and, going back further, the electrification of US factories between roughly 1900 and 1925. General-purpose technologies suppress measured productivity for a decade or two as firms tear up their workflows, and then the gains arrive in a rush.

The second piece of evidence is more uncomfortable. Brynjolfsson, with Bharat Chandar and Ruyu Chen, looked specifically at employment effects of AI in 2025 and found that recruitment for junior roles in AI-exposed sectors fell by roughly 16%, while employment for workers who used AI to augment their skills grew. That is the bifurcation in plain sight. Codified, entry-level tasks (the ones you used to give a graduate so they could learn the ropes) are the first to be absorbed. The rung at the bottom of the ladder is being sawn off.

The jobpocalypse narrative, in context

So is it a jobpocalypse? The FT's John Burn-Murdoch, writing in April 2026, spent a column gently dismantling the simple "AI can do this task, therefore this job dies" framing. His central observation, drawn from work by economist James Bessen, is that whether productivity gains lead to falling, flat, or rising employment depends on whether demand for the output is satiated. If it is not, cheaper production calls forth more demand, and employment can rise even as productivity per worker rises faster.

The historical receipts:

  • Software developers, 1990s onwards. Productivity per developer rose dramatically, but demand for software rose faster. Net effect: more developers.
  • Accountants and architects. Spreadsheets and CAD made each professional more productive. Demand for their services rose to absorb it. Net effect: more accountants, more architects.
  • Bookkeeping and accounting clerks. Same spreadsheet revolution. The clerks doing the rote ledger work, however, were obliterated.
  • Bank tellers. Survived the ATM. Killed by the smartphone, when mobile banking removed the reason to visit a branch at all.
  • Newspapers. Gutted not by a better newspaper, but by online classifieds and search engines eating the advertising business model from the side.
  • Radiologists. AI systems often outperform humans on imaging tasks. Radiology employment continues to rise, partly because regulation and insurance liability mean fully automated radiology is de facto prohibited.

The pattern is messier than the headlines. Knowledge-work productivity tends to expand the market for knowledge work, at least when there is pent-up demand. Manufacturing and other satiated markets convert productivity into job losses. The most dramatic shifts come sideways, from technologies that did not look like a threat at all (the smartphone, not the ATM, killed the teller).

Pulling Brynjolfsson and Burn-Murdoch together, the picture is something like: a 2.7% productivity bump is real and showing up in the macro data, junior knowledge-work hiring is soft right now, but the medium-run effect on senior knowledge workers is more likely to be augmentation than replacement, and the catastrophic shocks will arrive from places nobody is currently watching.

That is the news for adults. The news for our children is on a different order of magnitude.

Look one generation down

If you are a working knowledge professional today, the question is whether you can use AI to do more in less time. The answer, in the Brynjolfsson power-user data, is yes: the cohort that uses AI for end-to-end workflows (drafting whole marketing plans in hours instead of weeks, not just summarising documents) is pulling away from the cohort that uses AI as, in his phrase, a "glorified dictionary". For most adults reading this, the gap between those two cohorts is the gap that matters for the next five years of their career.

For an eight-year-old, the gap that matters is different. By the time my son sits his first job interview in roughly 2036, the productivity effects Brynjolfsson is currently estimating at 2.7% per year will have compounded for a decade. Today's "frontier" coding models will look like the spell-checker on a 2005 Nokia. The disappointing-junior-output that experienced software developers complain about now (slop, hallucinated APIs, code that does not run) is a feature of the 2026 capability frontier, not a feature of the technology. Capability is a moving curve. Today's seventy-year-olds were born before the transistor was invented; today's twenty-year-olds cannot remember a world without smartphones. The eight-year-olds entering the workforce in the mid-2030s will have spent their entire conscious life with conversational AI as a default tool, the way my generation grew up with calculators and the generation before that grew up with telephones.

The interesting question is not "will AI take my child's job". The interesting question is "what will the children who grew up fluent with AI tools do to the children who didn't?" And the early answer, from informal observation, is the same thing the kids who got home computers in the 1980s did to the ones who didn't. They will eat their lunch.

This is where I worry, gently, about the parents I see in my own circle who have set up explicit "no AI" rules for their primary-school children, in the spirit of the no-screens rule that worked tolerably well for a previous decade. The instincts are good. The execution misses the actual threat. A child who never touches a calculator until university will struggle, but the gap is closeable. A child who never touches an AI tool until university will arrive on day one of their first job competing with peers who have spent ten years iterating on prompt structures, learning when the model is confidently wrong, building intuitions for what to delegate and what to keep, and thinking architecturally about systems they could not yet write a single line of code for. That is not a closeable gap. That is the gap between the kid who learned to read at four and the kid who learned to read at fourteen.

The eight-year-old as architect

Robin cannot type fast enough to actually code. I spent five minutes explaining object-oriented programming to him as a concept (stripped of the jargon, it is a very simple idea) and he followed it, and from there he was perfectly happy engaging with the steady stream of "should I also do X?" follow-ups Claude kept offering. None of the rest turned out to matter. What he had was opinions, on every part of the game that wasn't code: the layout, the difficulty curve, what should feel fair and what should feel impossible, what the menus should do. He had drawn the thing already. The afternoon was about turning his opinions into something that ran.

That is the part worth pausing on. The watercolour is the design document. The execution layer (Pygame loop, lane speeds, separating entities into modules) is now something a model does in minutes. The bird's-eye view, the satisfaction curve, the iterating-from-feedback: those are where the value lives, and an eight-year-old can practice them as soon as he can hold a paintbrush.

The implementation, faithful to the concept, looked like this:

In-game screenshot. Brown spawn strip on the left, marker-blob trees with brown trunks, green grass strips edging the asphalt roads, red and blue cars travelling vertically through the lanes. The aesthetic deliberately echoes the original watercolour rather than a photoreal arcade game.

The full source, the design assets, and the smoke tests are on GitHub: zhubarb/chicken-impossible. For the avoidance of doubt: the game is licensed by Robin Sesen of London, age 8. Anyone wishing to copy, fork, repackage, reskin, or otherwise lift the gameplay concept will be billed a flat £100 million pounds (in cash, ideally in £2 coins), payable to Robin, on the understanding that this is a perfectly normal way to license a piece of intellectual property and that his solicitor will be in touch.

Implications for parenting and pedagogy

The instinct to keep small children away from AI tools is the same instinct that, a generation ago, kept some children away from the home computer because it was "just for games" or "rotting their brain". That instinct cost those children meaningfully when they hit the workforce in the late 1990s and discovered that the kids who had been "wasting time" with QBASIC and Excel macros had a five-year head start on the productivity tools that ran the corporate world.

The shape of the bet is the same now, only the gradient is steeper. A few practical positions worth considering, none of which I would defend as a complete parenting philosophy:

  • The model is a collaborator, not an oracle. The most damaging habit a child (or an adult) can form with AI is to type a question, paste the answer, and stop. The whole skill is in the back-and-forth: pushing back, asking for alternatives, spotting where the model has confidently bluffed. Robin's most common phrase during the Chicken Impossible session was "no, that's not what I meant, I want it to do X". That instinct, that the model is wrong sometimes and you are the one who decides, is the central skill.
  • Architecture before syntax. A child who can describe what they want a system to do, in plain English, has the only durable skill in this stack. The implementation tier is rapidly becoming a commodity. The "what should this be" tier is not. Drawing the level layout on paper before any code was written did more for Robin's understanding of game design than any amount of typing would have.
  • Supervised use, not banned use. Eight-year-olds should not be on Claude alone any more than they should be on YouTube alone. Sat next to a parent, reading the model's outputs aloud, deciding together what to keep and what to push back on, is a different activity entirely. The supervision is the pedagogy.
  • Voice is the unlock for pre-typing children. whisper-flow-linux and similar speech-to-text tools collapse the typing barrier entirely. A six-year-old who can speak in full sentences can now do things that a fourteen-year-old in 2015 could not.

None of this is a recipe. It is a direction of travel. The point is that the direction of travel is not "shield your children from this technology so they grow up uncorrupted". The direction of travel is "your children will live and work in a world where this is the substrate, and the kids whose parents helped them learn to use it well will have an advantage that the others will struggle to close".

The honest case for adapting

In some quarters, AI scepticism has become a kind of moral position. Refuse to use it, talk darkly about its dangers, picket its data centres, fold your arms. I understand the impulse. I share several of the concerns, particularly around concentration of power in the small handful of labs that can afford to train foundation models. But the consequences of that posture are not symmetric across generations.

If you are a working adult and you decide to opt out of AI tools, you are making a bet that the productivity gap between you and your AI-using colleagues will not be large enough to matter over the remainder of your career. That bet might come off. The Brynjolfsson power-user data suggests the gap is real but bounded, and a senior knowledge worker with two decades of domain experience may well stay competitive on judgement alone for a long time.

If you make the same bet for your child, you are making it with an entirely different time horizon. You are betting that in 2036, on day one of their first job, the gap between them and the kids whose parents helped them iterate with these tools from age six will be small enough not to matter. I would not take that bet. The compounding period is too long, and the technology is moving in the wrong direction for the bet to win.

The sober version of this argument is not "AI is wonderful, embrace it". The sober version is: AI is the next general-purpose technology, it follows the same J-curve as electrification and the computer, the displacement effects are real but mostly survivable for adults, and for children the right strategy is the strategy that worked for every previous general-purpose technology. Get them in early, supervise the use, teach them the architecture and the scepticism alongside the prompting, and trust them to figure out the parts you cannot anticipate.

Chicken Impossible is not the point. The game was a Saturday afternoon. The point is that an eight-year-old now knows what it feels like to want something to exist, talk to a model about it, push back when the model gets it wrong, and end up with the thing. That is the posture you cannot bolt on later: AI-native, comfortable speaking to a model, comfortable driving it as a tool. He has it now. By the time the kids who do not have it arrive at his job interview, the gap will be the only thing in the room.

Interactive Tools

  • LLM Cost Calculator: See live pricing across 300+ LLM APIs, including the small-model tier that makes home projects like this affordable
  • Explore Our Free Tools: Hands-on calculators and visualisers for statistics, machine learning, and quantitative finance
  • AI Adoption by US State: How many US businesses report using AI, state by state, from the Census Bureau's fortnightly survey

Stay in the loop

Get notified when new tutorials and deep-dives are published. No spam, unsubscribe anytime.

Related Posts

Frequently Asked Questions