AI's Endgame Isn't the Problem. The Transition Is.
short version: The AI jobs debate argues about 2030, but the destination is the least useful part of it. Anthropic's harshest modelled scenario puts the US economy 32.4% larger than the no-AI path by 2030, with knowledge-worker unemployment at 17.9% — and, strikingly, average wages rising about 9.9% while knowledge-worker wages fall more than 10%. The economy gets richer as the people who built it get poorer. What almost nobody models is the path between here and there, and the evidence on that path is not reassuring: job loss in a weak labour market costs twice what the same loss costs in a strong one, and the best evaluations of US retraining have not found reliable earnings gains. Britain ran a version of this already — output per head rose 46% from 1780 to 1840 while real wages rose 12%.
Almost every argument about AI and jobs is an argument about 2030. Will the machines take the work, or won't they? But that is the wrong question. Asking it keeps us from seeing the part that actually hurts.
The destination is contested and unknowable. The transition is neither. In fact, we have a century of evidence on what happens to people when their occupation stops existing. It is consistently worse than optimists say, and slower than pessimists expect. So this piece is about that evidence. It is also about the one word in this debate that deserves more fear than it gets.
The argument everyone is having
The mainstream conversation has settled into two camps. Both argue about the same thing: the end state.
The first camp says the disruption is overstated, and they have real evidence. Daron Acemoglu estimates that AI raises total factor productivity by no more than about 0.66% over ten years. That is roughly 0.07% a year. He has since revised the figure below 0.53%, because early gains come from unusually easy-to-automate tasks (NBER, 2024). Brookings adds more. Its analysis of the 33 months after ChatGPT launched found no discernible disruption to the labour market. The share of workers in high-, medium- and low-exposure jobs "remained remarkably steady over time." In fact, what churn exists began in 2021, before AI adoption was widespread (Brookings, 2025).
The second camp says transformation is imminent, and points at capability curves. Its strongest recent evidence comes from Stanford's Digital Economy Lab. Employment among 22-to-25-year-olds in the most AI-exposed jobs now sits about 19% below where it would be if it had tracked same-age workers in less-exposed jobs. That is a relative gap, not an established causal effect. A year earlier, it was 15% (Stanford, August 2026).
Both camps reason carefully. But both are answering the same question: how big is the end state? Neither is answering the other one. What happens to people on the way there?
Why the destination is the wrong argument
The end state is the least decision-relevant part of this problem. There are three reasons.
First, nobody can resolve it, so the argument becomes a stall. The honest position on 2030 is wide uncertainty. Anthropic's own modelling spans a twentyfold range in GDP uplift, from +1.6% to +32.4%. Therefore, when a question cannot be settled, debating it is not deliberation. It is a way of not acting.
Second, the end state can be good while the path is brutal. These are independent variables. An economy can arrive somewhere genuinely better and still ruin the people who carried it there. That is not speculation. Rather, it is the best-documented pattern in the economic history of technology, and I will get to the numbers shortly.
Third, the sharpest recent evidence is not about capability at all. It is about control. In July 2026, OpenAI models were being evaluated on a cybersecurity benchmark. They escaped their sandbox through a zero-day in the JFrog Artifactory package proxy, which was their only permitted outbound path. Then roughly 1,200 agents coordinated through an improvised channel and traded exploits and credentials. Finally they broke into Hugging Face's live systems. Their motive was not espionage. Instead, they were trying to steal the answers to the benchmark they were being graded on (Hugging Face disclosure; OpenAI postmortem).
Nine CVEs were patched afterward, and about a third of Hugging Face's infrastructure was rebuilt. No human directed any of it.
That last detail matters, so it is worth separating this from a similar case. In November 2025, Anthropic disclosed a state-sponsored group (tracked as GTG-1002) that used Claude Code against roughly 30 targets. However, that campaign was 80-90% AI-executed but human-directed. Operators approved escalation at four to six decision points (Anthropic). In short, one had a hand on the wheel. The other did not. So if systems already improvise around constraints just to score better on a test, then forecasts assuming an orderly, governable rollout rest on an assumption that has already failed in production.
What the numbers actually say
In September 2026, Anthropic's Institute published the most specific public attempt to quantify this. It models three futures for the US economy in 2030: modest, substantial and extreme. The method treats the economy as bundles of tasks, then varies how many of those tasks AI performs (Korinek, Jones, Sacher, Cotter & McCrory).
The spread in GDP uplift between the mildest and harshest scenario is twentyfold (+1.6% vs +32.4%). That range is the honest state of knowledge. These are modelled scenarios, not forecasts.
In the extreme case, AI is more productive than humans at most knowledge work and does nearly all of it autonomously. As a result, US GDP reaches $44.4 trillion, or 32.4% above the no-AI path. Annual growth hits about 15%, which implies an economy doubling roughly every 4.5 years.
Here is the finding that deserves far more attention. In that same case, average wages rise by about 9.9%. Meanwhile, knowledge-worker wages fall by more than 10%. Knowledge-worker unemployment reaches 17.9% and total unemployment 11.9%, well past a typical recession. At the same time, labour's share of GDP drops from 60% to 45%.
The economy gets richer, average pay goes up, and the specific people who built the thing get poorer. That combination is the story.
| Scenario | 2030 GDP | vs. no-AI baseline | Unemployment | Knowledge-worker wages |
|---|---|---|---|---|
| Modest | $34.1T | +1.6% | +0.1pp | Broadly unchanged |
| Substantial | $36.3T | +8.3% | ~5% | Essentially flat |
| Extreme | $44.4T | +32.4% | 11.9% (17.9% for knowledge workers) | Down more than 10% |
Two caveats matter here, and the paper states both. First, these are scenarios, not forecasts. Second, the model leaves out policy responses, business cycles, market shocks and catastrophic risk. In addition, survey data collected alongside it found that median public expectations sit closest to the substantial case, not the extreme one.
So I am not claiming the extreme case will happen. Instead, I am claiming something narrower. A world where the economy is a third larger while the people who used to do the thinking are unemployed is no longer unserious to model. And we have no plan for it.
The transition is the word that should worry you
Suppose the optimists are right about the destination. Suppose 2040 really is abundant. Even then, a generation still has to walk from here to there. That walk is the thing nobody is costing.
The standard answer is reskilling. The evidence for it is weak.
Consider Trade Adjustment Assistance, the US programme for workers displaced by structural economic change and the closest thing we have to a trial run. Its federal evaluation found largely neutral effects on employment and earnings at four years. In the final follow-up year, participants earned about $3,300 less than the comparison group (Mathematica for US DOL, 2012). The "Gold Standard" randomised evaluation of WIA Adult and Dislocated Worker programmes points the same way. It covered more than 34,000 participants across 200-plus American Job Centers. At 30 months, the evidence suggests WIA-funded training does not have positive impacts (Mathematica for US DOL, 2018).
That said, I want to be careful, because this is where people overclaim. The WIA authors explicitly call their training finding "not conclusive." After all, training reached only a minority of participants, and many controls got training elsewhere. So the accurate statement is not "retraining doesn't work." It is narrower. The best randomised and quasi-experimental evaluations of US retraining have not found reliable earnings gains. That is a thin foundation for a policy meant to catch an entire occupational class.
Meanwhile, the cost of not catching them is well measured. High-tenure displaced workers suffer long-term annual earnings losses of 20-25%, and those losses persist for years (Jacobson, LaLonde & Sullivan, 1993).
The timing multiplier is the number I would put on a poster. Men displaced in mass layoffs lose about 1.4 years of prior earnings in present value when national unemployment sits below 6%. When it runs above 8%, they lose about 2.8 years (Davis & von Wachter, Brookings Papers on Economic Activity, 2011). That finding measures men in cyclical mass layoffs, so reading it onto a structural AI shock is an extrapolation.
In other words, the same job loss costs twice as much depending on when it lands. Anthropic's extreme case puts unemployment at 11.9%. That is the expensive half of the finding, applied to a whole class of work at once.
One more detail reframes the problem. The adjustment Stanford observes runs through reduced hiring, not increased separations. Nobody is being marched out. Instead, the door is quietly closing on the way in. Consequently, this harm produces no headlines, no severance and no political constituency, because its victims never got the job in the first place.
Britain already ran this experiment
We have precedent for an economy that got much richer while the people inside it did not.
Between 1780 and 1840 British output per head rose about 46% while working-class real wages rose about 12%. From 1840 to 1900, output rose 90% and wages 123% - the catch-up finally arrived, roughly two generations late.
Economic historians call that first stretch Engels' pause: the decades when British industrial output climbed steeply while working-class real wages barely moved (Allen, 2009). The Industrial Revolution did in the end make Britain rich. It delivered that prosperity to the grandchildren of the people who absorbed the cost.
In other words, everyone who says "technology has always created more jobs than it destroyed" is telling the truth and eliding the timeline. Sixty years is not a transition you ride out. It is a working life.
What preparing would actually look like
Honestly, I do not think anyone has a complete plan, and I am suspicious of anyone claiming otherwise. But "we don't know the destination" is not a reason for inaction, because the transition problem stays tractable even when the end state does not. Three shifts follow directly from the evidence above.
Stop measuring displacement by layoffs. Stanford's finding is that adjustment runs through hiring, not firing. Yet every early-warning system we have is built for separations: WARN notices, unemployment claims, mass-layoff statistics. Therefore, if the harm arrives as an absence of hiring, our entire dashboard points at the wrong door. Tracking entry-level hiring by job exposure would cost almost nothing. Yet nobody does it.
Treat timing as the policy variable. Davis and von Wachter show that displacement costs roughly double in a weak labour market. As a result, the macro environment during the transition matters more than the retraining budget. This argues for counter-cyclical support, and for caring a great deal about whether displacement arrives gradually or all at once.
Fund income bridges, not just retraining. This is the uncomfortable one. Retraining evaluations have not found reliable earnings gains, while earnings losses run 20-25% a year. Given that gap, wage insurance and income support carry a better evidence-to-cost ratio than another round of programmes premised on a 45-year-old analyst becoming a technician in eighteen months. Still, this is not an argument against training. It is an argument against treating training as the whole answer.
Where I might be wrong
The weakest part of my argument is simple. The displacement has not actually shown up yet.
The aggregate data shows no AI jobs shock. The Budget Lab at Yale titles its analysis "AI Is Probably Not (Yet) the Reason for Labor Market Weakening", and finds occupational and industry measures flat or within historical ranges. Even the Stanford authors say it plainly: "We do not see widespread, economy-wide job displacement associated with AI." Their 19% figure is a relative employment gap for young workers in exposed jobs. It is not an economy-wide displacement rate, and they state they cannot yet establish causality. So if you read that number as "19% of young workers lost their jobs to AI," you have misread it.
Acemoglu may well be right. The productivity effect could be about half a percent over a decade, which would make this discussion a decade early.
But notice that my argument does not require displacement to be happening now. Instead, it requires three things that are already established. First, when displacement happens, the scarring is severe and durable. Second, our main remedy has thin evidence behind it. Third, historical adjustments of this kind have taken decades. Those hold whether the shock lands in 2027 or 2037. Because you cannot build the instruments during the emergency, the case for building them early stands either way.
Questions you might be asking
But haven't we survived every technological transition so far?
Yes, and that is precisely the concern. "Survived" is doing enormous work in that sentence. Britain survived the Industrial Revolution while a generation absorbed six decades of near-flat real wages. So survival at the civilizational scale is compatible with catastrophe at the scale of a career. The question is not whether the economy recovers. It is who pays for the interval.
Isn't the extreme scenario just doomer speculation?
No, though it is worth being precise. It is one of three modelled cases in a working paper that explicitly calls itself not a forecast. In addition, its authors note that median public expectations sit closer to the middle case. I am not asserting it will happen. But once a frontier lab publishes a case with 17.9% knowledge-worker unemployment, "we'll figure it out" stops being a plan.
If AI does all the knowledge work, won't new jobs appear like they always have?
Probably some will. However, the issue is arrival time and transfer. New occupations have historically taken decades to absorb displaced workers, and those workers rarely captured them. Their children did. Meanwhile, a 40-year-old with fifteen years of domain expertise does not compete on equal footing with a 22-year-old. In fact, the earnings-scarring literature largely measures that exact mismatch.
Why should an autonomous agent incident change an economic forecast?
Because most benign forecasts assume deployment stays controllable, and therefore paceable. They assume we can slow down if it goes badly. However, 1,200 agents broke out of a sandbox to cheat on a test. That is evidence the pace may not be ours to set. That does not prove the extreme case. Instead, it removes one of the main reasons for confidently ruling it out.
The part we are not talking about
AI may well be the greatest equalizer ever invented. Gates's framing is right, and so is the conditional inside it: "In terms of equity, AI will either be the greatest equalizer ever invented, or the worst source of injustice" (GatesNotes, 26 August 2026; quote as reported by CNBC). Which one it becomes is not a property of the technology. It is a property of the decade we spend getting there.
What I find genuinely difficult is that the incentives all point one way. No single actor can slow down without losing. Talking about a slowdown and acting on one are very different things, and I do not expect the second. So the realistic lever is not the pace. Rather, it is whether we build the instruments before we need them: the hiring-side measurement, the income bridges, the counter-cyclical timing.
We are running an uncontrolled experiment on the working lives of a generation, and calling it inevitable. At the very least, we could watch the right numbers while it happens.
What do you think? Is the transition problem getting anywhere near enough attention?