Quick Buzz Feed
Taxing the computing power behind artificial intelligence could slow innovation, reduce productivity gains, and weaken American competitiveness.
Anton Korinek and Lee Lockwood, researchers at the Brookings Institution, recently published a review of different ways AI might be taxed as it becomes a larger part of the economy. They make an important distinction between taxes applied only to the final users of AI content or services and taxes on the computing resources or equipment itself. The proposal that has recently gained traction in public debates (including an article in the Wall Street Journal) is of the second type and would be a major departure for the U.S. tax system. A direct tax on computing resources would most likely be assessed using a measure of AI computation known as “tokens.” Simple applications like text generation use fewer tokens, while coding, video, and image creation use more. Tokens themselves are essential for developers and users alike to measure and sometimes price the computing power they use. A direct tax on AI computing resources could either be charged to data center operators or to all AI users. In either case, the tax would be designed to make each unit of AI computation more expensive to produce and consume, disincentivizing investment in computing equipment and reducing the use of AI models in the production of products and services we buy.
Some supporters of a compute tax are motivated by a desire to slow the adoption of AI throughout the economy, seeking to reduce potential negative disruptive consequences such as worker displacement. In this formulation, a compute tax would work like an excise tax applied to a “vice” good. We sometimes tax things because we want people to do less of them. Many states and localities collect a tax of several dollars on every pack of cigarettes for precisely this reason. Proposals for a carbon tax also work in an equivalent manner. Raise the price of emitting a unit of greenhouse gas into the atmosphere, and manufacturers will likely substitute with energy sources that emit less greenhouse gas. However, an across-the-board reduction in the usage of AI would have profound costs that would far eclipse its imagined benefits. AI is a foundational technology, the type that inspires countless follow-on innovations as people originate or improve applications of the technology. Along with automating some white-collar work, AI is fueling medical advances, translating languages, and solving engineering problems, to name a few examples. It’s also contributing to the GDP (Gross Domestic Product), raising overall prosperity in ways that can offset disruptions such as job loss. Raising the cost of all those innovation- and productivity-enhancing uses of computing would unequivocally harm American productivity, innovation, and global competitiveness. Knowingly imposing these costs on ourselves because of concerns about certain disruptive consequences of the technology would be a grave error. The authors of the Brookings study argue strongly against taxing AI computing resources for precisely this reason, writing that it “would be like taxing steel during the industrial revolution—a self-defeating policy that could slow the productivity growth needed to fund public priorities.”
The second group interested in a compute tax sees it not as a tool to reduce or slow economy-wide adoption of AI but as a way to overhaul our tax code for a future dominated by the technology. Elon Musk, Marc Zuckerberg, and Andrew Yang have all dreamt at various times of an economy where a compute tax funds universal basic income. “Universal HIGH INCOME via checks issued by the Federal government is the best way to deal with unemployment caused by AI,” wrote Musk in April on X. A somewhat less utopian version of this argument predicts that AI will cause a major structural shift in the economy. America currently relies mostly on labor for its tax base, deriving over three-quarters of tax revenue from individual income and corporate payroll taxes. In a future scenario where AI significantly lowers total labor demand—both from displacement and labor-saving innovations—the resulting smaller tax base could indeed cause fiscal problems. However, as the Brookings authors point out, this argument doesn’t make sense under current conditions, which are defined by more limited worker displacement and people learning how to use AI to enhance human productivity. Instead, it’s a possible response to a future where AI agents and machines mostly do their work without human involvement. Simply put, we don’t know with anything approaching certainty that such a massive shift will happen. Even if it does, we know far too little about what such a future would look like to start rewriting our tax code now. Economists have long considered how our tax code might change in a future even partially less reliant on human labor, and several options are likely more viable than directly taxing AI computing resources. Moving from a tax code based on labor income to one with more prominent consumption or sales taxes (including taxing AI services at their point of final consumption, though not at earlier stages in the value chain) or changing how corporate profits are taxed would be more moderate responses to such a future. But once again, it remains too soon to know whether any of these large-scale changes is desirable, meaning that good tax policy should maintain flexibility rather than placing bets on exactly what will happen.
There’s a tendency when thinking about technology and the future of work to view “AI” and “workers” as faceless, homogenous substitutes. This is simply wrong, and it’s a mistake both groups of compute tax proponents make. AI is sometimes a substitute for human labor and at other times a complement. This relationship is subtle and will continue to evolve in unpredictable ways as people discover new or better applications of this foundational technology. Doomsday scenarios based on AI and other automation tend to imagine dropping the technology of tomorrow, fully formed, upon the world today. Utopian scenarios do much the same thing. But these changes, even when they look rapid in hindsight, take place incrementally. The importance of this distinction cannot be overstated, because it gives labor a chance to evolve, too. Those millions of human beings can and do adjust through retraining, ingenuity, and entrepreneurship. Proponents of a compute tax, whether optimistic or pessimistic about the future they see, are prone to vastly overestimate the clarity of their vision.