Bad Taxes Would Slow AI Innovation

Editor’s note: This column appeared in the South Korean publication the Chosun Ilbo and is published here in English with permission.

Artificial intelligence is already changing how firms produce, how workers perform tasks, and how economies generate value. As anxiety grows that AI could replace labor, policymakers are starting to worry that the tax systems built around human work may become less reliable.

That concern is understandable. Across OECD countries, personal income taxes account for roughly a quarter of total tax revenue, and social security contributions account for a similar share. In other words, modern welfare states still rely heavily on wage taxes. If AI significantly reduces labor income, governments could face pressure on the revenue streams that fund pensions, health care, unemployment support, and worker training.

That is why proposals to tax AI more directly are gaining attention. Anthropic CEO Dario Amodei has discussed the idea of a tax on AI model usage, often described as a token tax. OpenAI has suggested that policymakers may need to consider higher taxes on capital gains and corporate income, levies on AI-driven returns, and incentives for firms to retain and retrain workers.

A separate question is how governments should capture the value that AI creates. If AI generates enormous gains, how should societies ensure those gains are broadly shared?

The answer should not be to tax the use of AI itself.

There is a difference between modernizing the tax system and creating a new tax on AI adoption. Governments may need a more durable revenue base to fund welfare programs, retraining, and transition support. They may need to rethink how they tax capital gains, corporate profits, consumption, digital services, or tax expenditures. Those are real policy debates. But taxing AI usage, tokens, compute, or automation itself is a different matter. That would not simply raise revenue. It would raise the cost of adopting productivity-enhancing technology, which is the basis for national competitiveness.

The debate should also not be driven by the assumption that AI will quickly destroy the labor tax base. A recent ITIF model shows how extreme that scenario would have to be. If labor’s share of income started at 57 percent, it would need to fall to 52 percent in just one year for tax revenue to drop below the previous year’s level. For the tax base to keep weakening over a decade, labor’s share would have to collapse to roughly 21 percent—an outcome so extreme that it should not be treated as a realistic basis for policymaking.

The real task is to distinguish bad taxes from better tax policy. Bad taxes are those that make firms pay simply for using AI. A token tax, a compute tax, or a tax on automation would penalize firms that use AI to increase productivity. The more a company uses AI to improve operations, reduce waste, develop new products, or serve customers more efficiently, the more it would pay. That is not a neutral tax reform. It is a tax on innovation.

Such a tax would be especially damaging because AI adoption is still uneven. Large technology firms can absorb higher costs. Smaller firms often cannot. If governments raise the price of AI tools, the burden will fall hardest on the companies that most need productivity gains: small businesses, manufacturers, service firms, and startups. Instead of spreading AI across the economy, policymakers would make it more expensive for ordinary firms to catch up.

Better tax policy would start from a different premise. If governments need more revenue, they should broaden tax bases rather than single out AI. That could mean improving consumption or sales tax systems, reducing poorly designed or overlapping digital services taxes, reforming tax expenditures, and making sure corporate tax rules remain neutral across sectors and technologies. If governments want firms to help workers adjust, they can use targeted incentives for training, wage insurance, and job transition programs without penalizing the deployment of AI itself.

Many of the groups urging governments to rethink tax policy in response to AI are also advocating policies that would slow or limit AI deployment. For example, Windfall Trust—a nonprofit funded by the Future of Life Institute, which led the 2023 call to pause AI development—has argued that AI-driven labor displacement could erode government tax bases. If governments respond to such claims by imposing punitive taxes on the development or deployment of AI, they would end up constraining the technology through tax policy rather than through direct regulation.

AI may cause job displacement, but labor tax losses do not automatically translate into fiscal collapse. Indeed, the fiscal impact on most economies would be minimal, even with some job displacement, if the productivity gains are largely retained by domestic companies and consumers. This distinction matters for Korea. Korea’s problem is not too much AI diffusion. It is too little productivity growth, especially in services and among small and midsize enterprises. Korea needs AI to spread faster across the real economy, rather than remain concentrated in a few large companies or frontier labs. Treating AI mainly as a new tax base would move policy in the wrong direction.

Taxing compute as AI transforms the economy makes no more sense than taxing tractors would have during the agricultural revolution. The right goal is not an AI tax. It is an AI diffusion strategy paired with sound tax reform. The government should help firms adopt AI, help workers transition, and ensure that the gains from productivity growth are broadly shared. But it should not make the use of AI itself more expensive.

Tax policy should help society share the gains from innovation. It should not become a barrier to the innovation that creates those gains in the first place.