Frank Schmid · 8 November 2026 · The flow of the paper, section by section

Artificial Intelligence, Mass Unemployment, and the Optimal Financing of Universal Basic Income

Under conditions the paper states, among them that AI and labor are gross substitutes and that the supply of AI capital is not too elastic, rising AI intensity (the productivity-weighted share of AI-performed tasks) lowers the wage and raises non-participation (the share of the population that does not work). AI capital is supplied against scarce complementary inputs, so part of its return is a quasi-rent; for an instrument set that excludes subsidies, the computed optimum finances a universal basic income (UBI) by placing the AI capital tax on that quasi-rent, leaving the normal return to AI capital untaxed, and setting a labor income tax that is progressive with a zero rate at the participation margin. Along the computed transition, for that instrument set, the effective AI tax (the AI tax as a share of the AI return) rises with AI intensity and, because the planner values faster adoption at the margin, lies below the optimum at the same AI intensity without the laws of motion until the rate on the quasi-rent reaches its upper limit in year 24.

The displacement of labor

AI capital substitutes for labor at the task level. The theme carries the elasticity between the two as an equilibrium object, the response of the wage and of participation to rising AI intensity in general equilibrium, and the six conditions jointly sufficient for mass unemployment, a large value of non-participation.

The financing of UBI

UBI is a transfer and in part public insurance against the devaluation of skill as AI’s task reach expands. The theme carries the taxes that finance it, in steady state and along the transition, and the dependence of that financing on coordination across jurisdictions where AI capital is mobile.

The line of sight

1

Introduction

Displacement of laborFinancing of UBI
Establishes
For an instrument set that excludes subsidies, the optimal tax structure is time-varying and state-contingent: the tax on AI capital falls on the quasi-rent to scarce complementary inputs, and its effective rate rises as AI intensity rises, because the rent share of the AI return grows as demand for AI capital rises against an inelastic supply. Rising AI displacement generates both the fiscal requirement for UBI and the tax base to finance it.
Leads to
What does the literature on optimal taxation, technological unemployment, and UBI leave open?
2

Related Literature

Displacement of laborFinancing of UBI
Establishes
Guerreiro, Rebelo, and Teles show that a robot tax is part of the optimum during the transition when displaced workers cannot be fully compensated through the income tax, with the optimal rate declining toward zero in the long run. Thuemmel finds quantitatively small optimal robot taxes once the nonlinear income tax adjusts, and Costinot and Werning derive optimal technology regulation from sufficient statistics, obtaining modest optimal rates. This paper differs in the scale of its revenue requirement and in an upward-sloping AI capital supply with taxable quasi-rents. It takes from Acemoglu and Restrepo the distinction between displacement and reinstatement (the creation of new tasks assigned to labor at prevailing factor prices), with a microfoundation that allows reinstatement to be modeled explicitly. Jones takes tasks to be complements. The mapping of Section 3 from AI intensity to the share of tasks AI performs assumes equal task productivities and a task-level elasticity of substitution equal to the aggregate elasticity between AI and labor, 1.8 in the calibrated path; within that mapping, tasks are substitutes. In the experiment of Vivalt et al., an unconditional transfer to low-income adults in two U.S. states lowered employment; its effect on whether recipients work is the margin through which the UBI of the computed optimum raises non-participation at year 0 above the laissez-faire rate of 5 percent.
Leads to
What foundation gives the elasticity between AI capital and labor its content?
3

A Task-Based Foundation for AI Production

Displacement of labor
Establishes
The aggregate elasticity of substitution between AI capital and labor is an equilibrium object that rises as AI’s task reach expands. The strength of reinstatement is a modeled feature rather than an assumption; whether reinstatement is weak for AI is an empirical question the paper does not resolve. AI intensity is a distinct object from the share of tasks AI performs.
4.2share of tasks AI performs at the year-0 AI intensity of 0.15, an illustrative value, under the mapping of Section 3, percent
87.8the same at the adoption ceiling on AI intensity, 0.75, a scenario for which no source gives an estimate, percent
Exhibits
Equation (5), AI intensity and the share of tasks AI performs
Leads to
How does rising AI intensity act on the wage and on participation?
4

The Baseline Economy

Displacement of laborFinancing of UBI
Establishes
UBI raises the reservation wage, and with it the employment threshold: a worker participates where the wage times effective labor reaches the reservation wage. AI capital is supplied against scarce complementary inputs, so its supply slopes upward, and part of its return is a quasi-rent, which can be taxed without distorting the marginal investment decision. Under the conditions of Proposition 1, among them that AI and labor are gross substitutes and that the AI capital supply is not too elastic, rising AI intensity raises non-participation in general equilibrium. Six conditions are jointly sufficient for mass unemployment.
Exhibits
Equation (25), the response of non-participation to AI intensity
Leads to
What does a universal transfer add when the devaluation of skill cannot be insured?
5

Risk, Insurance, and the Welfare Role of UBI

Financing of UBI
Establishes
The risk that a worker’s skill is devalued as AI’s task reach expands is uninsurable by private markets, and UBI acts in part as public insurance against it. The welfare gain from UBI divides into a mean-income gain and an insurance value, positive under decreasing absolute risk aversion. A planner that ignores the insurance component understates the welfare gain from UBI and the optimal transfer. The log-normal approximation of the values below understates the insurance value where the pre-UBI distribution places mass near zero consumption. The path of Section 7 is computed without skill-devaluation risk.
0.028approximate insurance value of UBI at stipulated values (a coefficient of relative risk aversion of 2, a coefficient of variation of pre-UBI consumption of 0.6, and UBI of 0.23 of mean pre-UBI consumption), share of mean consumption per period
2.4–3.2the same across relative risk aversion from 1.5 to 3, at the other stipulated values, percent of mean consumption per period
Exhibits
Equation (27), the decomposition of the welfare gain from UBIEquation (28), the insurance value under constant relative risk aversion
Leads to
Which taxes finance the transfer at the least cost in welfare?
6

The Optimal Tax Structure in Steady State

Financing of UBI
Establishes
Raising the AI capital tax lowers the wage (Proposition 1) and erodes the labor tax base, and that revenue leakage by itself places the optimal AI capital tax below the inverse-elasticity level; the Pigouvian and quasi-rent components work in the opposite direction, so the position of the total rate relative to that benchmark is a quantitative question. The static optimum is the optimum at a given AI intensity without the laws of motion; at the adoption ceiling it is the steady state. For an instrument set that excludes subsidies, the computed static optimum sets the rate on the quasi-rent at its upper limit and the rate on the normal return at zero, its lower limit, at every AI intensity considered, so that the AI tax falls on the rent base alone. The labor income tax is progressive across the terciles of the employed. Its rate at the participation margin is zero, a lower limit that binds: the planner would choose a participation subsidy, the earnings subsidy of Section 7.9, if the instrument set admitted one. In the model as computed, the consumption tax is not identified as a separate instrument.
36–86effective AI tax in the static optimum, across AI intensities from 0.15 to 0.75, percent
23–54UBI in the static optimum, across the same intensities, percent of per-capita gross domestic product (GDP)
40–64revenue in the static optimum, across the same intensities, percent of output
Exhibits
Table 1, Computed Optimal Tax Structure by AI IntensityEquation (34), the rate on the normal return to AI capitalEquation (39), the participation response to the bottom-tercile labor rate
Leads to
How does the structure move as AI intensity rises over time?
7

The Dynamic Transition Path

Displacement of laborFinancing of UBI
Establishes
For an instrument set that excludes subsidies, along the computed path AI intensity rises from 0.15 to 0.49 over 30 years, and non-participation rises in every year and without a plateau. Its year-0 value exceeds the laissez-faire rate of 5 percent because the optimal UBI lifts the reservation wage, so the series measures displacement and policy-induced non-participation together. The rate on the quasi-rent reaches its limit of 100 percent in year 24. Until then the path taxes AI capital less than the static optimum at the same AI intensity. At year 0 a marginally higher AI intensity raises the planner’s objective, so the planner lowers the AI capital tax to speed adoption, although in years 0 through 3 adoption outpaces the institutional absorption rate and a congestion term works in the opposite direction. The effective AI tax rises to 78 percent at year 30. As the elasticity of substitution between AI capital and labor rises from 1.3 to 2.4, with the other parameters at their baseline values, non-participation at year 30 rises from 54 to 71 percent; the displacement result, that rising AI intensity raises non-participation, holds across that range, and its size depends on the elasticity. The more adoption responds to the tax, the lower the AI capital tax the planner sets and the slower the transition. At the baseline return to retraining and below it, retraining is not used. With an earnings subsidy at the participation margin admitted, the problem has more than one local optimum. Over the optima found, which use the subsidy, non-participation stays at or below 4.4 percent in every year from 0 through 30; none is established as the global one, and why non-participation settles at that low level is not established. The welfare function counts the consumption of a household that does not participate and attaches no value to its outside option, whose income equivalent is the reservation wage. A subsidy that draws workers into employment is valued accordingly, and a criterion that valued time out of work would weigh it differently.
16non-participation on the path at year 0, percent
64non-participation on the path at year 30, percent
36effective AI tax at year 0, static optimum at the year-0 AI intensity less the path (path 0 percent), percentage points
44UBI on the path at year 30, percent of per-capita GDP
Exhibits
Table 2, Computed Optimal Tax PathFigure 1, Computed Optimal Transition Path, Years 0 through 30Figure 2, Sensitivity of the Computed Path at Year 30 to Four ParametersTable 3, Comparative-Statics Summary: Sign of Endogenous ResponsesEquation (47), the optimal AI capital tax with its dynamic correction
Leads to
What changes when AI capital can move to another jurisdiction?
8

The International Dimension

Financing of UBI
Establishes
With mobile AI capital, the symmetric Nash rate of the reduced-form best response lies below the closed-economy optimum. The gap increases with capital mobility and decreases with the tightness of the complementary-input constraint, under which the tax falls partly on the immobile rent. In the computed two-country solution, at a stipulated mobility share of 0.5, the uncoordinated mean effective AI tax lies below the cooperative one, at a cost in welfare. Destination-based taxation and minimum-rate coordination partially restore the welfare losses from tax competition.
36mean effective AI tax under Nash over the reported years from 0 to 30, at the mobility share of 0.5, percent
23cooperative mean less the Nash mean, effective AI tax, percentage points
2.1welfare loss of the Nash outcome, equivalent to a reduction of every household’s consumption at every date of the cooperative path, percent
Exhibits
Equation (56), the best response in AI capital taxesEquation (57), the gain from coordination
Leads to
What separates the optimal structure from an enacted one?
9

Implementation Challenges

Financing of UBI
Establishes
Taxing AI capital and AI rents requires workable definitions of the base. The tax on the quasi-rent to the complementary inputs is suited to input-based implementation, because compute capacity, electricity, and specialized labor markets generate measurable flows, though input-based taxation introduces distortions of its own. A tax structure raising 40 through 64 percent of output requires sustained political support. Seven limitations, stated in Section 9.6, bound the framework’s conclusions.
Leads to
What does the framework conclude about mass unemployment and the financing of UBI?
10

Conclusion

Displacement of laborFinancing of UBI
Establishes
Mass unemployment from AI adoption is not inevitable but emerges under six conditions; whether they hold now is an empirical question the framework does not settle. For an instrument set that excludes subsidies, the computed financing places the AI capital tax on the quasi-rent to scarce complementary inputs and sets the tax on normal AI returns to zero, with a labor income tax progressive across terciles; the structure finances UBI of 19 through 44 percent of per-capita GDP along the path. The structure is time-varying and state-contingent, and it depends on international coordination of AI capital taxation.
Appendices
Appendix A, Derivation of the Aggregate CES from Task-Level PrimitivesAppendix B, Notation SummaryAppendix C, Numerical Solution and Calibration