Korinek & Lockwood — Public Finance in the Age of AI: A Primer

Summary

"Public Finance in the Age of AI: A Primer" (NBER Working Paper 34873, February 2026; prepared for the Brookings Center on Regulation and Markets and forthcoming in The Economics of Transformative AI, NBER/University of Chicago Press) is Anton Korinek (University of Virginia, Brookings, NBER) and Lee Lockwood's (University of Virginia, NBER) framework for how optimal-taxation theory should adapt as AI progressively displaces human labor and, eventually, human production itself.[1] It is the direct instrument-design companion to Korinek & Vipra's market-concentration diagnosis and to Korinek & Stiglitz's long-run "innovator rents" theory — where those two papers ask whether AI generates rent, this one asks how to tax it once it does, and maps every AI-specific tax proposal currently circulating (robot taxes, compute taxes, token taxes, digital-services taxes, sovereign wealth funds, windfall clauses, universal basic capital) onto a single, rent-versus-normal-return distinction the wiki already treats as foundational for LVT and ACE/cash-flow taxation.

The Core Distinction — and Its Georgist Structure

The paper's entire analytical apparatus rests on one line, stated as a design principle rather than a Georgist argument: "taxing rents involves no efficiency cost, while taxing normal returns distorts capital accumulation."[1] The authors define rent conventionally — "returns above what is necessary to induce an activity such as a particular investment" — and draw the practical implication directly: "where feasible, tax super-normal returns (rents) heavily since there is no distortion."[1] Structurally, this is the law of rent and the logic behind ATCOR generalized to any fixed factor, arrived at by two public-finance economists with no Georgist framing.

The paper names land as its own leading example of a taxable fixed factor. Listing the categories of return that theory says should be taxed hardest, the authors write: "Tax negative externalities and the rents of fixed factors (such as unimproved land)... [and] where feasible, tax super-normal returns (rents) heavily since there is no distortion."[1] Later, discussing the AI transition specifically: "Fixed factor taxation (e.g., land, spectrum rights) gains importance as identifying and taxing economic rents becomes more valuable."[1] The wiki's own rent gradient — land as the clean case, everything else contested — is, in this paper, the authors' own organizing principle for where AI taxation should head as labor's share erodes: toward exactly the fixed-factor logic land already satisfies perfectly.

The mechanism for isolating rent is the cash-flow tax the wiki already covers. The authors note that "a cash-flow tax with full expensing and no interest deductions can isolate economic rents for taxation," immediately adding the same caveat the wiki's cash-flow tax page and the quasi-rents objection both carry: this requires "symmetric treatment of gains and losses, which is difficult to implement," and faces "the conceptual challenge that high realized returns under uncertainty may reflect compensation for risk-bearing rather than true rents."[1]

The Two-Stage Framework

The paper's substantive contribution is a formal two-stage model of how the optimal mix of tax instruments should shift as AI advances:

• Stage 1 ("the twilight of labor"): AI reduces labor's role in production but humans remain the primary consumers. The authors show formally that "maximum labor-tax revenue as a share of output approaches zero as the capital share approaches one," so consumption taxation becomes the primary revenue instrument, and — because the Atkinson–Stiglitz case for uniform commodity taxation depends on labor being the binding distortion — differential commodity taxation regains relevance once labor stops constraining the tax system.[1] Fixed-factor (rent) taxation "gains importance" in this stage for the same reason.
• Stage 2 ("the AGI-dominated economy"): if autonomous AGI systems both produce most economic value and absorb a growing share of resources for non-human purposes, even consumption taxation becomes inadequate — nothing is being consumed by a taxable human. The authors frame taxing AGI capital as an optimal harvesting problem: how much of AGI's growing capital stock should society "harvest" for current human benefit versus let accumulate for the future. In their baseline specification, the optimal tax rate on AGI capital equals the human discount rate (which they note is "commonly estimated around 4 percent"), reflecting a pure trade-off between current human consumption and future growth.[1] They explicitly carry the rent lesson into this stage too: "if part of the market returns are rents from fixed factors, or if AGI development is characterized by significant market concentration, then the returns to AGI capital may substantially exceed the normal competitive return... they could be taxed at rates approaching 100 percent without distorting the AGI's investment decisions."[1]

Mapping the Live AI-Tax Proposals

Section 5.2 is, in effect, a graded instrument table for AI-specific taxes — directly relevant to the wiki's own Taxing Tech Rents instrument comparison, though built independently and for a different (still-speculative) target. The authors sort each proposal by which of their three core instruments (labor tax, consumption tax, capital tax) it economically resembles:

The general principle: taxes that hit AI services at the point of final consumption (token taxes, robot-services taxes, DSTs) align with sound Stage-1 design and can ride on existing VAT/sales-tax infrastructure; taxes on the capital goods themselves (compute, robots) distort investment during Stage 1 and are "only a secondary option... if consumption taxation is limited for some reason," becoming appropriate only once AGI entities, not humans, are the thing being taxed.[1] The authors separately flag that producers of compute could still owe a rent tax if they earn above-normal returns — "if producers earn significant rents above and beyond the normal rate of return... our earlier lessons on rent taxation apply: taxing rents is an economically efficient way of raising revenue" — which is the compute-layer concentration Korinek & Vipra documents (Nvidia at 92–98% of the GPU market), read through this paper's own tax-design lens.[1]

Equity-Based Alternatives to Taxation

Beyond taxation, the paper takes seriously a family of predistribution and insurance mechanisms already circulating in AI-governance debates, evaluated on the same efficiency logic:

• Sovereign wealth funds / public equity stakes in AI companies — government ownership positions that "capture AI-generated returns without distorting investment," functioning as non-distortionary revenue that scales automatically with AI productivity.
• Windfall clauses — voluntary commitments by AI companies to share exceptional returns broadly, which the authors connect to their discussion of nonprofit AGI systems that optimally deploy resources for human benefit without formal taxation.
• Universal Basic Capital (UBC) — broadly distributed equity ownership of AI companies themselves, "predistributing" AI gains through ownership rather than ex-post redistribution.[1]

The authors' case for these mechanisms is explicitly about insurance under radical uncertainty: equity positions rise automatically if AI turns out to be as transformative as optimists expect and stay modest if it does not, in contrast to a tax rate that must be set (and re-set) by legislators guessing at the scale of a technology whose trajectory nobody currently knows. They also flag a second, distinct instrument for steering rather than raising revenue — Korinek & Stiglitz's (2025) "Steering Technological Progress" argument that a tax on labor-displacing technologies (e.g., robots) can be worthwhile purely for its predistributive effect of pushing R&D toward labor-complementing designs, "besides raising revenue" — a use of an AI tax the wiki's existing instrument-comparison table does not currently carry for any tech-rent instrument.[1]

Relation to the Georgist Case — Both Ways

• The strongest available restatement of the Georgist tax principle from mainstream public finance, applied prospectively to AI. The authors did not set out to make a Georgist argument — the paper cites Atkinson–Stiglitz, Chamley–Judd, and Diamond–Mirrlees, not George — yet arrive independently at "tax the rents of fixed factors like land... heavily, since there is no distortion" as their efficiency benchmark, and use it as the yardstick for every AI-specific proposal. That a rigorous NBER optimal-taxation paper reaches for land as its own canonical fixed-factor example, while building the AI case, is outside corroboration for the wiki's central claim that the land case is uniquely clean.
• But the paper's own instrument table shows how far AI still is from land's clarity. Every AI-specific proposal the authors map — compute, token, robot, DST — is classified by its formal economic resemblance to labor/consumption/capital taxation, not by whether the base it hits is actually rent. The authors are explicit that identifying real AI rents "in practice... lies in distinguishing true rents from returns that reflect compensation for risk-bearing or that are necessary to incentivize continued innovation" — the same unresolved valuation problem the wiki carries at data-rents and the quasi-rents objection. No AI tax proposal here clears the bar an LVT clears for land.
• The paper is speculative by its own design, and says so. Both stages describe futures that may not arrive; the authors build the theory "proactively" precisely because they think institutional adaptation is cheaper before disruption than reactive crisis response, not because Stage 1 or Stage 2 is imminent or certain. The 4%-discount-rate AGI tax-rate result is explicitly "a simple specification," not a policy-ready number.

Limits

• An unrefereed NBER/Brookings working paper, though slated for a peer-reviewed NBER/ University of Chicago Press volume (The Economics of Transformative AI); the formal results (Lemma 1, Propositions 1–2, the harvesting-problem solution) are derived under stated simplifying assumptions — homothetic preferences, time-invariant heterogeneity, no borrowing constraints — that the authors themselves flag as load-bearing for the Pareto-improvement results.
• The whole framework is conditional on transformative AI actually arriving in something like the form defined (machines performing "essentially all economically valuable work"); the paper explicitly brackets the empirical debate over whether or when this happens, citing both optimistic and skeptical camps (Goldman Sachs' "GenAI: Too much spend, too little benefit," Acemoglu's 0.07%/year growth estimate) without adjudicating between them.
• Does not itself supply new evidence that AI profit is currently rent — for that, see the companion diagnosis paper Korinek & Vipra, "Concentrating Intelligence", read separately for this wiki; this page covers only the tax-design theory, which is agnostic to how large any actual AI rent turns out to be.
• A primer, explicitly. The authors describe the paper as providing "a primer for economists and policymakers," not a fully worked policy proposal; specific parameter values (the 11% illustrative AGI capital tax rate under one calibration; the 4% discount-rate benchmark) are worked examples, not recommendations.
• Provenance. All quotations, table content, and results verified against the NBER Working Paper 34873 PDF (February 2026), fetched and read in substantial part (sections 1, 2, 5, and the conclusion) this session; sections 3 and 4's full formal derivations were not exhaustively read and are not relied on here beyond the authors' own prose summaries of their results.

See Also

• Korinek & Vipra — Concentrating Intelligence — the diagnosis companion this paper's tax design responds to
• Taxing Tech Rents — Instrument Comparison — the wiki's own graded synthesis; this paper's Table 3 (compute/token/robot/DST mapping) is directly relevant source material for any future revision
• Korinek & Stiglitz — AI and income distribution — the long-run "innovator rents" theory this paper's tax design implements
• Cash-Flow Tax · Allowance for Corporate Equity — the rent-isolating mechanisms this paper invokes directly
• Land Value Tax · Law of Rent · ATCOR — the land-case template the authors generalize
• Citizen's Dividend — the redistribution analogue to the paper's sovereign-wealth-fund and UBC proposals
• Digital Services Taxes and Their Incidence — the real-world DST incidence record against which this paper's "integrate with VAT" Stage-1 guidance should be checked
• Objection: Taxing quasi-rents kills innovation — the rent-vs-normal-return identification problem this paper concedes is unsolved in practice
• Economic Rent · Quasi-Rent · Geoism

Sources

1. Anton Korinek & Lee M. Lockwood (2026), "Public Finance in the Age of AI: A Primer," NBER Working Paper 34873, February 2026 (JEL H21, H24, O33); prepared for the Brookings Center on Regulation and Markets ("The Future of Tax Policy: A Public Finance Framework for the Age of AI," Brookings, January 2026); forthcoming in The Economics of Transformative AI (A. Agrawal, E. Brynjolfsson & A. Korinek, eds.), NBER/University of Chicago Press. NBER PDF · Brookings — used for the core rent/normal-return distinction and "taxing rents involves no efficiency cost" formulation; the fixed-factor examples (unimproved land, spectrum rights, unique datasets); the two-stage (post-labor / AGI-dominated) framework and its formal results; the optimal-harvesting-problem framing and the discount-rate tax-rate result; Table 3's mapping of compute/token/robot/digital-services taxes onto capital/consumption tax instruments; and the sovereign-wealth-fund, windfall-clause, and Universal Basic Capital discussion (B/C-claims — an unrefereed NBER working paper offering formal theory and policy-instrument mapping, not an empirical estimate; fetched and read in substantial part — sections 1, 2, 5, and the conclusion — this session).