Taxing AI to Help Workers Sounds Good, But Public Deserves More

Rep. Greg Casar’s (D-Texas) proposed AI Tax and Work Protection Act has a solid premise: Firms capturing gains from AI-driven labor force reductions should carry some of its social costs. But the proposed tax on AI tokens illustrates why taxing artificial intelligence use is a poor proxy for taxing automation itself. Lawmakers should instead give the public an equity stake in the firms that capture its economic gains.

Casar’s proposal is an attempt to correct an imbalance by levying a tax on AI companies and using the proceeds to support workers. It’s a sound idea. To the extent AI allows firms to replace human labor while shifting some of the resulting costs — unemployment, lost tax revenue, and the like — onto workers and society writ large, those costs resemble an externality.

We’ve seen this type of tax before: Governments tax cigarettes in part for their negative health impacts and tax gasoline to finance the wear and tear on infrastructure vehicles cause. And we’ve spilled much ink debating carbon taxes as a method of internalizing carbon emissions in manufacturing.

The problem with Casar’s bill is in the particulars, as it doesn’t really tax worker displacement. It imposes a levy equal to the greater of two amounts: a percentage of the fair market value of tokens processed in covered transactions, or a percentage of the revenue and related-party value associated with those transactions.

The tax rate is keyed to the unemployment rate, which produces an appealing feedback loop: As AI contributes to greater unemployment, the tax rises and generates more money to put people back to work.

Unemployment may spike for various reasons, though. Absent some mechanism to account for that, AI companies may end up bearing the costs of labor market contractions they didn’t cause. The proposed bill includes a safety valve, allowing the Treasury Department to adjust the escalator where unemployment traces to war, pandemic, or an unrelated shock.

But that requires someone at the Treasury to determine — in real time and on a political calendar — how much of a given labor contraction is because of AI. That’s a causal allocation the tax code doesn’t otherwise have much occasion to make. More problematically, Casar’s proposal requires the Treasury to put a value on an extraordinarily unstable unit of measurement. A token is a technical unit, not a standardized commodity.

AI services increasingly encompass text, code, images, audio, and video. And tokens aren’t necessarily sold individually in an arm’s-length market that would supply the Treasury with a convenient unit price.

Rather than trying to determine the value (and implicitly, the social cost) of each instance of AI use, governments could capture a continuing share of the economic returns AI produces by taking an equity stake in the companies engaged in producing it. Professors Jeremy Bearer-Friend and Sarah Polcz proposed such an approach, and Sen. Bernie Sanders (I-Vt.) recently incorporated a more aggressive version into legislation aimed at creating a sovereign wealth fund.

The underlying architecture deserves serious consideration. Instead of trying to tax the meter, government can own a piece of the franchise.

The recent embrace of public ownership makes Bearer-Friend and Polcz’s approach less exotic than it may have sounded even five years ago. An equity-based approach ducks much of the mismatch inherent in trying to tax AI use as a proxy for AI-driven automation. Under Casar’s bill, a company selling access to a foundation model can generate covered transactions whether its customers use that model to displace workers, augment them, or do something with no meaningful effect on labor.

If we were to focus solely on internalizing the externalities of automation, we might consider taxing labor displacement itself. But that creates an even less enviable position for the Treasury, which would need to determine whether a given employee was terminated because of AI, conventional software, outsourcing, falling demand, or an ordinary business decision. And what about the employee who is never terminated because they are simply never hired — are we prepared to tax firms for failing to expand their labor force at the rate a schedule suggests they should?

An equity-based approach sidesteps much of that exercise. Bearer-Friend and Polcz propose taxing generative AI companies in kind — with firms transferring equity to the government rather than shelling out cash.

If AI merely augments workers and produces modest gains, the value captured by the public would remain equally modest. If it instead produces extraordinary gains, the public would participate in the resulting increase in value.

Equity is ultimately a proxy, too. Share price tracks profit, not AI-generated pink slips. A firm could labor-augment its way to a fortune and hand the public a windfall for causing no harm whatsoever. But it’s a better proxy than token usage for the portion of economic gains potentially associated with AI-driven labor displacement, because it scales with the size of economic reallocation and asks the government only to value a company once rather than to value a unit of computation continually.

Sanders took that architecture considerably further by proposing that covered AI companies transfer a 50% stake to a federal sovereign wealth fund. Congress needn’t conclude that Uncle Sam should own half of an AI company to adopt the underlying idea.

A smaller equity assessment, limited to the largest firms and held through an independently managed investment vehicle, could allow the public to participate in AI’s upside without tasking a staff-reduced Treasury Department with continually valuing and revaluing tokens or determining whether a prompt cost a job.

Congress would still need to determine which firms are covered, how private companies are valued to determine equity shares, how public ownership rights are executed, and whether the government should be a passive shareholder. The relevant question is whether those problems are more tractable than trying to identify the economic and labor consequences of billions of individual AI transactions.

Andrew Leahey is an assistant professor of law at Drexel Kline School of Law, where he teaches classes on tax, technology, and regulation. Follow him on Mastodon at @andrew@esq.social.

Read More Technically Speaking

Learn more about Bloomberg Tax or Log In to keep reading: