So you want to have an AI token tax?

There are better ways to raise revenue and counter job loss

There are better ways to tax AI. (Image Source: Shuttershock)

There are a lot of reasons why someone might be anxious about artificial intelligence’s rapid penetration into society: it could lead to job loss, exacerbate income inequality, and maybe eliminate our ability to independently reason and critically think. In the tax policy world, specifically, people are worried that job displacement could lead to lower income tax revenue —which makes up over half of the federal tax base —while higher unemployment rates simultaneously increase demand for social assistance programs.

These uncertainties could impact our economy at the same time other fiscal issues continue to close in. The impending insolvency of Social Security, the worrying magnitude of national debt, and the need for reforms to our social welfare programs all put the federal purse in a precarious position. This makes AI’s threat to federal revenues an even more serious problem.

One of the tax proposals touted as a solution by numerous people, including former California gubernatorial candidate Tom Steyer and Anthropic CEO Dario Amodei, is a token tax.

Part of the allure of this tax is its supposed simplicity. A token tax would take a set dollar amount of tax per AI usage token purchased. Tokens are the units of data AI providers use to measure input and output. They can be measured in words, characters, pixels, or lines of code. What counts as a token can also vary by model and what exactly you are asking the AI to do. When you pay $20 a month for Claude, you’re functionally paying for a certain number of tokens.

Depending on who you ask, the intended purpose could be to slow down AI development and use, prevent downstream harms from AI, or capture tax revenue connected to the products offered by AI companies.

Despite how simple it sounds, there are a lot of potential drawbacks to this approach. It’s important for policymakers and tech leaders to understand these complications as they begin to think through potential tax reforms. Instead of complicated “whack-a-mole” policies like a token tax, policymakers should focus on leveraging taxes that are easier to administer and capture more predictable revenue from an AI-driven economy.

It likely won’t slow down AI adoption

One of the primary reasons why some policymakers are drawn to a token tax is its potential to slow down the pace of AI adoption. But would a token tax significantly reduce AI uptake?

A tax-driven ‘slow down’ of AI itself could be slow or non-apparent. That’s because the hyperscalers driving the AI boom are already accustomed to operating at a loss for AI development. So if a token tax hits, AI providers could undershift, meaning they would absorb the cost of the tax to maintain their current prices and user base until AI is even harder to replace in day-to-day life and to regulate. Even if the tax is passed down to consumers in the form of higher prices, most may just eat the cost because the technology is seen as integral in their day-to-day life (in some cases, people may not change their behavior at all because they are genuinely addicted).

Even if most consumers don’t change their behavior, it’s worth considering which Americans would be most likely to reduce their AI usage. If, or when, a token tax hits consumer pocketbooks, it would likely be regressive, and could price out lower-income Americans who may otherwise use these tools in their work and personal lives. This specific type of slow down would be a problem. That’s because what sets AI apart from other goods taxed in comparable ways, like tobacco and alcohol, is that it can benefit the user in the form of greater productivity and efficiency. Those who can afford the tax could continue to adopt AI tools and benefit from the efficiency gains —an advantage that would build on their existing income and wealth.

Similarly, a token tax could present a competitive setback to small American businesses. Just like in the individual case, a large business that is better able to afford the cost of a token tax may use AI to outcompete the small business that has to pass those costs along to consumers or operate with less efficient technology. Meanwhile, this same effect may hamper all American businesses competing abroad, raising the cost of American goods and making them less competitive against foreign companies that aren’t subject to such a tax.

It may not be simple for a token tax to reduce AI’s harmful effects

Suppose proponents want to use taxes as a tool to reduce the harmful effects of AI, including job displacement. These concerns deserve timely, effective solutions that may be difficult to accomplish with a token tax.

A common argument is that the tax should be applied to the end use only, hitting providers or consumers when prompting AI, but excluding business uses. However, that tax design would miss upstream externalities. By exempting business uses, the tax would miss AI use within companies or between companies for functions where human labor is replaced by AI.

Other proposals center around an intermediate transaction token tax on business AI use. This would target AI uses including inventory management, document processing, creative work, and customer service bots. There are downstream consequences to consider here too. Rather than discouraging automation and lower utility usage, a token tax applied along the production chain could primarily show up as lower profits and deadweight loss, where companies maintain their current AI usage in exchange for less available revenue that could be used for future hiring and investment. Alternatively, an intermediate token tax could result in tax compounding. This means if a token tax increases costs along the production process, prices on final goods could rise without any compensatory increase to jobs or decrease in business utility usage. As a result, taxes aren’t borne by the businesses, but passed along to the final consumer through higher prices.

Rather than pivoting away from AI, businesses may continue on their current trajectory, gambling that AI will generate future gains that offset current costs, rather than make a meaningful change away from AI. In short, how business behavior changes in response to a token tax could be quite unpredictable.

Challenges raising revenue with a token tax

Returning to the revenue question at the start of the piece: Can a token tax effectively capture revenue and help offset potential erosion to the income tax base? There is reason to believe token taxes cannot because they are challenging to define and cumbersome to validate.

In general, it is best to establish a specific definition for an economic activity before it is taxed; a token is not a well-defined economic activity. Right now, tokens function as non-standard units that aren’t uniformly tied to consumption or output, making them an unsuitable target for taxation. With small tweaks, the same amount of work can be done by the same model but with drastically different token use. As AI companies work to minimize token use and improve back-end efficiency, they would be able to achieve the same results with fewer tokens. A use-based token tax may very well then collect less money over time, even if AI usage continues to grow.

Even if we were able to define what comprises a token as a taxable unit, validating its use would remain difficult. Some have argued for tax rates by model, with each getting its own tax rate. This would require distinct validation for each model type, and each model change or update may necessitate a new tax assessment. This would be fodder for endless tax gaming, not to mention incredibly difficult to accurately administer. Outside users and hyperscalers themselves are already leveraging ‘loopholes’ in how tokens are counted to minimize use. Even without willful tax avoidance, future innovations and market changes could erode the base of a token tax and limit the amount of revenue raised.

A path forward that makes more cents

To be clear: the tax system should be updated no matter which path AI deployment takes. However, lawmakers should focus first and foremost on generating revenue through reforms to existing taxes and proposals with greater ease of administration. A robust plan for AI taxation would target corporations and the avenues that wealthy individuals lean on to shrink their tax burden, include clear incentives for responsible use, and be paired with meaningful AI regulation.

At the top of the list should be an overhaul of the treatment of business profits. A token tax misses the bigger picture. Most glaringly, a token tax would leave increased profits from upstream operations like model development or chipmaking, the most profitable parts of the AI race, untouched. Additionally, token taxes wouldn’t generally scale with business profits, which, with AI’s help, can grow without hiring any workers. These issues would be better solved through broader corporate tax reforms.

Policymakers also need to look at personal capital and income – including closing the step-up-in-basis loophole, security-backed lines of credit, and pass-through business taxation – to help compensate for gains from AI on the individual side.

A good state and local tax step would be re-examining preferential tax breaks for data centers, the revenue losses from which have ballooned while providing relatively few permanent jobs. This would make the tax treatment of data centers neutral to that of other businesses and recover billions of dollars for state and local governments. This money could then be used to modernize the grid.

To preserve tax revenue, lawmakers could also consider applying a broad-based consumption tax. While this wouldn’t target AI specifically, it would help maintain federal tax revenues for important programs if AI contributes to mass layoffs. These taxes can be progressive, would tap into a more stable stream of revenue than token taxes, and capture the ‘value’ added to a good by AI as well as any subsequent technologies. If we are concerned about the erosion of a wage tax base, a turn back to consumption is sensible.

It’s possible to capture revenue from AI-driven growth without turning to untested token tax proposals. While intriguing at face value, a token tax could be regressive, difficult to implement, and likely ineffective at raising revenue. Reforms should focus on tax proposals that can better withstand whatever the future of AI development has in store.