How potential AI futures would play out in the current tax system

Putting all these pieces together, we now show estimates of revenue responses under the scenarios described above.7

Figure 1 shows the overall revenue effects across scenarios. In the Slow variant, federal revenues in 2030 barely change, the result of a nearly identical path when compared to the baseline level of economic growth. In the Moderate scenario, we see revenues are higher in 2030 by between $85 and $127 billion, depending on how labor income inequality changes. (The “Compressive” assumption is that labor income inequality declines; the “Expansive” assumption is that it increases.) Finally, we see the largest revenue gains in the Rapid growth scenario; a maximum of $216 billion with Rapid growth and rising labor income inequality, or an increase in federal revenues of 3.3 percent.8

Figure 2 shows how total revenues from corporate income tax, capital taxes, and labor taxes change in each scenario. Moving from the Slow to Moderate to Rapid scenarios, productivity growth and capital share both rise, causing revenue from corporate income tax (dark blue) and capital taxes (orange) to increase as well. In the Rapid-Expansive scenario, the collective increase in those tax revenues amounts to about $216 billion more annually than in the status quo. 

More complicated is the effect of AI growth on the revenue from labor taxation across scenarios. Fundamentally, there are two forces at work: a) the degree to which the labor share falls relative to the output increase and b) the change in labor income inequality. If the labor share falls sufficiently relative to GDP growth—as is the case in the Slow and Rapid scenarios—then we estimate there will be a cumulative reduction in labor income relative to the CBO baseline. In those two scenarios, labor income shrinks by between 0.3 and 1.0 percentage points and tax revenue from labor income declines (in most cases). On the other hand, tax revenue from labor income tends to increase in the Moderate scenario because the projected labor share decline is relatively small. Within any of the Slow, Moderate, or Rapid scenarios, labor tax revenue is rising in the amount of labor income inequality, due to the progressive nature of the income tax system. These two forces can move either in the same direction or can oppose each other, depending on the mix of assumptions. 

Figure 3 provides a different perspective, decomposing revenue effects by tax instruments, allowing us to observe how payroll and individual income tax receipts vary across AI scenarios, along with corporate tax and tax credit outlays.9 Payroll and individual income tax revenues change in very different ways across the scenarios. As inequality rises (left to right within a cluster of bars), revenue from the income tax rises, reflecting the progressivity of that tax schedule. In sharp contrast, payroll tax revenue declines as one moves left to right within a cluster of bars. When inequality is higher, more labor income exceeds the Social Security earnings cap and is consequently subject to a much lower tax rate. 

Simulating corporate and capital tax receipts requires several additional assumptions, described in the next section and appendix. One of them is that we assume that the additional capital income, generated by the AI shock, is realized at the same rate as capital income at baseline. We make this assumption because our estimate of the excess flow of capital income is based on the observed (realized) capital income in our tax data that grows as specified in our scenarios. 

Guided by survey respondents’ expectations about how the labor share will fall, our AI scenarios imply that excess economic growth skews heavily towards capital and away from labor. This has a pronounced effect on our revenue calculations, as shown in Figure 5. In the most extreme case (with relatively slow GDP growth and a reduction in labor income inequality), the revenue increase above baseline is 82% lower than it would have been if the labor share and distribution had not changed. In every scenario variant, total revenues are lower than they would be had we assumed that the labor share remained constant, and excess economic growth was more balanced between labor and capital. 

The stark difference in these revenue projections is driven by two factors. First, capital income faces a much lower average statutory tax rate than labor. This factor is explicit in our tax simulations. Second, implicit in our model is the fact that many forms of capital income are not subject to taxation at all: this includes unrealized gains, preferential treatment for retirement income, and other factors like the holding of equities by non-US citizens.  

With the mix of income and tax revenue changes in each scenario, it is not immediately obvious how average tax rates might change across the distribution of taxpayers. Figure 7 shows the percentage point change in average tax rate for the Moderate scenario, assuming a proportional shock to the distribution of labor income. Average tax rates grow modestly for almost all deciles, reflecting the accumulation of income in higher tax brackets. We also see that this effect tapers off at the top of the income distribution due to the income mix for tax filters in the top decile being especially capital heavy. 

The analysis above is focused on the revenue and distributional implications of different AI shocks. This leaves unanswered some important questions about the overall fiscal picture that depend on how net interest payments and non-interest outlays respond to the AI shock. In Appendix Figure 4, we show how debt-to-GDP shares in 2030 deviate from baseline in each AI scenario.13 The most important driver of those deviations is the assumed output growth rate through 2030. However, we do not attempt to model how federal expenditures would respond in AI scenarios, which would have important implications for the debt trajectory.14

This appendix provides additional details that are helpful for understanding and interpreting the results. For the full details, see the accompanying methodology document. The model is available on GitHub here, but the data used to analyze the scenarios cannot be released publicly. 

We begin with the tax-unit level data created for and used by The Budget Lab’s tax microsimulation model, based on the 2015 IRS Public Use File, which is aged to the baseline year (2030 in our case) through SSA population weights and the CBO economic forecast. We impute asset values onto these tax units using the Survey of Consumer Finances (SCF) and a random-forest machine-learning approach. Notably, our baseline incorporates the CBO’s assessment of the growth effects of AI, which in 2026 were approximately 10 basis points per year on average. 

Based on the particular scenario we adjust overall labor and capital income to account for AI effects on the GDP growth path, labor share of income, and distribution of labor income. We then apportion these changes to individual tax returns on the PUF. Using our tax microsimulation model and a simple representation of the corporate income tax, we then obtain revenue and distributional estimates by comparing the data before and after these adjustments. 

Below, we show some of the key calculations for implementing the Karger et al. (2026) scenarios. Second, we describe additional economic assumptions not discussed above. Third, we discuss our process for combining these results with our macro modeling capacity to estimate debt-to-GDP ratios. Finally, we present a few additional figures.

Table 1 (reproduced below with variable symbols) provides the values for key variables that are implied by our implementation of Karger et al. (2026) economic forecasts. These include the output increase and the decline in labor share. Because survey respondents were asked about three consistent scenarios (Slow, Moderate, and Rapid), we can link the labor share and economic growth changes together. Further, while our calculations for the change in labor income inequality were not directly taken from the Karger et al. (2026) paper, we use their GDP growth projections as inputs to estimate the scale of labor income compression or expansion. The calculations underlying Table 1 values were conducted as follows:

We begin with a CBO baseline level of cumulative GDP growth through 2030 of 9.7%, or 2.2% for 2026 and 1.8% for all following years (\(G_{CBO})\). We also have pre-shock labor (\(\theta_{0}^{L})\) and capital (\(\theta_{0}^{K} = 1 - \theta_{0}^{L})\) shares of factor income. From the aged PUF, we have a pre-shock distribution of labor income. 

For any scenario, we begin by scaling up the cumulative CBO growth rate by the annualized rate (\(r_{ai})\) from Karger et al. (2026).

\(g_{Y} = \frac{\left( 1 + r_{ai} \right)^{5}}{G_{CBO}} - 1\).

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 labor growth rate (\(g_{L})\) is negative for Slow and Rapid because the labor share (\(\theta_{1}^{L})\) falls fast enough to outweigh productivity gains over the five-year horizon. 

The aggregate capital income flow added by the shock is:

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

where \(Y_{0}^{K}\) is the baseline level of realized, taxable capital income observed on the PUF. Because we are defining our AI shock as a function of taxable income, we are assuming that the rate of realization of capital income is unchanged by the AI shock. We then apportion \(Y_{1}^{K}\) to individuals based on their total assets (imputed from the SCF) and within individuals across types of capital income (capital gains, dividends, interest, retirement accounts, etc.) proportional to types of assets (equities, bonds, etc.). 

The aggregate labor income flow added by the shock is:

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

The "labor income inequality parameter (λ)" rows govern the within-tax-unit dispersion of the labor shock under the compressive or expanding specifications. For the proportional model, we simply scale up all positive labor income by a single multiplier to target post-shock aggregate labor income (a function of \(g_{L}\)). 

For the expansive and compressive scenarios, we re-estimate labor income for tax unit i  \((Y_{1,i}^{L}\)) using the following specification:

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

Where \(\lambda\) is the ratio of the post- to pre-shock standard deviation of \(\ln Y_{j,i}^{L}\) restricting attention to only those tax units with positive labor income. Here, \(\mu_{0}\) is the weighted log-mean of positive baseline labor income and \(\mu_{1}\) is solved so the positive-subset aggregate matches \(Y_{1}^{L}\) after scaling. 

Under the assumption that any shift in distribution will be correlated with the size of the AI shock, we estimate \(\lambda\) as a function of the GDP shock \(g_{Y}\) so that:  \(\lambda = 1 \pm k \cdot g_{Y}\), where \(1 - k \cdot g_{Y}\) is compressive and \(1 + k \cdot g_{Y}\) is expansive. The default \(k = 1\) is the proportional-to-\(g_{Y}\) mapping.

To further explore the fiscal picture, we calculate scenario-specific changes in revenue-to-GDP shares using the approach described above. We then use the Budget Lab’s Small Macro Model (BLSMM) to determine what these revenue changes mean for the debt-to-GDP ratio in 2030. To keep things as consistent as possible between the BLSMM and tax models, we simultaneously raise the potential output growth rate assumed in BLSMM until actual GDP growth in 2030 is equal to the rate assumed in the scenario-specific Karger et al. (2026) forecast.16 Importantly, the BLSMM model assumes CBO’s so-called rules of thumb, which dictate how outlays (as a share of GDP) respond to changes in potential output. As discussed in prior Budget Lab analysis, an AI shock could be different than others in the sense that policymakers might want to spend on displaced worker support. If that occurs, outlays would likely rise more than expected by BLSMM, increasing debt relative to our calculations.