AI tax debate misses the threat that’s already here

On Thursday, Senator Wyden (D-Ore.) proposed limiting tax benefits for new data centers and levying a new excise tax “to help the communities and workers most impacted as data center construction continues.”

He is one of many lawmakers who want to use taxes to address worries that AI could eliminate jobs, inflict environmental harms, and have broader negative consequences for society and global security. Proposals so far range from taxing the computing power to build or use AI models, the energy they consume, or the windfall profits accruing to AI companies themselves.

But the debate has jumped to “which tax?” before answering two fundamental questions: What are lawmakers trying to accomplish? And what can a tax actually do about it?

For many who want to tax AI, the answer to the first question is straightforward and sympathetic: raise money to help workers displaced by AI. The second question is harder. A specific tax on AI would likely do far less than its supporters hope. And the danger lawmakers are most focused on, widespread job loss, is less certain than one they rarely mention: the nation’s perilous fiscal path.

Whether the goal is to protect workers or improve the fiscal outlook, shoring up the taxation of capital income, where much of AI’s gains are likely to accrue, would likely be more effective than a new tax that singles out AI.

Even strong AI growth would leave debt rising

A new National Bureau of Economic Research (NBER) working paper from economists Karen Dynan, Douglas Elmendorf, and Louise Sheiner examines how AI could affect the U.S. fiscal trajectory under different assumptions about productivity, jobs, wages, inequality, and returns to capital. The scenarios range from a rising economic tide that lifts everyone to a wave of automation that displaces millions of workers as financial gains flow to a wealthy few.

At first, the news is encouraging. Faster growth would mean more income to tax and a bigger economy to carry the debt the country already owes. In every scenario, AI improves the budget outlook, trimming projected federal debt by 39 to 49 percentage points of gross domestic product (GDP) three decades from now.

But even those reductions would leave federal debt rising sharply: the Congressional Budget Office (CBO) currently projects federal debt will climb from 101% of GDP today to 175% by 2056. Even the paper’s most optimistic AI scenario would leave debt near 126% of GDP, higher than at any point in American history, including the peak after World War II.

AI, in other words, might slow the climb but would not reverse it. In every scenario the authors model, federal debt would “continue to rise significantly relative to GDP.”

Debt falls most when income shifts from labor to capital

The deepest debt reduction comes not when AI’s gains are broadly shared, but when AI costs workers their jobs and sends resulting gains from cost savings to the top. Even then, the tax system reaches only a fraction of that capital income because that income is lightly taxed.

Were those gains taxed at ordinary rates, federal debt would fall below its current share of GDP. What moves debt is less about whether and how to tax AI and more about how we tax capital.

There are plenty of options for strengthening capital income taxation. They range from raising corporate and capital gains tax rates and ending step-up in basis to more far-reaching approaches, such as taxing wealth.

Then there is the perennial favorite of many public finance economists: a broad based consumption tax. It would tax income, including income derived from capital, when it is ultimately spent.

Proposed AI taxes come with the same tradeoff

A specific tax on AI does two things at once. It slows the adoption of AI, since taxing anything gives you less of it, and it raises revenue. But these two objectives pull in opposite directions: A tax that sharply reduces AI use shrinks the tax base and forgoes gains from growth and productivity. A tax that raises real money, on the other hand, can likely do little to slow AI down.

Consider the leading proposals. That conflict is embedded in their designs.

A compute tax, levied on the computing power used to train models, discourages the very training that makes models better, and its base shrinks as that training grows more efficient.

A token tax, on each use of a model, falls on all users, both productive and wasteful. To the extent that AI is an input in the production of goods and services, the costs of a token tax can spread across the economy.

Other proposals would target profits instead. A windfall profit tax would not discourage AI use, since taxing a true windfall leaves the investment behind it intact. But such windfalls are hard to capture in today’s tax system, and any real revenue may be years away.

Even working as designed, none of these levers are likely to raise enough revenue, or with enough certainty, to meaningfully change the debt trajectory, risking the “fiscal illusion” that a narrow tax on AI has solved what only broader capital income taxation can fix.

Prepare for AI by taxing capital income more effectively

The worry driving the debate over taxing AI is real. AI may displace workers on a large scale as well as the labor income on which our existing tax system relies.

But the danger of the national debt—a financial hole that an AI tax cannot fill—is already here. Rising interest costs consume a growing share of the federal budget. That leaves less for everything else the government does, or can do—like respond to the next recession, pandemic, or the very wave of displacement that AI taxes are meant to address.

Lawmakers can prepare for both dangers. AI’s gains are likely to keep flowing disproportionately to capital, income that’s already taxed more lightly than labor. Closing that gap would likely do more for the fiscal outlook than a narrow tax on AI and help fund responses to its costs.