Ready for 2030? What work looks like in 2030
Illustrative stress test
Do you want to see what work looks like in 2030?
A one-minute story using published AI and robotics scenarios, with assumed retraining capacity. It plays by itself. No sound.
Your job is gone. The bills aren’t.
If work changes faster than people can retrain, how long could you wait? Pick a month and compare course completion with today’s assumed capacity and an optimistic EduMark scenario.
Illustrative stress test · the paper’s most disruptive scenario · no assigned probability.
Loading the scenario…
The assumptions behind the numbers.
An illustration, not a forecast. We extend a US economic scenario to a worldwide workforce and add a queue for retraining. The page shows the paper’s most disruptive scenario, selected by EduMark; the paper assigns no probabilities.
The 80m annual starts are an assumed capacity benchmark. EduMark’s extra intake and shorter courses are optimistic assumptions, not measured performance. Earnings are in PPP international dollars; finishing a course means being ready for a role, not being hired.
Where the scenario comes fromThe paper, our global extrapolation and our choice of scenario
Korinek, Jones, Sacher, Cotter and McCrory (2026), Economic Scenarios for Transformative AI, sets out modest, substantial and extreme paths for AI’s effect on the US economy to 2030. We take from it the 62.4% share of workers in cognitive occupations, each scenario’s January 2030 change in cognitive employment (−0.5%, −3.9% and −21.5%), and its scenario inputs: how much work AI can do, how widely it is adopted, how much of that is automated, the productivity gain and how hard it is to move into other work.
The paper is calibrated to the United States. We apply its job mix and scenario changes to a rounded worldwide workforce of 3.6bn, inferred from the ILO’s 2026 projection of 186m unemployed at a 4.9% rate. That is our extrapolation, not a global result from the paper. We treat the employment contraction as people who need formal retraining, and we time their arrival using the paper’s AI adoption paths. We do not model each country’s occupations or education systems, so a pooled world total can understate local bottlenecks.
Why we show one future. The paper sets out three scenarios and assigns no probabilities to them. EduMark selected its most disruptive one, which it calls “extreme”, to stress-test education capacity. This choice does not assign it a probability. Under the paper’s modest and substantial scenarios, the 80m benchmark below keeps up with demand: people still have a modelled earnings gap while they retrain, but nobody queues for a place.
The assumed 80m course startsHow we set capacity, course length and retries
There is no single statistic for the world’s spare capacity to retrain adults. We use a benchmark drawn from formal education instead. UNESCO reports 269m tertiary students worldwide in 2024. In China (2024) and the UK (2023/24), annual entrants were 30.1% of enrolled students when pooled. Applying that ratio gives about 81m starts a year, which we round to 80m course starts a year.
That is a scale benchmark, not a count of free seats. The model assumes resources for 80m starts a year are funded or redirected to displaced workers; existing students do not vacate their places. Courses last six months, 85% of learners pass, and the rest rejoin the queue and use another place on their retry.
Your course and earnings estimateFirst come, first served; an illustrative PPP earnings benchmark
Course places are allocated first come, first served. When you choose the month your job ends, we place you in the middle of everyone who loses their job that month and find the first month in which enough places have opened to reach you. Your course length is fixed when you start. We show your path if you pass first time.
The modelled earnings gap runs from losing your job to finishing your course. We value it at the ILO’s 2021 global median full-time-equivalent wage of 846 international dollars a month. Those dollars compare buying power across countries; they are not cash US dollars, and your own pay, savings and benefits will differ. We stop counting at course completion, which is optimistic: being ready for a new role is not the same as being hired.
What we assume EduMark changesTwo optimistic assumptions, with the evidence and limits
Two separate, optimistic assumptions. Neither is measured EduMark performance. More places: staff time saved across the whole teaching, marking and admin workflow equals the scenario’s automation rate, and all of it becomes extra intake. Faster learning: the paper’s productivity gain is applied to how quickly learners reach the same standard, with the same 85% pass rate. Both phase in from nothing in July 2026 to full effect in January 2030, following the paper’s adoption curve. Shorter courses never multiply the yearly start budget a second time.
| Future | Automation | Intake | Starts a year | Course length |
|---|---|---|---|---|
| Modest | 50% | 2× | 160m | 4.4 months |
| Substantial | 75% | 4× | 320m | 3.8 months |
| Extreme | 90% | 10× | 800m | 2.7 months |
What the evidence supports today is narrower. An EEF/NFER randomised trial found teachers spent 31% less time preparing specified science lessons with generative AI and guidance. A Harvard randomised trial found students learned more, in less time, with a well-designed AI tutor across two physics lessons. Neither shows tenfold intake or shorter qualifications; that would need qualification-level evaluation. The model shows why capacity matters; it does not prove EduMark is indispensable.
The robotics assumptionA published shipment forecast and the workers we assume it affects
The paper covers AI in cognitive work only. We add physical work from Goldman Sachs Research’s September 2026 humanoid robot forecast: about 75,000 robots shipped in 2026, 890,000 in 2030 and 6.5m in 2035. We treat each year’s figure as the shipment rate at mid-year and grow it steadily between those points. On that path the rate passes a quarter of a million robots a year in 2028, which is when we say robots arrive at scale.
Our assumptions: each robot covers two workers’ shifts, and every displaced worker needs formal retraining, like everyone else in the queue. By January 2030 that adds about 1.9m people to the 483m from knowledge work; on the same path it reaches about 25m by January 2035, beyond the queue model’s horizon. Robots come later and start smaller than AI, but they are accelerating. Industrial robots already number 4.66m worldwide (IFR, 2024); we count only the new general-purpose wave, not factory automation that labour markets already absorb.
Goldman Sachs forecast (reported) ↗IFR World Robotics 2025 ↗
Worldwide modelled earningsHow we total the benchmark, and what it leaves out
We add up everyone waiting for a place or learning, month by month, and multiply by the same 846-dollar monthly benchmark. The average per affected worker divides that total by everyone displaced so far. It is modelled foregone earnings under these assumptions, not GDP lost or a measured personal wage loss. Benefits, severance and part-time work are not included.
Limits of the modelNew jobs, local bottlenecks and what happens after 2030
We model AI in knowledge work and the general-purpose robot wave. We do not model other physical automation, self-driving vehicles or new jobs that AI creates, and we stop the queue in January 2030 even though the robot wave keeps growing after it.
Long job loss harms more than income. In one long-running US study, workers displaced in mass layoffs had 15–20% higher death rates over the following 20 years. We cite that research for context. We do not convert it into a death toll for these scenarios: the model does not support that figure.