Rep. Greg Casar (D-TX), with Reps. Valerie Foushee (D-NC) and Sara Jacobs (D-CA) · 2026
Rep. Greg Casar — AI Tax and Work Protection Act (H.R. 10044)
Casar AI Tax
Introduced August 6, 2026 and referred jointly to Education and Workforce and to Ways and Means, this is the first bill to put actual statutory text behind the idea of taxing AI computation and routing the proceeds into public jobs. Title I adds a new subchapter to the Internal Revenue Code imposing an excise tax on foundation model developers, calculated as the greater of a percentage of the fair market value of tokens processed or a percentage of revenue from AI services — with rates that automatically escalate as unemployment rises. Titles II through IV take 100 percent of that revenue into a trust fund and spend it through a Work Protection Administration inside the Department of Labor, whose acronym is a deliberate echo of the New Deal's Works Progress Administration. The WPA awards competitive grants to states, localities, tribes, school districts, universities, and nonprofits to hire people directly into 17 enumerated categories of public work — child care, education, health, elder and disability care, housing, homelessness services, violence prevention, scientific research, infrastructure, arts and libraries, conservation, wildfire and disaster preparedness, facility repair, parks, local journalism, and WIOA training. The bill is unusual in coupling the tax trigger to the labor market it is meant to protect: the levy is imposed only on models used commercially or used to shrink a workforce, and its rate is indexed to the U-4 unemployment measure.
Key Provisions
New IRC §4491 excise tax on 'covered persons' — anyone who develops a foundation model, sells access to one, or modifies an open-weight model, and who either earns revenue from it or uses it to reduce their own workforce
Tax equals the greater of (a) fair market value of tokens processed × the applicable token percentage, or (b) AI service revenue plus related-party transaction value × the applicable transaction percentage — a belt-and-suspenders design to stop firms from shifting value between the two bases
Rates indexed to unemployment: 2% on tokens / 3% on transactions when U-4 is at or below 5%, rising point-for-point between 5% and 7%, and at double rate above 7% — so the tax hits hardest exactly when displacement is worst
Foundation model defined by a 10^25 FLOP training-compute threshold, with authority for the Secretary to reset the threshold annually to track comparable capability
Covered transactions include internal use: a company that runs a model on its own workers' jobs owes the tax even if it never sells access
R&D carve-out for governments, institutions of higher education, FFRDCs, and 501(c)(3) organizations; Treasury may discount unemployment attributable to war, pandemic, or other shocks unrelated to AI
Trust fund capturing 100% of the tax from FY2027 onward, spendable only on the jobs program and the administering office; up to 20% may be routed to WIOA programs
Work Protection Administration established within DOL within 90 days, awarding competitive grants that prioritize permanent, full-time jobs and state or local governments for governmental functions
17 categories of grant-funded public work, from child care and school staffing to wildfire fuel reduction, broadband construction, and local news and journalism capacity
Job quality floor: prevailing wage (or the higher of federal, state, or local minimum), FEHB-comparable health coverage, 12 weeks of paid family and medical leave at full pay, and one hour of paid any-purpose leave per 30 hours worked
Labor conditions on every grant: NLRA §7 protections, posted rights notices, majority-sign-up union recognition, and local hiring preference — with tribal governments exempt
Nondisplacement rules barring grant-funded hires that cause layoffs or partial displacement, that take over the duties of striking workers, or that replace laid-off workers who have not been offered restoration
Advisory committee spanning Labor, Commerce, Education, the Federal Reserve, the NLRB, NSF, CEA, OSTP, every state governor, large-union representatives, and two labor economists — exempted from the Federal Advisory Committee Act
$20M per year for FY2027–2031 for the Bureau of Labor Statistics to measure not just AI-driven job loss but job degradation — falling pay, reduced hours, and conversion of permanent roles to temporary ones
Tax applies only to covered transactions occurring more than one year after enactment
Regulatory Philosophy
Tax the substitution, then buy back the jobs. Casar's bill declines to regulate AI as a safety problem at all — there is no testing regime, no licensing, no liability framework, and no preemption language anywhere in the text. It treats AI purely as an economic event in which capital is being substituted for labor, and it responds with the two most conventional tools in the fiscal toolkit: an excise tax on the substituting technology and a direct federal employment program funded by the proceeds. The design embeds a countercyclical automatic stabilizer — because rates rise with unemployment, the revenue and the jobs program both expand precisely when the labor market deteriorates, without requiring Congress to act again. The job quality conditions reveal the second half of the theory: the bill is not only replacing lost employment but attempting to set a floor on what replacement employment looks like, using federal grant conditions to spread prevailing wages, paid leave, and union access into sectors where they are rare.
Tokens processed and AI service revenue, assessed on the model developer
Why it lands there
The distinction that separates this from the Bores and Khanna token taxes is who the statute names. Section 4491 assesses the covered person — whoever develops a foundation model, sells access to one, or modifies an open-weight model — rather than the buyer, which makes it the closest thing in this landscape to a tax whose legal incidence sits on the frontier layer itself. Economic incidence still passes to the app layer through inference prices, and the internal-use trigger pulls in any enterprise that self-hosts a model to cut headcount, making it a covered person in its own right. Training compute remains untaxed.
Two token taxes, two theories of what to do with the money
Casar and Bores converge on the same novel tax base — computation measured in tokens — and diverge completely on what it buys. Bores routes revenue to direct payments, treating displacement as an income problem to be solved with cash, and pairs the token tax with federal equity warrants so the public captures upside from the firms doing the displacing. Casar routes every dollar to wages for public work, treating displacement as an employment problem to be solved with jobs, and adds no equity claim at all. The rate structures encode the same split: Bores triggers payments on economic conditions after the fact, while Casar's rate schedule escalates continuously with U-4 unemployment so the tax itself is the trigger. Casar also has what Bores does not — an actual bill number, committee referrals, and statutory text defining the tax base, which is where the token-tax concept has always been most vulnerable.
+This is legislative text, not a white paper — a scored, referred bill with a section-by-section tax mechanism, which puts it ahead of every other AI tax proposal in the landscape, including the Bores AI Dividend and Khanna's Future Workforce Administration
+Indexing the rate to U-4 unemployment makes the tax a genuine automatic stabilizer: it stays modest while AI is complementing labor and escalates sharply only if the displacement thesis actually materializes, which answers the standard objection that automation taxes penalize productivity growth
+The dual tax base — tokens or revenue, whichever is greater — anticipates the obvious avoidance strategy of giving inference away cheaply while monetizing elsewhere, and the internal-use trigger closes the loophole that would let firms self-host models to escape a sales-based tax entirely
+Funding public goods that are chronically understaffed and structurally resistant to automation — child care, elder care, home health, wildfire mitigation, local journalism — is a coherent match of revenue source to spending target, moving labor toward work AI cannot do rather than subsidizing competition with it
+The nondisplacement and prevailing wage conditions are drawn from decades of experience with federal jobs programs and directly address the standard critique that public employment schemes undercut existing workers and depress wages
+The BLS mandate to measure job degradation, not just job loss, targets the most likely real-world outcome of AI adoption — hours cut, pay reduced, permanent roles made contingent — which current statistics do not capture
Weaknesses
From the perspective of political opposition
−Taxing tokens is taxing an accounting fiction — 'fair market value of tokens processed' has no established measure, tokenization schemes are chosen by the developer, and a company can restructure batching, context windows, or model architecture to shrink its taxable unit count without changing a thing about what the model does
−The 10^25 FLOP threshold and the internal-use trigger together create a large avoidance surface: train just under the line, fine-tune abroad, buy inference from a foreign provider outside U.S. taxing jurisdiction, or characterize a workforce reduction as a reorganization unrelated to the model that quietly performs the work
−Escalating the tax as unemployment rises is procyclical in the wrong direction — it raises the cost of the most productive technology in the economy exactly when the economy is weakest, and the war-and-pandemic escape hatch gives Treasury broad discretion to gut the escalator whenever it becomes politically inconvenient
−A new federal agency standing up a competitive grant program across 17 program areas and hiring at prevailing wage with federal-grade health benefits will spend years on rulemaking and application review before the first worker is hired, while the displacement it responds to would already have occurred
−The bill is silent on every AI governance question other than money — no safety testing, no transparency, no liability, no child protection, no preemption stance — so a Congress that passed it would have taxed frontier AI without having regulated it in any respect
−The drafting is visibly unfinished: Title III is captioned 'Office of Job Creation' while its operative text creates a 'Work Protection Administration,' the definitions section defines a Director of an office the bill never establishes, the effective-date clause cross-references a subsection that does not contain the definition it cites, and the rate-suspension sentence in §4491(c) is grammatically incoherent
−Exempting the advisory committee from FACA while seating every governor, federal agency designees, and large-union representatives creates a closed body allocating billions in employment grants with no open-meeting or public-record obligations