The IMF War-Gamed What Happens to Tax Revenue When AI Takes the Jobs

On a December morning in Washington, roughly 50 economists and technologists sat down at the International Monetary Fund and tried to break the government.

Over two days — December 10 and 11, 2025, under the Chatham House Rule — IMF staff and outside experts from the Board of Governors of the Federal Reserve System, Google DeepMind, the RAND Corporation, MIT and the Peterson Institute split into six groups and war-gamed what happens to public finance if artificial intelligence does what its most bullish builders promise. The resulting synthesis, published as IMF Note 2026/002 and led by IMF economist Natalia Tamirisa under the guidance of Era Dabla-Norris, is one of the few documents in which a major multilateral institution treats mass labor displacement as a budgeting problem rather than a thought experiment.

Bloomberg surfaced the exercise on July 19 in a Forecast newsletter by David Ramli and Walter Frick, under a blunt framing: fiscal crisis. The IMF is more measured, but not by much.

What the exercise found

Both scenarios assumed AI reaches human expert-level capability within five years and, by the early 2030s, can perform most cognitive and physical tasks. They differed only in absorption speed. In the baseline path, adoption is slow and uneven, throttled by regulation, infrastructure bottlenecks and public pushback. In the runaway path, it is rapid and near-frictionless, producing economy-wide automation and a concentration of power among a few large AI firms.

Under the second path, the tax system does not simply strain. It loses its foundation.

“Erosion of the labor tax base will necessitate adaptation of tax systems,” the note states. “Declines in employment and wage income weaken personal income tax bases, particularly under rapid diffusion, even after accounting for potentially higher wages for remaining workers.”

The obvious fix — tax capital instead — does not clear. Shifting taxation toward corporate income “may only partially offset these losses given capital mobility, market concentration, and international tax competition,” the note warns, adding that the problem is “most acute for countries that primarily consume AI services, rather than produce them.” Participants counseled against blunt instruments, urging governments to avoid “robot taxes” in favor of stronger capital income taxation at the individual level.

Spending moves the wrong way at the same moment. Employment-linked social insurance, the note argues, becomes “less effective under rapid AI diffusion as job opportunities decline,” pushing governments toward transfers decoupled from employment status entirely — up to and including universal basic income.

The numbers live elsewhere — and they cut both ways

The IMF published no figures: no GDP paths, no revenue estimates, no debt ratios. For quantification, turn to a July 1 working paper from Brookings and Federal Reserve economists Ben Harris, Neil R. Mehrotra and William Overcash. Their headline is optimistic: a once-in-a-generation AI productivity shock could cut the annual U.S. deficit from roughly 6% of GDP to as low as 2% — about 2.2 trillion dollars off the bill by 2036.

Then they spend the paper taking most of it back. Five compounding side effects claw back more than half the gain, and in the worst case roughly two-thirds. Longer lives from AI-driven medicine add an estimated 3 million more retirement-age Americans by 2036. A 3% drop in labor force participation — about 6 million fewer workers, a COVID-scale hit likely to be permanent — shrinks payroll receipts while swelling SNAP and disability rolls. Higher neutral rates add roughly 60 billion dollars in debt service, and an AI arms race could add over 350 billion in defense spending.

And the tax base itself shifts. Individual income taxes currently supply 52% of federal revenue against roughly 6% from corporate taxes, per Treasury data; a 2024 IRS study put the effective rate on capital gains near 5%. If AI moves national income from paychecks toward profits and rents, the average effective tax rate falls even as total income rises. The economy gets bigger and the government gets poorer.

Why this is now an AI policy question

The optimistic case has a loud champion. Elon Musk has argued for months that AI and robotics at scale are “pretty much the only thing that is going to solve the US debt crisis.” That is not a fringe view. The 1990s internet boom raised revenue by 2.2 percentage points of GDP and cut the deficit roughly 60% between 1992 and 2002.

The rebuttal is that the 1990s boom taxed workers. This one might not. Anton Korinek of the University of Virginia — a co-organizer of the IMF workshop — and Lee Lockwood argue in a February NBER paper that transformative AI “may gradually erode the two main tax bases that underpin modern tax systems: labor income and human consumption,” concluding that consumption taxation “may serve as a primary revenue instrument” in the first stage of the transition.

This reframes AI policy. Most regulatory debate runs on safety, copyright and competition. The fiscal channel makes diffusion speed itself a budget variable: the faster the adoption, the wider the gap between revenue loss and institutional adaptation. The IMF calls state capacity “a binding constraint,” warning that weak tax administration risks locking countries into low-growth equilibria with eroded tax bases.

What to watch

Three signals. First, payroll composition — Bloomberg reports tech and finance shedding roughly 28,000 jobs a month, and whether they reappear elsewhere determines which scenario is running. Second, federal receipt mix: a sustained fall in the individual income tax share below 52% would be the earliest hard evidence of base erosion. Third, international coordination on capital and rent taxation, which the IMF flags as “pivotal to avoid bifurcation.”

The exercise produced no forecast. It produced a warning that the U.S. and its peers have built revenue systems on an assumption — that most income arrives as wages — that a decade of AI diffusion could quietly retire.