Methodology: How potential AI futures interact with the current tax system

All tax-side computations except the estimation of the corporate income tax use The Budget Lab’s tax microsimulation. The base file is the 2015 IRS Public Use File (PUF, ~150,000 records), supplemented with non-filer imputations and aged to the baseline year through SSA population weights and CBO economic scaling. The Tax-Simulator calculator handles individual income tax (AGI, standard and itemized deductions, ordinary and preferential rates, AMT, NIIT, EITC, CTC, CDCTC, education credits, Saver’s, QBI) and the payroll-tax module. It does not handle the corporate income tax or the estate tax; we address the CIT through the wedge described below.

The tax microsimulation alone does not carry asset-level wealth, which we need to allocate AI-driven capital income across households. To address this, we match tax units in the PUF with similar units in the Federal Reserve’s Survey of Consumer Finances (SCF). This process involves imputing asset values onto tax units using a random-forest machine-learning approach. This means that for every tax unit, we have an estimate of cash, equities, bonds, retirement balances, life insurance, annuities, trusts, real-estate funds, primary and other home equity, and pass-through equity, as well as active/passive splits for S-Corp and partnership profit. The PUF-SCF match is validated against aggregates from NIPA, SOI, and the Distributional Financial Accounts (DFA) upstream in the data pipeline. This project uses the PUF-SCF matched file as the ground truth.

We specify a tax unit’s income as \(Y_{j,i}^{X}\), where X specifies the type of income (labor, capital, etc.), \(j \in \{ 0,1\}\) specifies before or after the shock, and i is a tax-unit index. The aggregate income values lack the i subscript (i.e., \(Y_{j}^{X})\).  Each tax unit has a PUF weight \(w_{i}\) that has been adjusted to match the SSA and CBO targets discussed above. 

Once the merged file is loaded, we group a tax filer’s pre-shock income into two buckets:

• Labor income (\(Y_{0,i}^{L}\)): wages, sole-proprietor and farm income, and the labor portion of each unit’s S-Corp and partnership net profit.
• Capital income (\(Y_{0,i}^{K}\)): interest (taxable and exempt), dividends (ordinary and preferential), realized capital gains, other gains, net rent and estate income, taxable IRA and pension distributions, plus the capital portion of S-Corp and partnership net profit.

These values do not flow into the tax simulation but are only used to design the AI shock.

S-Corp and partnership profit blends true entrepreneurial labor (“a partner is paid for working”) with returns on invested capital (“a partner is paid for owning”). The split matters: a shift from wages to capital income carries different tax consequences than a shift from wages to pass-through active income, even when the dollar amount is the same. To address this we follow Saez and Zucman (2020) in allocating profits across labor and capital income differentially across the income distribution. For each tax unit, let \(W^{\star}\) be the 99.99th percentile of positive pre-shock wages (\(Y_{0,i}^{w})\) and \(Y_{j,i}^{PT}\) be their pass-through income of type \(PT\) (S-Corp active, partnership active, and passive). 

First, a tax unit’s passive labor pass-through income is \(Y_{j,i}^{L,Pass} = Y_{j,i}^{Pass} \cdot 0.25\) with, \(Y_{j,i}^{K,Pass} = Y_{j,i}^{Pass} \cdot 0.75\). Second, for both S-Corp and partnership active income, we divide up income between capital and labor differentially based on income relative to \(W^{\star}\):

\[Y_{j,i}^{L,PT} = 0.75 \cdot \min\left( Y_{j,i}^{PT},W^{\star} \right) + 0.25 \cdot \max\left( Y_{j,i}^{PT} - W^{\star},0 \right)\]

And 

\[Y_{j,i}^{K,PT} = Y_{j,i}^{PT} - Y_{j,i}^{L,PT}\]

The intuition is that an active partner earning at or below typical top-end wages looks like they are being paid mostly for their labor, while an active partner whose draw vastly exceeds top wages is more plausibly receiving rents on firm capital.1

Let \(\theta_{0}^{L}\) and \(\theta_{1}^{L}\) be the baseline and post-shock labor shares of factor income, with capital shares \(\theta_{t}^{K} = 1 - \theta_{t}^{L}\). Let \(H = 5\) be the Karger horizon and

\[G_{CBO}\mspace{6mu} = \mspace{6mu}\left( 1 + r_{CBO}^{2026} \right) \cdot \left( 1 + r_{CBO}^{2027+} \right)^{H - 1}\]

be cumulative CBO growth from 2025 to 2030. Then

\[g_{Y}\mspace{6mu} = \mspace{6mu}\frac{\left( 1 + r_{ai} \right)^{H}}{G_{CBO}} - 1\]

is the additional 5-year growth in nominal factor income caused by AI. For the three variants this deviation works out to roughly +0.6% (S), +3.6% (M), and +7.2% (R) above the CBO baseline at year 2030.

The growth rates of capital (\(g_{K})\) and labor (\(g_{L})\) follow from \(g_{Y}\) and the pre-and post-shock factor income shares of capital and labor; (\(\theta_{0}^{K},\theta_{1}^{K},\theta_{0}^{L},\theta_{1}^{L})\): 

\[g_{K} = \frac{\theta_{1}^{K}\left( 1 + g_{Y} \right) - \theta_{0}^{K}}{\theta_{0}^{K}},\quad\quad g_{L} = \frac{\theta_{1}^{L}\left( 1 + g_{Y} \right) - \theta_{0}^{L}}{\theta_{0}^{L}}\]

The microsimulation’s measured labor share differs from the NIPA share Karger reports — the microsim aggregates exclude employer FICA, undistributed C-corp profits, and imputed housing rent — but the growth rate of each factor is comparable across data sources, so the shock translates cleanly.

We hold constant the labor income for tax units with negative baseline labor income and scale the labor income of all tax units with positive baseline labor income by a single multiplier \(\rho\) so the aggregate equals \(Y_{1}^{L}\):

\[Y_{1,i}^{L}\mspace{6mu} = \mspace{6mu}\left\{ \begin{matrix} \rho \cdot Y_{0,i}^{L} & \text{if }Y_{0,i}^{L} > 0, \\ Y_{0,i}^{L} & \text{otherwise.} \end{matrix} \right.\]

By construction, \(\sum_{i}^{}w_{i} \cdot Y_{1,i}^{L} = Y_{1}^{L}\) (limited to positive values). The proportional scenario preserves the baseline distribution of positive labor income.

It is empirically uncertain whether the labor-side adjustment induced by AI will compress (S2) or stretch (S3) the distribution of labor income.2 We consider both and implement them via opposing dispersion-shift specifications using a log-affine transformation applied to tax units with positive labor income:

\[Y_{1,i}^{L}\mspace{6mu} = \mspace{6mu}\mu_{1} + \lambda \cdot \left( \ln Y_{0,i}^{L} - \mu_{0} \right),\]

where \(\mu_{0}\) is the weighted log-mean of positive baseline labor income, \(\mu_{1}\) is solved so the positive-subset aggregate matches \(Y_{1}^{L}\) after scaling, and \(\lambda\) is defined as the ratio of post- to pre-shock standard deviation of \(\ln Y_{j,i}^{L}\) on the positive subset. We tie \(\lambda\) to the productivity shock \(g_{Y}\) through

\[\lambda_{S2}\mspace{6mu} = \mspace{6mu} 1 - k \cdot g_{Y}\quad\text{(compressive)},\quad\quad\lambda_{S3}\mspace{6mu} = \mspace{6mu} 1 + k \cdot g_{Y}\quad\text{(expansive)}.\]

For now we assume \(k = 1\), which results in a proportional mapping between growth and dispersion. The code supports a scalar multiplier if a different calibration becomes appropriate.

The compressive scenario captures the world in which AI’s labor-side cost falls hardest on the top end of the wage distribution (substitution at the high end, e.g. cognitive professional work). The expansive scenario captures the opposite — AI raises the productivity (and pay) of high-skill labor while displacing or stagnating mid- and bottom-wages. We do not take a view on which is more likely.

We run every variant under two different factor-share settings (\(\chi\)) to let readers separate the productivity and reallocation channels:

• Reallocate (R) — use \(\theta_{1}^{K}\) from Karger so capital grows faster than labor.
• Fixed share (F) — pin \(\theta_{1}^{K} := \theta_{0}^{K}\) so the post-shock factor split equals the baseline. Here, we use the same \(g_{Y}\), but \(g_{K} = g_{Y}\) so labor and capital both grow at the productivity rate.

The fixed-share option isolates the productivity increase from the factor reallocation. Together, the two share modes let us decompose “AI’s effects on revenue and distribution” into “more taxable output” and “the mix of who earns that output has shifted.”

The aggregate capital income flow added by the shock is

\[Y_{1}^{K} = \mspace{6mu} Y_{0}^{K} \cdot \left( 1 + g_{K} \right)\]

For ease of notation, we define the excess capital income as:

\[X = Y_{1}^{K} - Y_{0}^{K} = Y_{0}^{K} \cdot g_{K}\]

Corporate income tax operates upstream of household realizations, so the AI capital flow \(X\) reaches tax units in full. In order to capture the effect of an AI shock on corporate income tax revenue, we need to estimate what growth in C-corporation income is implied by \(X\).  We proceed simply by multiplying the change in capital income by the ratio of CBO’s baseline level of CIT revenue \((R_{CBO}^{CIT})\) to baseline (2030 in this case) realized capital income (\(Y_{0}^{K})\), implying that the AI CIT delta then scales linearly with \(X\):

\[\Delta R^{CIT} = X \cdot \frac{R_{CBO}^{CIT}}{Y_{0}^{K}}\]

We can back-out the implicit assumptions implied by this set-up as follows. Let \(\kappa\) be the share of \(Y_{0}^{K}\) that flows through C-Corporations (versus entities that are not exposed to the corporate income tax, like partnerships and S-Corporations), \(\tau_{C}^{stat}\) be the statutory corporate tax rate (\(21\%)\), and \(\eta\) be an adjustment factor absorbing the statutory-vs-effective gap (avoidance, NOLs, credits, profit shifting, etc..). Then: 

\[R_{CBO}^{CIT} = Y_{0}^{K} \cdot \tau_{C}^{stat} \cdot \eta \cdot \kappa\]

Our formulation above then assumes that \(\tau_{C}^{stat},\eta,\kappa\) are constant across AI-induced shocks. 

\[\Delta R^{CIT} = X \cdot \frac{R_{CBO}^{CIT}}{Y_{0}^{K}} = X \cdot \tau_{C}^{stat} \cdot \eta \cdot \kappa = X \cdot \tau_{C}^{eff} \cdot \kappa\]

Where \(\tau_{C}^{eff} := \tau_{C}^{stat} \cdot \eta\). The absence of change in the statutory tax rate is true by assumption, and future work will investigate how AI-induced capital growth might differentially impact the take-up of various adjustments that make up \(\eta\) and vary across corporate form, impacting \(\kappa\). 

With \(\tau_{C}^{stat} = 21\%\) (TCJA), \(\kappa \approx 0.50\) (via NIPA), and CBO’s 2030 CIT anchor of \(\sim \$ 486\)B (or a CIT-to-GDP equaling \(0.013\) times GDP), \(\eta\) falls out to roughly one in practice; its exact value is reported on the parameters sheet of every run.

The corporate-tax delta enters aggregate revenue as a single number layered onto the micro-simulation revenue total; it is not attributed back to individual households. As a result, every distributional figure in our output reflects the full capital flow \(X\) that reaches households. This is a deliberate choice: we do not adopt a per-household corporate-incidence assumption. 

We distribute \(X\) across tax units in proportion to each unit’s share of pre-shock total household wealth:

\[X_{i}\mspace{6mu} = \mspace{6mu} X \cdot \frac{A_{0,i}}{\sum_{k}^{}w_{k}A_{0,k}},\]

where \(A_{0,i}\) is the sum of the unit’s SCF-imputed wealth columns (cash, equities, bonds, retirement balances, life insurance, annuities, trusts, real-estate funds, primary and other home equity, pass-through equity, and miscellaneous non-financial assets). Allocating to total wealth assumes the AI shock raises returns proportionally across asset classes; pinning to a narrower base would push the distributional incidence further toward the top and is an area we plan to explore in the future.

The asset base contains both income-bearing and non-income-bearing classes. Of the income-bearing classes, four matter for tax purposes: taxable equities, bonds, pass-through equity, and retirement balances. Let \(A_{j,i}^{Inc}\) be their sum for a given tax unit. We allocate each unit’s \(X_{i}\) across the four classes proportional to its baseline holdings within \(A_{0,i}^{Inc}\).

Units with zero income-bearing holdings route their entire \(X_{i}\) to retirement, preserving the aggregate identity \(\sum_{i}^{}w_{i}X_{i} = X\). This fallback affects roughly 33% of tax units in 2030 but only 6.7% of \(X\); most of the affected mass sits in tax units whose wealth is only in housing. If we redefine the asset base to exclude housing, only 2.4% of \(X\) is routed through retirement via this fallback.

Each class then flows through to specific taxable-income types:

• Taxable public-equity flow splits between qualified dividends and gross long-term capital gains using a fixed share table (30% dividend / 70% gross LTCG).
• Fixed-income flow splits between taxable and tax-exempt interest at each unit’s baseline ratio.
• Pass-through equity flow becomes pass-through ordinary income, further allocated across S-Corp and partnership active/passive sub-buckets using baseline positive capital holdings as weights.
• Retirement flow runs through the procedure described in the next section.

Two separate channels move revenue in a typical scenario: shifting labor income and adding capital income. The interaction between them is non-linear (because the tax schedule itself is non-linear: brackets, phase-outs, AMT), so a clean “labor versus capital” attribution requires care.

Each cell produces three Tax-Simulator runs:

• Both — labor and capital shocks both applied (the standard scenario).
• Labor-only (LO) — labor side shocked; capital flows held at baseline.
• Capital-only (CO) — capital side shocked; labor flows held at baseline.

We then decompose the total revenue change as

\[\Delta T_{both}\mspace{6mu} = \mspace{6mu}\Delta T_{LO} + \Delta T_{CO} + \text{interaction},\]

where

\[\text{interaction}\mspace{6mu} = \mspace{6mu} T_{both} - T_{LO} - T_{CO} + T_{base}.\]

The interaction term captures non-linear bracket effects (the combined shock can push a unit across an AMT or NIIT threshold that neither side would have triggered alone). Payroll tax counts mechanically as labor: the capital-only run leaves wages at baseline, so its payroll delta is zero.

We use the Budget Lab Small Macro Model (BLSMM) to estimate the effect of these shocks on the debt-to-GDP ratio. To do this, we need to provide the model with the change in the revenue-to-GDP ratio.  Because the modeling above omits various forms of tax revenue not directly estimated by either our macro wedge or our individual tax simulations, the model’s internal revenue / GDP ratio understates the published level by roughly 1.3 percentage points in 2030. For the publishable revenue-to-GDP comparison we therefore convert the baseline to CBO’s published 2030 ratio (17.7% from CBO publication 62105) into dollars and adjust this dollar amount by the estimated change in revenue from our simulations. 

Represent the revenue-to-GDP ratio as \(Z_{j} = \frac{R_{j}}{Y_{j}}\), where j indicates the source of the data, with \(j \in \{ 0,1\}\) following the notation above. We want to estimate \(\Delta Z\), or the change in the revenue-to-GDP ratio due to a given AI shock. As noted above, defining  \(R_{1}\) based solely on the change in revenue estimated above under-estimates total revenue, so we define: 

\[R_{1}\mspace{6mu} = \mspace{6mu} Z_{0} \cdot Y_{0} + \Delta R_{TBL}\]

The estimated change in revenue (\(\Delta R_{TBL}\)) includes both the Tax-Simulator change in income and payroll tax (net of refundable credits) plus the corporate-tax computed outside the microsim. The baseline revenue-to-GDP ratio \(Z_{0}\) and the GDP level \(Y_{0}\) come from the 2026 Budget and Economic Outlook.

Given this, we can estimate the change in the revenue-to-GDP ratio: 

\[\Delta Z = Z_{1} - Z_{0} = \frac{R_{1}}{Y_{1}} - \frac{R_{0}}{Y_{0}} = \frac{Z_{0} \cdot Y_{0} + \Delta R_{TBL}}{Y_{0}(1 + g_{Y})} - \frac{R_{0}}{Y_{0}}\]

We can re-write this as: 

\[\Delta Z = \underbrace{\frac{\Delta R_{TBL}}{Y_{0}(1 + g_{Y})}}_{\text{Revenue}} + \underbrace{Z_{0}\frac{g_{Y}}{(1 + g_{Y})}}_{\text{Macro}}\]

Where the first term represents the change in revenue, and the second represents the shock to GDP. Even with no revenue change, GDP growth alone would push revenue / GDP down. Anchoring to CBO applies that drag to a realistic revenue base, so the published response number is meaningful relative to the forecast level. The maintained assumption is that revenue streams the model omits (notably the baseline CIT level and any pieces of “other” revenue not captured by Tax-Simulator) are unchanged in the scenario except through channels already in \(\Delta R\). 

With this input, we work through the following process:

1. Linearly scale the revenue-to-GDP ratio so that by the baseline year, the full \(\Delta Z\) is applied. The adjustment is zero in 2025, 0.2 of the target in 2026, and reaches full level in the baseline year, and held at full level thereafter.
2. Solves for a constant per-year productivity bump (added to BLSMM’s potential-output path) such that the annualized 2025–baseline-year real GDP growth in BLSMM matches the Karger variant’s \(r_{ai}\) target.
3. Reads off BLSMM’s 2030 federal debt and debt/GDP

BLSMM’s CBO-style fiscal feedback (the model’s \(\psi_{1}\), \(\psi_{2}\) elasticities) mechanically lowers outlays as potential output rises. Under an AI shock policymakers may instead raise outlays for displaced-worker support; the BLSMM number therefore reads as a baseline-feedback estimate of debt/GDP rather than an all-things-considered projection. 

A natural fourth labor-incidence rule scales each unit’s labor income by an occupation’s exposure to AI: \(Y_{1,i}^{L} = \left( 1 + \beta e_{i} \right)Y_{0,i}^{L}\) where \(e_{i}\) is an occupation-level AI-exposure index, and \(\beta\) is solved so the aggregate matches \(Y_{1}^{L}\). The mechanic is straightforward; the missing piece is an occupation-level exposure imputation onto the tax-unit file. Once that imputation exists, the AI-exposure scenario slots into the labor-scenario axis alongside the existing proportional / compressive / expansive rules.

The current specification assumes that the AI effect is applied on the intensive margin: all tax units have their labor income adjusted by some parameter. However, it is highly likely that some of the effect of AI on labor income will be felt on the extensive margin via job loss or creation. A future extension would allow individual income to adjust along this margin to match \(Y_{1}^{L}\).

Currently, the AI shock is applied to both labor and capital as a function of observed income on the PUF and imputed assets via the SCF. However, this bakes in a number of assumptions, including the realization rate of capital, the shares of corporate income that is held by C-Corporations, the treatment of assets held in retirement accounts, and the way in which we implement the corporate income tax effect. Ideally, we would begin with a model of the macro-economy that could be subjected to an AI shock. That shock would flow down to different types of income (capital, labor, etc.) via a series of parameters, allowing a far greater degree of flexibility in defining the structure of a potential shock. We discuss a few examples of these sorts of extensions below: 

• If the macro-shock is applied in such a way to allow us to observe an impact of AI on capital income that flows through C-Corporations, then we could specify a more extensive corporate income tax model, rather than our simple macro-wedge approach detailed above. 
• The current release treats AI capital gains as recognized immediately at accrual, because the capital income shock is built from the on-1040 realized base and re-applying a realization discount would double-count. A natural extension is to begin with a definition of \(X\) that includes unrealized capital gains and apply an explicit realization rate — the share of the annual gain stock realized each year — as a separate parameter: This requires a setting where \(X\) captures unrealized as well as realized gains and an estimate of the realization rate from the literature. Further work would condition the realization rate on individual or economy-level characteristics: the marginal tax rates a unit faces, the composition of its wealth and income, and the relative size of its capital versus labor base.
• For retirement income specifically, we would ideally model not just the realization rate but the share of realizations that are taxable. Most retirement wealth sits in tax-deferred vehicles (401(k), IRA, traditional pensions); the marginal AI dollar flowing into those accounts would not be taxed in-year.

The single-year flow treatment misses a longer-term dimension of capital-gains taxation: gains can compound unrealized, get realized at some hazard rate, or escape tax entirely via step-up at death. A present-value formulation — unrealized pool compounding at \(\rho\), realized at hazard rate \(r\), discounted at \(\delta\), with a step-up haircut \(\varphi\) on the residual at horizon \(T\) — captures this. Tax-Simulator is an annual-flow object and cannot natively consume a lifetime-PV input, so this extension requires either an out-of-model aggregation step or a redesign of the Tax-Simulator handoff.

The current specification is mechanical. Two behavioral channels are worth modeling: labor-supply responses to the change in net-of-tax returns (Elasticity of Taxable Income \(\sim 0.25\) per Saez-Slemrod-Giertz), and realization-rate responses to capital gains tax exposure.

The current model holds prices, wages, and the asset stock \(A_{i}\) at baseline and distributes the flow \(X\) without revaluing the stock. A full GE pass would let asset prices re-equilibrate as AI capital returns rise, pushing the distributional results further toward the top.