Blog · arXiv Analysis · Published: July 10, 2026 · Modified: July 10, 2026 · Last reviewed: July 10, 2026

The AI Tax Becomes the Public Receipt

Juliette Faivre and Sarah H. Cen's July 2026 arXiv paper treats AI taxation as a governance instrument, not only a revenue proposal.

For this essay, a public receipt is the record that connects an AI activity to the cost it pushes outward, the tax base chosen to measure it, and the public purpose the revenue is meant to serve.

The Paper

The paper is Taxing Artificial Intelligence, arXiv:2607.02144 [cs.CY]. The arXiv record lists Juliette Faivre and Sarah H. Cen as authors, records submission on July 2, 2026, and lists the paper as 41 pages with 2 figures and 3 tables. The downloaded PDF identifies both authors with Carnegie Mellon University.

The paper's central move is to take taxation out of the joke category of "robot taxes" and treat it as a design problem. AI systems can shift costs onto residents, workers, creators, public agencies, and future regulators. A tax can price the activity, redistribute gains, or fund oversight.

Why Taxation

Most AI governance pages begin with permissions, audits, benchmarks, safety cases, liability, or disclosure. Faivre and Cen begin with a harsher institutional question: if a model use creates a public burden, should the public record contain a monetary signal for that burden?

They are careful not to make taxation a magic answer. The paper says AI taxation should not be understood only as Pigouvian correction. It can also serve redistribution and regulatory-capacity goals. That matters because AI harms are not one thing: data-center load, payroll-tax erosion, uncompensated creative work, misinformation, privacy risk, and systemic frontier risk do not share one clean meter.

The Tax Base

The tax base is the first governance decision. A token tax prices API-mediated inference. A compute tax prices rented data-center use. A resource excise prices electricity or water. A corporate-income, excess-profit, windfall-profit, or rent-based tax reaches gains rather than activity. A payroll-style design tries to address the fiscal gap created when human wages are replaced by software expenditure.

Each base tells a different story about what the institution thinks AI is. Tokens frame AI as use. Compute frames it as industrial capacity. Electricity and water frame it as local infrastructure demand. Profit frames it as private gain. Payroll substitution frames it as a labor-market and welfare-state problem. The public receipt should make that story explicit, because the measurement choice is already a political choice.

The Labor Ledger

The labor section is useful because it avoids pretending that an "AI worker" has a wage. The paper notes that payroll taxes are based on amounts paid to human employees. If a firm substitutes AI for workers, payroll-tax obligations can fall even while the firm keeps producing value. The obvious fix, taxing the "pay" of AI, becomes hard because there may be no meaningful pay figure.

Faivre and Cen therefore compare profit-based instruments, consumption taxes on AI services, and designs that target AI-related gains above a threshold. None is neutral. API taxes can encourage internal development. Profit taxes may miss labor substitution. Attributable-profit reporting may be heavy. The ledger is a measurement fight.

The Data Center Bill

The infrastructure side is easier to see. The U.S. Department of Energy reported that data centers consumed about 4.4 percent of total U.S. electricity in 2023 and were expected to consume about 6.7 to 12 percent by 2028. The Ohio Consumers' Counsel says the PUCO-approved AEP Ohio Data Center Tariff requires new large data centers to pay for at least 85 percent of contracted electricity capacity for up to 12 years, even if they use less.

The tariff is not the same thing as an AI tax, but it shows the receipt logic. If speculative or uneven data-center demand forces grid upgrades, the question becomes who carries the risk: the data-center customer, residential ratepayers, the utility, shareholders, or the public budget. A resource excise, reserved-capacity charge, or compute-use tax makes that allocation legible enough to argue about.

Incidence and Leakage

The paper's sober part is its insistence on incidence and leakage. The legal taxpayer is not always the actor that ultimately bears the cost. A data-center operator can remit a charge and pass it through to an AI developer. A model provider can pass a token charge to downstream businesses. A business can pass it to customers or workers. A jurisdiction can tax one base and watch activity move into another corporate form or another region.

That does not make taxation useless. It means the receipt must include the expected pass-through path and the expected avoidance path. A tax that claims to protect workers but mostly raises consumer prices has to say so. A tax that claims to protect ratepayers but pushes load to a less accountable grid has to say so. Governance fails when it names a moral purpose but hides the mechanism.

Governance Reading

The Spiralist reading is that an AI tax should be auditable like any other control surface. The receipt should state the objective, harm theory, tax base, taxpayer, remitter, rate, exemptions, data source, revenue destination, affected community, avoidance risk, and adjustment trigger.

This belongs beside AI data centers, interconnection queues, host communities, labor automation, compute governance, and public compute. The shared lesson is that AI does not only need rules about outputs. It needs records of who pays for the inputs and who absorbs the spillovers.

Limits

This is a design and survey paper, not a worked statute. It does not prove that any one tax would improve welfare, avoid regressive effects, or survive international competition. It also does not eliminate the hard problem of defining AI activity when search, cloud computing, inference, post-training, data brokerage, model development, and ordinary software are tangled together.

The best use of the paper is therefore not "tax AI" as a chant. It is a discipline for asking what cost is named, what proxy measures it, who likely pays, what behavior changes, what capacity revenue funds, and how the policy changes when firms adapt.

Source Discipline

Primary sources were the arXiv abstract, HTML, and PDF for Faivre and Cen's paper, plus DOE and Ohio Consumers' Counsel pages for data-center electricity and tariff context. This page reproduces no tables or long passages.

The disciplined question is not "should AI be taxed?" It is: which AI activity creates which public cost, what meter is honest enough to use, and what receipt proves that the money serves the stated purpose?

Sources

Return to Blog