America's first AI tax: taxing tokens to fund the jobs AI takes

This week, three House Democrats introduced the first serious federal proposal to tax AI usage itself — a levy on the tokens companies sell and the revenue they take in, bankrolling a New Deal-style jobs program. The bill has essentially no chance of passing a Republican-controlled House. It matters anyway: it is the moment Washington started trying to put a price on AI-driven displacement, and the fight over how to tax AI is fast becoming the Democratic Party's liveliest policy argument heading into the midterms and 2028. We covered the rollout in this morning's brief — AI Brief: House Democrats propose AI tax to fund worker protections.

What the bill actually does

The AI Tax and Work Protection Act, unveiled Thursday by Rep. Greg Casar (D-TX), chair of the Congressional Progressive Caucus, with Reps. Valerie Foushee (D-NC) and Sara Jacobs (D-CA), is deliberately simple in structure. Big AI companies pay a tax calculated on the higher of two amounts: the value of the tokens they sell, or the revenue they generate from AI products. A token, as Bloomberg Tax's explainer notes, is the unit of data — a word, an image, an audio clip — that a language model processes, which makes the tax base essentially metered usage. The rate is not fixed: it rises and falls with the national unemployment rate, so the program automatically grows when job losses grow.

The revenue funds a new Work Protection Administration, explicitly modeled on the Works Progress Administration Franklin Roosevelt created in 1935. Grants would flow to states, cities, tribes, localities, nonprofits, unions and educational institutions to create jobs in housing, infrastructure, child care and elder care. Two clauses are worth reading twice. First, for open-weight models, the tax falls on whichever company deploys the model to reduce workforce costs — a direct attempt to reach the AI that can't be metered at the source. Second, the bill would cancel out the tax advantage companies pick up when they replace wages with software, which sponsors describe as the tax break big companies currently get when they automate a job away.

"This bill says: we will not let AI billionaires get rich by putting you out of work," Casar said in announcing the legislation, flanked by AFT president Randi Weingarten and economic expert Gene Sperling. "Right now, the path we are on is clear: AI will turn a couple of billionaires into trillionaires but leave millions without work." The bill has the backing of AFT, AFSCME, Groundwork Action and Demand Progress Action.

The backdrop: returns that haven't arrived, anxiety that has

The bill is a bet that the labor economists warning about AI are right. Its sponsors lean on Anthropic CEO Dario Amodei's repeated prediction of unemployment not seen since the Great Depression, and on a visible run of AI-attributed layoffs at tech companies, law firms and banks. The other half of the argument is that the promised returns are still mostly missing. In PwC's 29th Global CEO Survey of about 4,500 leaders, only roughly a third of CEOs reported revenue gains, cost reductions, or both from AI; the follow-up snapshot this summer found the aggregate picture essentially unchanged, with 39 percent of CEOs reporting a positive AI impact on revenue or costs and 16 percent reporting negative impacts. PwC's separate work on AI return on investment finds the gains concentrating hard: just 20 percent of companies capture 74 percent of all AI-driven value. The technology is real; the distribution is not.

That asymmetry is what makes the politics combustible. Gallup polling this year found 7 in 10 Americans oppose data center construction near them, and the share of Americans who believe AI does more harm than good jumped from 31 percent last year to 39 percent this year. Meanwhile the stock market's AI run has made Elon Musk the world's first trillionaire, and data center construction costs rose an estimated $130 billion in the last year alone. Progressives look at that spread — record wealth creation on one side, a majority of CEOs unable to show returns on the other, public anxiety climbing — and see a political gift and a policy emergency at the same time.

The Democratic menu: half a dozen ways to tax the machine

Casar's bill is not the first Democratic AI tax idea; it is the most ambitious of a pile. The intra-party menu now includes: Sen. Bernie Sanders' American AI Sovereign Wealth Fund Act, which would impose a one-time 50 percent tax on the stock of leading AI companies to seed a $7 trillion sovereign wealth fund; a Sanders–Ocasio-Cortez push for a nationwide moratorium on new data center construction; Sen. Ron Wyden's proposal, unveiled the same day as Casar's, to strip data centers of tax incentives like Opportunity Zone benefits and bonus depreciation and add a low single-digit excise tax on their revenues; Sen. Elizabeth Warren's teased tax on the energy data centers consume; Rep. Ro Khanna's "Data Center Bill of Rights"; and Sen. Ruben Gallego's plan to end data center tax breaks and make companies pay for the grid upgrades they force.

That is roughly half a dozen competing answers to the same question — what should the government do about the concentrated gains and diffuse costs of AI? — and the differences are not cosmetic. Sanders wants to seize equity and halt construction. Wyden wants to tax the physical plant. Warren wants to tax the electricity. Casar wants to tax the usage itself, which his earlier op-ed in The American Prospect argued would let federal revenue scale with AI adoption rather than lag it. The split is a feature of a party still forming its position with an eye on 2028, as NOTUS put it — AI could be an area of division within the party in the run-up to the next presidential race. "Democrats need a vision here," said Rob Flaherty, a strategist who worked in the Biden White House. "We have to figure out what we are for fast — capture this moment, or Republicans will do it for you."

Republicans are not idle on AI either, which sharpens the contrast. Their frame is security and competition rather than distribution: Rep. Jay Obernolte's bill with Rep. Lori Trahan would preempt state AI regulations and give the federal government power to restrict catastrophic-risk models, with a committee vote floated for September, and the White House sent major labs pre-release review guidelines this month. The administration's posture is protectionist in a different direction — MIT Technology Review recently detailed how Trump's AI protectionism has reached robotics. The emerging shape of 2026 politics is both parties agreeing Washington must act on AI, and disagreeing completely about what kind of act.

The hard parts nobody has solved

The bill's weaknesses are real, and its supporters know it. Start with the definition. What counts as an "AI product" at the margin — a customer-support chatbot, a spreadsheet with an inference button, a code assistant embedded in an IDE? The tax base is "the higher of" tokens or product revenue, which invites companies to structure pricing so neither number is comfortable to audit. Tokens are at least objectively metered, but they are an input, not a profit: two companies selling the same tokens can have wildly different margins, and a usage tax falls hardest on the cheapest, most efficient providers. Tax lawyers have already started probing where the definition of a taxable token begins and ends, and every answer produces a loophole.

Then there is pass-through. Skeptics of token taxes warn the cost doesn't stay with the labs — it flows to the businesses and consumers buying the tokens, which makes the tax a levy on AI adoption itself. That is the deepest critique: if AI is a general-purpose technology whose benefits are still mostly ahead of it, taxing its metered use is a tax on the future, paid partly by the very workers the bill means to help, in the form of slower diffusion and higher prices. The Brown University computer scientist Serena Booth, quoted in the bill's rollout, framed the wager honestly: if Silicon Valley is right about AI replacing millions of jobs, we need a plan; if they're wrong, the tax raises little revenue.

The open-weight clause shows how hard it is to make the tax reach the actual disruption. Moonshot, Meta and Mistral don't sell tokens; they publish weights. The bill's answer — tax the company that deploys the model to cut workforce costs — requires proving deployment intent to reduce headcount, which is precisely the kind of fact no auditor can establish. A company can adopt an open-weight model for "efficiency" and lay people off for "restructuring." And a model deployed by a foreign company to cut costs outside US jurisdiction is entirely out of reach. The clause is a flag planted on a hill the IRS cannot hold.

Finally, the politics: with Republicans controlling the House, the bill has no path, and industry will frame any AI tax as a China subsidy. That critique is not wrong — every dollar of tax on US AI development is a relative advantage for labs that don't pay it. The sponsors' answer is that the alternative — doing nothing while displacement happens — is not neutral either. This is a marker bill, and its function is to force the distribution question onto the federal table with a concrete mechanism, the way the first minimum-wage bills forced a different question in a different century.

Why it matters anyway

Dead bills can still change what is thinkable. In 2017, Bill Gates floated the idea of taxing robots and was widely treated as a curiosity; the European Parliament rejected a robot tax the same year. Nine years later, a usage tax on AI is a House bill with major union backing, a Senate counterpart in Sanders' equity tax, and a tax-writing Democrat in Wyden proposing his own variant on the same day. The arc from joke to caucus position to legislation is the real news here — the Overton window on taxing automation has moved from closed to wide open in under a decade.

The design debate is also genuinely useful even without a vote. A token tax, a data-center excise, an energy levy and an equity tax are four different theories of where AI's taxable value lives: in usage, in physical capital, in power, or in windfall. Democrats are effectively running a live experiment in which theory survives contact with tax lawyers, and the winner will be the template for the 2027 Congress if Democrats take the House in November. Casar's version has one structural advantage the others lack: a usage tax scales automatically with adoption, which is exactly the property a safety net needs if the doomsday scenario arrives, and a tax that raises little revenue if it doesn't.

That is the case for treating the bill as more than performance. "The most direct policy response to mass unemployment is mass employment," said Asad Ramzanali of the Vanderbilt Policy Accelerator. "That was true in the 1930s, when FDR launched the Works Progress Administration, and it's true today." The WPA comparison is doing real work here: it is an argument that the answer to machine-driven displacement is not redistribution checks but public employment — jobs in housing, care and infrastructure, precisely what the Work Protection Administration would fund. Whether that is economically wise or politically possible is arguable. That it is now a live federal proposal, with a tax base attached, is not.

What to watch

Three things, in order of likelihood. First, whether the bill gets any hearing before the midterms — unlikely, but a single hearing would legitimize the tax-base debate overnight. Second, which variant of the AI tax survives the Democratic primary season: watch whether 2026 candidates run on tokens, data centers, energy or Sanders' equity tax, because the primary winner becomes the 2027 default. Third, the technical fight over the definition — how "AI product" and "token value" survive real scrutiny will determine whether any future version is collectible, and whether the open-weight clause gets a workable enforcement mechanism or gets dropped. The state level bears watching too: more than a dozen states have considered data center moratoriums, and state action could outrun the federal debate entirely. A bill that cannot pass this year can still define the terms of the fight that does.

Should the government tax AI usage to fund worker protections — or would the cost just flow through to customers? Tell us in the comments.

Sources: Rep. Casar press release · NBC News · Politico · Bloomberg Tax · NOTUS · PwC CEO Survey Snapshot · PwC — Decoding ROI from AI · Gallup — Americans cool toward AI · Gallup — data center opposition · Sanders press release · Wyden white paper · Warren — Why We Need to Tax AI (Time) · Casar — The American Prospect op-ed · Axios — Amodei on unemployment · Quartz — Bill Gates on robot taxes · Politico — Obernolte-Trahan AI bill · MIT Technology Review — Trump's AI protectionism