The Case for an AI Token Tax

May 28, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 28, 2026

MAIN STORY

The Case for an AI Token Tax (and the case against it)

The idea of taxing AI tokens has been around for a while, but it's getting serious political momentum now — with a Senate candidate releasing a comprehensive worker protection plan featuring a token tax, and Elizabeth Warren publishing an op-ed in Time calling for taxing AI directly. The industry's instinct will be to dismiss this out of hand, but that's the wrong call. If we are genuinely entering a categorically different era, it follows that policies designed for the old one may simply not fit anymore — and the healthiest stance is open engagement, not reflexive rejection.

DO WE NEED AN AI TAX

The Political Moment
Warren, a Senate candidate, and Mark Cuban are all calling for it.
Michigan Senate candidate Mallory McMorrow released a worker protection plan this week that includes a token tax — described as "a modest fee on commercial companies' AI usage" that, at a fraction of a cent per token, becomes a meaningful funding stream without raising taxes on any individual worker. The same day, Elizabeth Warren published her Time op-ed arguing that taxing AI is how we ensure its winnings benefit all Americans rather than only the wealthy — starting with an excise tax on data center energy use and leaving the door open to "bigger and bolder proposals." Also in the mix: DuckDuckGo's Gabriel Weinberg, who argued at the end of April for a 10% surcharge on token charges — roughly matching what employers already pay in payroll taxes — held in a lockbox for displaced workers. And Dario Amodei floated something similar to Axios last year: a 3% revenue levy on model usage that he acknowledged wasn't in his economic interest but called "a reasonable solution."

 The First Principles Case For
When AI does the work, the tax base shifts — tokens could fix that.
Across the OECD, the average single worker costs about 35% of labor costs in taxes. When a human does a job, that income flows through income and payroll taxes. When an AI agent does the same job, the value shows up as lower costs, higher margins, or capital gains — taxed less, or not at all. The IMF flagged this as far back as 2024: labor substitution could erode the income tax base if capital income is taxed at lower rates than labor. Tokens are an imperfect proxy for AI labor, but they're observable and already metered by providers — which makes them administratively tractable in a way that other measures of AI output are not. The deeper philosophical argument: if the obligation to fund public goods has historically attached to productive human labor, and AI agents increasingly do that labor, maybe the obligation should follow the capacity, not just the toil.

The Case Against
Tokens are a terrible proxy for value — and the math doesn't hold up.
David Friedman wrote a detailed response to Mark Cuban's proposal laying out several serious problems. First, tokens are an extremely poor proxy for economic value — a million tokens might generate spam or perform high-value legal analysis, and many tokens consumed aren't in the service of work at all. Second, there's what Friedman calls the tokenizer endogeneity problem: different providers tokenize the same content differently (Mandarin runs 2–3x more tokens than English; some low-resource languages run 10–15x more), so a flat per-token tax discriminates between providers on a basis completely unrelated to any externality being taxed. Third, token prices have fallen roughly 200x annually for two years running — a fixed tax would either become confiscatory as prices fall or collapse in revenue if indexed downward. And fourth, a tax at the provider level is structurally a subsidy for foreign inference: American enterprise customers would increasingly route through non-US API providers to avoid it.

The Experimentation Problem
A token tax would deepen exactly the dynamic already hurting AI adoption.
We're already in a period where token prices are rising and companies are restricting experimentation to use cases with clear, known ROI — prioritizing efficiency AI (cheaper customer support, faster analyst decks) over the harder, higher-upside work of discovering genuinely new applications. A flat token tax would make that worse, and it wouldn't fall equally: large firms would negotiate discounts, reserve capacity, self-host, and amortize experimentation costs over massive revenue bases. The result would be significant entrenchment of incumbents and a further tax on exactly the kind of open-ended exploration that surfaces the highest-value uses.

Where This Lands
The first principles question is worth taking seriously; the proposals so far, less so.
The underlying logic — that a world where productive capacity shifts from human labor to agentic labor might require rethinking how we structure the tax base — is coherent and worth engaging with. But a flat per-token tax as currently proposed has serious design problems that the academic work (including the Brookings paper) is already flagging. The Brookings authors suggest the right answer in the current phase isn't a production-side token tax but a consumption-side tax integrated into VAT and sales tax infrastructure, with B2B use exempted to avoid cascading distortions. That's a much more nuanced conversation than "tax the tokens" — and it's one the industry should be willing to show up for rather than dismiss.