Specialization Tax: AI and Labor Dynamics

• Specialization Tax is a state-contingent levy on AI capital that is activated when cognitive workers begin switching to manual jobs.
• The model uses a dynamic four-factor growth framework with incentive-compatibility constraints to balance AI adoption and occupational specialization.
• Quantitative insights indicate that modest AI tax rates can delay cognitive-to-manual migration, preserving labor specialization in evolving economies.

Searching arXiv for the specified paper and closely related work to ground the article. The specialization tax is a state-contingent tax on AI capital, denoted $\tau_A(t)$, that arises endogenously in a dynamic four-factor growth model with human manual labor, human cognitive labor, physical capital, and artificial intelligence. In the model, the planner uses this instrument to preserve occupational assignment when incentive-compatibility constraints would otherwise induce workers to mimic another type. The central result is threshold-based: it is optimal to start taxing AI when cognitive workers begin to consider switching to manual jobs, a regime change that may occur once AI becomes sufficiently capable in substituting humans across cognitive tasks (Growiec et al., 18 Mar 2026).

1. Dynamic environment and planner’s problem

The economy is organized around an aggregate production function

$Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$

where $L_{c,t}$ and $L_{m,t}$ are cognitive and manual labor inputs, $K_t$ is physical capital, and $A_t$ is AI capital. Labor inputs are given by

$$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$

with fixed population shares $\pi_h$ and skill-productivity shifters $z_h$. Firms rent physical capital and AI capital at net marginal products $F_K\equiv \partial F/\partial K$ and $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$0, and hire labor at wages $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$1 (Growiec et al., 18 Mar 2026).

Type-$Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$2 households have utility

$Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$3

with $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$4, $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$5, $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$6, $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$7, and $Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$8. Feasibility is governed by the resource constraint

$Y_t = F(L_{c,t},L_{m,t},K_t,A_t),$9

The planner maximizes social welfare

$L_{c,t}$0

subject to feasibility, factor accumulation, the initial conditions $L_{c,t}$1, and an incentive-compatibility constraint designed to prevent type-$L_{c,t}$2 agents from pretending to be type-$L_{c,t}$3 agents:

$L_{c,t}$4

The current-value Lagrangian introduces multipliers $L_{c,t}$5 on feasibility and $L_{c,t}$6 on the binding incentive-compatibility constraint. From the first-order conditions with respect to consumption and labor, one recovers intratemporal Euler equations; from the first-order conditions with respect to $L_{c,t}$7 and $L_{c,t}$8, one obtains the intertemporal conditions for physical-capital investment and AI adoption. This setup places occupational specialization, rather than only factor accumulation, at the center of optimal taxation.

2. Incentive constraints and optimal tax wedges

The planner-imposed intertemporal wedge on factor $L_{c,t}$9 for a type-$L_{m,t}$0 agent is defined as

$L_{m,t}$1

Under the case in which the incentive-compatibility constraint binds for cognitive workers, the first-order conditions imply

$L_{m,t}$2

and

$L_{m,t}$3

Here $L_{m,t}$4 and $L_{m,t}$5 capture the effect of an additional unit of $L_{m,t}$6 or $L_{m,t}$7 on the multiplier of the incentive-compatibility constraint through its effect on the wage ratio $L_{m,t}$8 (Growiec et al., 18 Mar 2026).

In this regime, the planner taxes physical capital but subsidizes AI. The sign pattern reverses if the incentive-compatibility constraint binds for manual workers. Under that alternative assumption, the optimal wedge on AI becomes a positive tax. The same logic extends intratemporally to labor supply through

$L_{m,t}$9

which is negative, that is, a subsidy, for the type whose incentive-compatibility constraint is binding.

This sign structure is the paper’s main conceptual departure from uniform-factor-tax reasoning. The optimal treatment of AI is not fixed ex ante; it depends on which worker type is at the margin of occupational deviation.

3. Threshold logic and the switch in the binding constraint

The central state variable governing the regime change is the manual-to-cognitive wage ratio

$K_t$0

By Assumption 2, $K_t$1, so more AI raises the manual-to-cognitive wage ratio and makes manual work relatively more attractive. At low $K_t$2, the incentive-compatibility constraint on cognitive workers binds. Once AI becomes sufficiently productive in cognitive tasks, however, $K_t$3 crosses a threshold $K_t$4 at which the inequality in the incentive-compatibility condition flips (Growiec et al., 18 Mar 2026).

Formally, the threshold is characterized by

$K_t$5

In a static approximation, the cutoff $K_t$6 satisfies

$K_t$7

The note also provides a CES sub-aggregate for cognitive labor and AI:

$K_t$8

In this case,

$K_t$9

and the threshold becomes

$A_t$0

This makes the cutoff transparent in the elasticity parameter $A_t$1 and the CES weights $A_t$2.

A plausible implication is that the specialization tax is fundamentally a threshold policy, not a continuously increasing penalty on AI. The model’s comparative statics are organized around the crossing of $A_t$3.

4. Mechanism of the specialization tax

Once AI passes $A_t$4, the binding incentive-compatibility condition shifts to the manual type, and the optimal $A_t$5 becomes strictly positive. The note identifies this positive tax on AI capital as the specialization tax. Its mechanism is to reduce the after-tax marginal productivity of AI,

$A_t$6

which slows AI adoption by firms (Growiec et al., 18 Mar 2026).

The model ties this directly to occupational incentives. Because $A_t$7, unchecked AI accumulation pushes up the manual-to-cognitive wage ratio and makes departure from cognitive specialization more attractive. With the tax in place, the planner instead works with

$A_t$8

so the wage ratio becomes

$A_t$9

and its derivative with respect to $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$0 carries the extra factor $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$1. Solving for the time at which $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$2 reaches $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$3 yields the delayed crossing

$$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$4

The mechanism is therefore not framed as a correction for a technological externality in the abstract. It is a distortion chosen by the planner to preserve specialization by postponing the point at which cognitive workers would be tempted out of their assigned occupation.

5. Regime-dependent tax structure

The model implies a sharp before-and-after pattern in optimal wedges. The specialization tax is not active throughout the entire transition path. Rather, $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$5 is zero until $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$6 reaches $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$7 and then jumps to the optimal positive path implied by $$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$8 (Growiec et al., 18 Mar 2026).

Regime	AI wedge	Other wedge implications
Before the threshold	$$L_{h,t} = \pi_h\, l_{h,t}\, z_h \qquad (h\in\{c,m\}),$$9	Cognitive labor is subsidized: $\pi_h$0
After the threshold	$\pi_h$1	Traditional capital faces a subsidy: $\pi_h$2; manual labor is subsidized: $\pi_h$3

This regime dependence clarifies two common misconceptions. First, the framework does not imply that AI should always be taxed; before the threshold, the planner subsidizes AI. Second, the specialization tax is not exogenous. It arises endogenously from the interaction between the wage ratio $\pi_h$4, the binding incentive-compatibility condition, and the planner’s desire to preserve the occupational division of labor.

The note states that the level of the tax trades off the static efficiency losses from higher AI costs against the dynamic benefit of preserving cognitive specialization. This suggests that the tax is designed as a second-best instrument in an incentive-constrained allocation rather than as a permanent anti-adoption measure.

6. Quantitative illustration, scope, and policy interpretation

The note does not calibrate the model to data. It nonetheless provides an illustrative CES calculation: if $\pi_h$5 and $\pi_h$6, then increasing $\pi_h$7 from $\pi_h$8 to $\pi_h$9 percent postpones the threshold $z_h$0 by roughly $z_h$1--$z_h$2 percent (Growiec et al., 18 Mar 2026). The same passage states that, in more elaborate task-based calibrations, implied optimal AI taxes remain in the single-digit percent range but can slow cognitive-to-manual migration by several years.

The policy implications are correspondingly conditional. A positive tax on AI capital is called for exactly once AI becomes sufficiently substitutable for cognitive work so that cognitive workers would otherwise want to switch to manual tasks. Before that threshold, the planner subsidizes AI and cognitive labor to encourage cognitive R&D and employment; after the crossing, the sign of all wedges flips. The note further states that governments could begin modestly taxing revenues from AI services or robots once enterprise-level AI penetration hits industry-specific benchmarks, for example AI-substitution rates above $z_h$3 percent in cognitive occupations, with revenue recycled into R&D or cognitive-skill training subsidies (Growiec et al., 18 Mar 2026).

The scope of these conclusions is limited by the paper’s own structure. Because the analysis is conducted in a dynamic four-factor growth model with explicit incentive-compatibility constraints and without a calibration to data, the specialization tax should be understood as a theoretically optimal wedge within that environment. A plausible implication is that empirical implementation would require identifying when AI adoption has become sufficiently substitutable for cognitive work to shift the relevant binding constraint.