Kenneth Rogoff doubts that AI technology's contribution to tax revenues will outweigh massive adjustment costs

By Kenneth Rogoff*

Many people seem convinced that the productivity gains from AI will solve the problem of unsustainable budget deficits in advanced economies. By showering governments with higher tax revenues, the thinking goes, AI will leave even the most profligate countries with shrinking deficits.

That could be right. But there are many more reasons to think that such expectations are dangerously optimistic. For starters, AI seems likely to increase the share of capital in output while reducing the share of labour, which tends to reduce the tax intake. Absent a determined effort to increase taxes on capital income—which is becoming ever more difficult as wealth becomes more concentrated, politically powerful, and mobile—it is unlikely that tax revenue will grow as quickly as output.

Moreover, even if revenues do grow, what assurance is there that the political system won’t respond by just ramping up spending and deficits even more? After all, the advanced economies are already very rich. In principle, they could have easily managed their finances better if their leaders had found it politically expedient to do so. But being rich has never been an obstacle to going bankrupt. Voters have hardly begun to adjust to the realities of a higher-interest-rate world, where demographic decline and vastly increased needs for defense spending have introduced major new costs. As the median voter grows older, systems already biased against young people will become only more so.

A more immediate concern is that the transition to AI will not be smooth sailing. Many fear a sharp rise in unemployment as firms shed white-collar workers. Some commentators dismiss such warnings as “doomist,” but they might ask themselves why so many top AI researchers subscribe to them. (While some may have a financial interest in hyping the power of their models, plenty of independent experts share the same view.)

The counterpoint is that there will still be many goods and especially services that require a large human component for the foreseeable future. Even if AI can perform some tasks at low cost, humans will be very well compensated to do what AI and robots cannot. This, too, might be right, but the speed at which the AI revolution seems to be unfolding raises doubts about how quickly workers will be able and willing to adapt. Maybe unemployed white-collar workers will prove more flexible than those who lost jobs in manufacturing in recent decades. Even so, the change could come faster than anything we have dealt with before.

Perhaps the most serious and immediate issue is that despite AI’s massive potential to make life better, it could also cause profound problems in the absence of proper regulation. The financial industry is only just waking up to the risk that new AI coding tools may help bad actors penetrate systems that were thought to be safe. Likewise, deepfakes are becoming ubiquitous online, where they could threaten political stability by making it impossible for anyone to trust what they see.

In principle, AI firms’ own safety controls could forestall such problems. But with government regulation virtually non-existent, model developers are focused more on gaining a lead in what they see as a winner-take-all race than they are on safety.

Even more sinister are the military uses of AI, which could undo all the technology’s benefits in the absence of some kind of international agreement on the issue. Much has been written about the risk of drones and robots that are programmed to kill an enemy independently, without human guidance. Already, not even a world champion stands a chance against a computer in chess. If human generals cannot match AI generals’ reaction times and deep-planning capacity, the result could be an uncontrolled escalation of conflicts where human judgment on both sides might have produced a more peaceful outcome.

If AI is not an unalloyed benefit for advanced economies, the calculus is even more fraught for developing countries. Still-poor India has been one of the world’s fastest-growing large economies in recent years, far surpassing China’s growth rate. But service exports (outsourcing), perhaps the crown jewel of the Indian economy, are now acutely vulnerable to competition from AI. Even if white-collar Indians fulfilling back-office roles remotely are paid far below the level of white-collar workers in rich countries, AI could cost even less.

Of course, some countries could emerge as big winners despite all these costs. South Korea, for example, has found a niche in manufacturing the memory chips essential to developing AI, and Japan has pursued similar opportunities.

What about the United States? As the country driving AI development (alongside China), it may seem safe to assume that it will be a sure winner, and that is certainly what US stock markets think. But if that is the case, the US will likely find itself at the cutting edge of the job losses and social disruption caused by AI. Given America’s profound political divisions, there is no reason to believe the transition will be well-managed. And as Iran’s use of drones has shown, the evolution of warfare in an AI-driven world may erode America’s military advantage and force a sharp increase in defense expenditures.

Although AI could potentially help solve the problem of unsustainable budget deficits in the US and elsewhere, there is a greater likelihood that it will make things worse before they get better. Faced with society-wide disruptions, policymakers may find that fiscal prudence is the last thing on their minds.

*Kenneth Rogoff, a former chief economist of the International Monetary Fund, is Professor of Economics and Public Policy at Harvard University and the recipient of the 2011 Deutsche Bank Prize in Financial Economics. He is the co-author (with Carmen M. Reinhart) of This Time is Different: Eight Centuries of Financial Folly (Princeton University Press, 2011) and the author of Our Dollar, Your Problem (Yale University Press, 2025). Copyright: Project Syndicate, 2026. www.project-syndicate.org