Suffering on the road to abundance
The First Industrial Revolution created unprecedented wealth – but for the first 60 years, workers saw little of it. We know of this as Engels’ pause.
Although the general safety of AI remains the most significant concern, we must consider technology’s best outcomes: Where it is safe and provides meaningful economic value. Specifically, we face a significant problem where displaced workers will likely lose income prior to materializing the benefits of AI.
The average worker will benefit from this technology in the long run, however the short- to medium-term loss of income can be long and severe. The political fallout of such a loss may be amplified because it is felt most acutely by a politically engaged professional worker class. Other workers will see immediate economic benefits.
Closing the gap in lost living standards will be one of the defining political issues of our era.
The modelThis is a tool that lets you predict the impact of AI on workers’ economic wellbeing across the world. Choose how markets and governments respond to explore outcomes.
How to use itPick an AI takeoff, a country and a redeployment policy in the assumptions panel. The verdict, cards and chart update instantly. Hover any dotted-underlined term for its definition; finer controls are under Fine-tune and Advanced.
The dates are illustrative, not forecasts. They move by years with assumptions nobody can pin down today, especially how much growth AI brings and how much new work it creates. The range above shows how far.
Adoption follows a logistic curve S(t), delayed by each country's adoption lag.
Labor share follows the task model of Acemoglu and Restrepo: AI automates a share D = Dmax·S of exposed work, and new tasks give back a share r of it, so L = L₀ × (1 − exposure × D × (1 − r)). Output per worker grows at the baseline rate plus the AI boost, which is set for advanced economies and scaled by each country's exposure. In Baseline that is 1 point for advanced economies and about 0.6–0.7 for the world, depending on how much is redeployed, near the top of the IMF's estimate that AI adds 0.1–0.8 points a year to global growth over the medium term (IMF, January 2026, via Reuters). Countries that lag still get the global spillover share of the world's AI-driven growth.
Wages are reconciled to the economy's total labor income L·Y: everyone working earns one going wage, re-hired workers earn the new-job share of it, and displaced workers who aren't re-hired earn nothing and draw the existing safety net: its first-year rate while newly jobless, then its long-run rate once insurance runs out, with benefits capped at 1.5× typical pay. Workers paid above typical pay lose most of their skill premium when displaced, because AI automated the skill that earned it; only the premium-kept share carries over to the new job. Wages across all workers therefore always add up to labor's share of output. The default worker is the average across exposed jobs at the income you set; the other fates follow one person who loses their job at the country's adoption midpoint.
Redeploying AI profits has costs. National taxes on them push reported profits abroad at 0.8% per point of tax, the meta-analysis consensus; shifted profits escape the tax but still go to their owners. The pooled share is globally coordinated and can't be shifted. Redeployment also slows AI investment by the investment response. Pooled money from abroad loses value as it grows relative to GDP, through inflation and a stronger currency, losing half once inflows reach the absorption limit.
New capital income is (1−L)·Y − (1−L₀)·Y_cf. The redeployed share of the domestically taxable part is split equally among exposed workers, capped by tax headroom. A pooled share is split equally among exposed workers worldwide. Both are converted to dollars at output per worker (GDP per person over the employment rate) and arrive as flat amounts, so they replace more of a low income than a high one.
AI-driven profits are split: the value-chain share goes to the global AI value chain and returns to each country by its share of AI-firm ownership, financial wealth, manufacturing, resources and data center capacity under the chosen mix; the rest stays with the domestic firms that adopted AI. The country can tax and redeploy what it captures.
Investment income is explicit: wealth earns the base return plus an AI uplift equal to owners' share of AI-driven capital income, after redeployment, divided by the capital it accrues to. Home-captured AI income is spread over the country's own capital; the capital-pools share is spread over all the world's financial wealth. Wealth is held at its 2026 real value, without capital gains.
What reaches workers is scaled by delivery efficiency. Well-being counts transfers at a discount and subtracts a non-income cost of joblessness. That cost doesn't fade with time, since people adapt little to unemployment, but it eases when joblessness is common. Cheaper AI-made goods are part of real output, so they reach workers as higher real income. The purchasing-power line adjusts for income: lower earners buy fewer of the goods AI cheapens most. Recovery dates are computed out to 2100, and the range reruns the model with the growth boost halved or raised by half, new work 25 points lower or higher, and re-employment, new-job pay and the skill premium kept shifted with them. "Without the dividend" reruns the model with nothing redeployed, so it also counts the faster growth and higher capital income that come without the levy.
Exposure follows the IMF's estimate that about 60% of jobs in advanced economies, 40% in emerging markets and 26% in low-income countries are exposed, with about half of exposed jobs likely to be hurt rather than helped, through lower pay as well as lost jobs. Baseline therefore assumes 30% of exposed jobs are eliminated outright (IMF, 2024).
The world labor share starts at 52.3%, the ILO's 2024 estimate (ILO via VOA). Country labor shares, tax ratios, GDP per person, employment rates and typical pay are rounded approximations of recent data.
Channel shares are rough estimates. Capital-market profits are assigned to the firms' owners: domestic investors keep what foreigners don't own, and the foreign-held share is split by how much each country's investors hold abroad, which favors deep capital markets like the US. Capital pools aren't assigned to countries: that share of AI profits raises the return on every dollar of the world's financial wealth (about 2.6 times world GDP, per Allianz) by the same amount. Manufacturing follows manufacturing value added, data centers follow hosted capacity (the US holds about 54% of hyperscale capacity, per Statista), and resource rents stay where the resources are, though states have historically struggled to capture them.
Tax headroom is the gap to a 45% tax-to-GDP ceiling, about Denmark's 2024 level (OECD Revenue Statistics 2025), scaled by delivery efficiency.
Re-employment (65%) follows the BLS Displaced Worker Survey (BLS, 2026); new-job pay (80%) follows lasting earnings losses of 15–25% after displacement (Lachowska, Mas and Woodbury). New work created follows Acemoglu and Restrepo (2019). Safety-net rates for OECD countries follow the OECD's net replacement rates in unemployment for 2024–25, for a single person without children on average pay, with social assistance and housing benefits, so that child benefits paid regardless of work don't count as replacing lost pay (OECD Benefits and Wages); means tests and incomplete take-up mean many people get less. Profit shifting follows Heckemeyer and Overesch. Inflow absorption follows aid studies that find diminishing returns around 20% of GDP (UK aid-absorption review). Well-being follows Clark et al. (2008) on adaptation and Clark (2003) on unemployment as a social norm.
Adoption lag, delivery efficiency, safety-net replacement outside the OECD, the investment response and the size of the jobless cost remain judgment calls, which is why they stay adjustable.
The "for scale" line compares this worker's worst year and time behind with past downturns, each measured as the fall from the prior peak: US median household income in the Great Recession and the pandemic (Census via EPI; Census via AP), US real disposable income per person in the Great Depression (BEA via FRED), British real weekly pay after 2008 (ONS), Greek real wages (ZEW) and the earnings of laid-off long-tenured workers (Davis and von Wachter). Against the no-AI path the model measures a shortfall from trend, which is a stricter test than a fall from the peak: those downturns left people even further behind their trend.
Cheaper AI-made goods are counted, because output and wages are measured in real terms. Still left out: rising asset prices, migration, the effect of transfers on spending, and differences within the exposed group beyond the four worker fates.
The AI uplift to returns counts all of owners' AI-driven capital income as a higher return on existing wealth. Part of it would really be the normal return on new investment in AI, so the uplift is an upper bound for someone who doesn't add to their savings.
The world fund assumes every country joins and pays in. For poorer countries it can be most of the dividend, so their results depend heavily on that cooperation.