Vector 1
AI Strategy & Economics

The Payroll Paradox: Demographics, AI, and the American Economy to 2100

Brian J. Curry
Brian J. Curry
September 23, 2026 · 25 min read

How a more productive economy can leave Social Security’s payroll base behind

Vector1 Research | September 2026

Everything behind this analysis is available to paid subscribers at brianjamescurry.substack.com. That includes the full V1-DAI model (the US demographics and AI temporal structural causal model), the analysis pipeline that reproduces every figure and table, the 1,200-draw Monte Carlo dataset, the 720-combination policy search, and the dated forecast claims file for scoring the model as new data arrives. Run it, change the assumptions, and test your own scenarios.

Abstract

An economy can produce more while the share of income supporting its retirement system declines. This paper examines that possibility through V1-DAI, an annual, age-structured scenario model of the U.S. economy from 2026 to 2100. The framework connects fertility, mortality and migration to labor supply, AI diffusion, output, labor compensation and Social Security financing.

The central mechanism is the payroll paradox: AI can increase output while reducing total labor compensation relative to a counterfactual without further AI adoption. Whether that happens depends on the productivity gains from AI, its effect on labor’s income share, and the response of capital investment. A larger economy does not automatically generate a proportionately larger payroll tax base.

The reported simulations also suggest that immigration strongly affects aggregate economic size, while productivity assumptions dominate the range of outcomes for GDP per person. Expanding the Social Security tax base to include capital income can improve financing in the model, although the required coverage depends on tax design and behavioral responses.

Research status. The quantitative V1-DAI results below are retained from the initial analysis and remain provisional. The underlying code, run outputs and figures were not supplied with the manuscript reviewed here. External benchmarks have been checked, but the simulations have not been independently reproduced. A mismatch between the stated fertility assumptions and the 2026 Trustees Report prevents the original validation claim from being accepted without a rerun.

1. Why examine the United States separately?

Benzell, Kotlikoff and Ye study demography and automation in a global model covering 17 regions and more than 150 countries. Replacing the UN’s 2017 demographic projections with its 2024 projections changes the model’s allocation of world output in 2100: the U.S. share rises from 11.2% to 14.4%, while China’s falls from 25.6% to 14.9%. These are comparisons between forecast vintages for the same future year, not changes from today’s GDP shares. [1]

The paper also reports a U.S. share of 9.2% in 2100 if future U.S. immigration is eliminated, and 25.3% under its transformative automation scenario with the 2024 demographics. Its framework includes internationally mobile capital and pension financing. [1]

Those results motivate a more specific domestic question. How much of an AI-driven increase in American output becomes taxable earnings, and how much becomes income outside Social Security’s principal financing base?

V1-DAI examines that question through a simplified U.S. model. It is a separate scenario exercise, not a replication of the NBER study or an empirically identified estimate of AI’s future causal effects.

2. The empirical record as of September 2026

Demographics

CBO’s January 2026 outlook projects the Social Security area population rising from 349 million in 2026 to 364 million in 2056. This population measure is broader than the Census Bureau’s resident population. Deaths exceed births beginning in 2030. Fertility falls from 1.58 births per woman in 2026 to approximately 1.53 in 2036, then remains near that level through 2056. The ratio of people aged 25 to 64 to those aged 65 and older falls from 2.7 to 2.2. [2]

The 2026 Social Security Trustees assume fertility reaches 1.75 by 2050. Their previous report assumed an eventual rate of 1.90. CBO’s January comparison used that earlier Trustees vintage, which explains the 1.90 figure in the initial manuscript. Comparing current projections requires using the updated assumption. [3]

Immigration

The sources agree that migration slowed sharply, although they disagree about its level.

  • CBO: Estimated net immigration of 410,000 in 2025, down from the 2.0 million it had projected a year earlier. [2]
  • Brookings and AEI researchers: Estimated calendar-year 2025 net migration between negative 295,000 and negative 10,000. Their January outlook for 2026 ranged from negative 925,000 to positive 185,000. Negative migration was therefore a scenario, not a settled observation for the full year. [4]
  • Census Bureau: Estimated approximately 1.3 million net international migrants between July 1, 2024, and June 30, 2025, and projected roughly 321,000 for the following year if prevailing trends continued. [5]

These figures use different reference periods and estimation methods. They should not be presented as interchangeable measurements of the same flow.

A Federal Reserve staff note estimates that potential labor-force growth could fall below 10,000 people per month in 2026, bringing the employment growth consistent with a stable unemployment rate close to zero. That is a conditional estimate of labor supply, not evidence that every weak payroll report reflects healthy conditions. [6]

AI adoption

The Census Bureau’s Business Trends and Outlook Survey recorded AI use among 17% to 20% of businesses from December 2025 through May 2026. Use reached 37% among firms with at least 250 employees. The question changed in November 2025 to cover AI use in any business function, which limits comparisons with earlier survey readings. [7]

Worker surveys measure something different. The Real-Time Population Survey found that 39.2% of employed adults used generative AI for work in the previous week in the second quarter of 2026. Reported hours assisted by AI reached 6.3% of work hours, while reported hours saved reached 2.2%. These are self-reported measures. Hours saved do not directly measure the share of tasks automated or the increase in economy-wide productivity. [8]

A Boston College research brief cites earlier survey evidence that 18% of workers aged 50 to 64 used generative AI, below adoption among workers in their 30s and 40s. This supports investigating age differences, but it does not establish a permanent adoption penalty for every worker aged 55 or older. [9]

Productivity and income distribution

BLS reports that nonfarm business labor productivity increased 2.2% over the year to the second quarter of 2026. Labor compensation accounted for 52.8% of nonfarm business output, the lowest share in that series, which begins in 1947. [10]

That labor share includes compensation beyond cash wages and covers a particular sector. Applying it to total GDP or taxable earnings requires an explicit accounting bridge. Its recent decline also does not, by itself, identify AI as the cause.

Social Security

The 2026 Trustees project that the Old-Age and Survivors Insurance fund will exhaust its reserves in the fourth quarter of 2032, when continuing income would cover 78% of scheduled benefits. On a hypothetical combined OASDI basis, depletion occurs in the third quarter of 2034, with 83% payable. The funds are legally separate. The combined 75-year actuarial deficit is 4.42% of taxable payroll, up from 3.82% in the preceding report. [11]

Evidence behind the mechanisms

Maestas, Mullen and Powell estimate that a 10% increase in the population share aged 60 and older reduced GDP per person by 5.5% in their historical U.S. state-level analysis. About two-thirds of the estimated effect operates through labor productivity. Extending that relationship to national TFP over the rest of this century requires additional assumptions. [12]

Acemoglu and Restrepo find evidence that population aging encourages industrial automation, with robots substituting especially for workers aged 36 to 55. Whether the same age mechanism governs generative AI remains an open empirical question. [13]

Acemoglu estimates a ten-year AI contribution to TFP of no more than 0.66% under his assumptions. Penn Wharton Budget Model projects TFP and GDP levels approximately 1.5% higher by 2035, nearly 3% higher by 2055 and 3.7% higher by 2075. These are conditional level effects, not annual growth rates. [14, 15]

3. The V1-DAI framework

3.1 Structure and assumptions

V1-DAI is designed as a temporal structural scenario model: population, labor, technology and fiscal outcomes evolve through specified annual mechanisms. Interventions change those mechanisms or their parameters. The structure makes assumptions inspectable, but does not establish their causal validity.

Population. A single-year cohort model tracks ages 0 to 100. The stated specification uses a Gompertz mortality schedule, fertility scaled to each scenario’s total fertility rate, and a young-adult migrant age profile. Life expectancy follows the CBO path through 2056, then rises by an assumed 0.08 years per year. The extension beyond 2056 is a model assumption. Matching life expectancy alone does not validate age-specific survival, especially at infancy and very old ages.

Labor. Age-specific participation and efficiency weights translate population into effective labor. A policy lever raises participation at ages 55 to 74 and reduces claiming at ages 62 to 69. The one-for-one link between participation and claiming is imposed. Working and collecting benefits are not mutually exclusive, so this lever needs a more detailed interpretation before it can represent a specific retirement reform.

AI diffusion. The model uses τ, pronounced tau, as an effective AI task share. It starts at 2.2% in 2026 and follows logistic diffusion. That starting value was motivated by the RPS hours-saved measure, but equating hours saved with task exposure is an unvalidated mapping. If τ already measures realized savings, multiplying it by another cost-saving parameter can double count the same mechanism. A reproducible calibration must define τ as an economically weighted exposure measure and explain how it is inferred from the survey.

Two demographic channels influence diffusion. An induced-automation term responds to the ratio of people aged 56 to 75 to those aged 36 to 55. An adoption term assigns workers aged 55 and older a relative uptake rate of 0.5. Both are scenario assumptions informed by earlier research, rather than estimated U.S. generative-AI relationships. The model also lowers the eventual AI ceiling when older-worker uptake is lower, a stronger assumption than merely delaying adoption.

AI and distribution. Incremental TFP is specified as the increase in τ multiplied by average task-level cost savings, π. The labor-income share falls by λ for each unit increase in τ. These relationships impose a productivity effect and a distributional effect separately. Neither is inferred from the recent labor-share data.

TFP and output. Trend TFP growth is assumed to be 0.9% annually. An aging term responds to the log change in the population share aged 60 and older. The central elasticity is 0.183, half the manuscript’s productivity-channel interpretation of the Maestas estimate. Halving it is a sensitivity choice, not an estimated national correction.

Output uses a Cobb-Douglas production function with capital elasticity α = 0.35. Capital adjusts to an imposed real-return path falling from 4.0% to 2.5% by 2100. The implementation must specify whether this return is gross or net of depreciation. It must also reconcile the separate, changing labor-share rule with fixed production elasticities. Under competitive factor pricing and constant returns, a conventional Cobb-Douglas model would imply a constant labor share of 1 − α. V1-DAI therefore needs an explicit distributional wedge or a different production specification.

Social Security. Beneficiary counts combine age-specific claiming with disability and survivor recipients. Average benefits follow a 12-year moving average of wages. Taxable payroll is set at 64% of modeled labor income. A benefit parameter is calibrated to the official 2026 OASDI cost rate of 15.37% of taxable payroll. [16]

This is a simplified fiscal block. Actual retirement benefits depend on individual earnings histories, wage indexing, a progressive benefit formula, claiming age and subsequent cost-of-living adjustments. The taxable maximum also means that earnings distribution matters. A fixed payroll fraction and moving-average benefit rule cannot capture all these responses. [17]

3.2 AI scenarios

The stated scenario inputs are:

  • No additional AI: The task index remains at 2.2%. Incremental cost savings and TFP gains are zero. This freezes further adoption; it does not remove AI already in use.
  • Baseline AI: Nominal task ceiling of 25% and task-level cost savings of 15%. The stated formula gives a maximum incremental TFP effect of approximately 3.42% before any age adjustment to the ceiling.
  • Accelerated AI: Nominal task ceiling of 40% and cost savings of 30%, implying approximately 11.34% incremental TFP before that adjustment.
  • Transformative AI: Nominal task ceiling of 60% and cost savings of 50%, implying approximately 28.9% incremental TFP before that adjustment.

These ceilings are assumptions, not measured automation rates. The reported labor shares in 2056 are 52.8%, 47.1%, 43.4% and 38.4%, respectively. Their consistency with the age-adjusted ceilings must be checked against the implementation.

The baseline’s near-term TFP effect is compared with PWBM. Similarity at one date does not establish agreement with its full path or GDP response, particularly because V1-DAI adds its own capital adjustment. The stronger scenarios are independent parameterizations.

3.3 Benchmark checks

The initial manuscript reports these comparisons:

  • People aged 25 to 64 per person aged 65 or older, 2026: CBO, 2.7; V1-DAI, 2.7.
  • The same ratio in 2056: CBO, 2.2; V1-DAI, 2.2.
  • Population in 2056: CBO, 364 million; V1-DAI, 372 million.
  • First year deaths exceed births: CBO, 2030; V1-DAI, 2032.
  • OASDI cost rate in 2056: Trustees, 17.61% of taxable payroll; model comparison run, 16.85%.
  • OASDI cost rate in 2100: Trustees, 20.02%; model comparison run, 19.97%.
  • 75-year financing measure: Trustees’ actuarial deficit, 4.42%; model shortfall measure, 4.43%.
  • AI-related TFP level gain by 2035: PWBM, 1.5%; V1-DAI, 1.42%.

Sources for the external benchmarks are CBO, SSA and PWBM. [2, 15, 16]

The demographic discrepancies are material to a model whose results depend on cohort composition. Population is about 2.2% above the CBO benchmark in 2056, and the natural-decrease date is two years late. The fiscal path also undershoots the Trustees’ 2056 cost rate by 0.76 percentage points.

The near-match between 4.43% and 4.42% cannot establish actuarial replication. The manuscript states that its measure excludes initial trust-fund reserves, whereas the official actuarial balance incorporates reserves and a terminal reserve requirement. Discounting, income components and the projection window must also match. [18]

More fundamentally, the stated comparison run uses fertility of 1.90, while the 2026 Trustees use 1.75. Until its actual inputs are inspected and the comparison rerun, it should not be described as reproducing the 2026 Trustees’ assumptions.

4. Scenario results

4.1 Immigration and aggregate economic size

The initial simulations report a population peak near 376 million in the late 2070s under the model’s CBO-informed path. A restrictive migration path peaks around 2040 and falls to 301 million by 2100. Zero net migration produces a peak around 2030 and a population of 256 million in 2100.

These are V1-DAI projections. The path after 2056 is an extrapolation, not a published CBO forecast through 2100.

Relative to the model’s CBO-informed path, zero net migration reduces reported GDP by 15.7% in 2056 and 36.2% in 2100. The restrictive path reduces 2056 GDP by 10.2%. Zero migration also raises the reported 2075 OASDI cost rate by 3.35 percentage points.

The resemblance to the NBER result requires care. Dividing its zero-immigration U.S. world-output share, 9.2%, by its comparison share, 14.4%, gives approximately 63.9%. V1-DAI reports a U.S. GDP-level ratio of 63.8%. [1]

Those ratios are not equivalent. A country’s share of world output changes when its own output changes, when world output changes, or both. Converting the NBER share ratio into a U.S. output ratio requires the ratio of world GDP across its scenarios. The numerical resemblance is not independent validation.

The reported effect on GDP per person is smaller: zero migration lowers it by 6.8% in 2075. Within this specification, immigration affects aggregate economic size more strongly than average output per resident. Neither measure, on its own, identifies the distribution of gains among existing residents or the full fiscal contribution of migrants.

4.2 Age structure and AI diffusion

The model’s population ratio of ages 56 to 75 to ages 36 to 55 falls from 0.92 in 2026 to 0.79 in 2040. This is a population ratio, not automatically a ratio of employed workers.

Under the specified induced-automation mechanism, that decline weakens the incentive to automate. Turning the channel off raises the reported 2040 AI task index from 19% to 20%. Raising older workers’ relative uptake from 0.5 to 0.85 lifts it to 22% and increases reported 2075 GDP per person by 0.6%.

The result supports testing whether training can reduce adoption gaps. It does not establish that workforce age is the binding constraint on U.S. AI adoption through 2040. Occupation, organizational investment, technology costs, management decisions and task reallocation are not separately identified here.

The model also reduces the eventual AI task index by roughly one to two percentage points through its age-uptake rule. Firms’ ability to automate tasks without requiring each incumbent worker to adopt AI could weaken that channel.

4.3 AI and living standards

Across the reported 1,200 Monte Carlo draws, median annual real GDP-per-person growth from 2026 to 2075 is 1.51%. The central 90% of simulated outcomes runs from 0.91% to 2.16% annually. The initial analysis attributes approximately 0.1 percentage point of annual growth to baseline AI and 0.8 percentage points to transformative AI, relative to frozen adoption under the same demographics.

This interval describes variation under the chosen parameter distributions. It is not a validated 90% confidence interval for the actual economy.

The manuscript reports the following standardized sensitivity scores for GDP per person in 2075:

  • Trend TFP growth: +0.83.
  • AI capability: +0.40.
  • Aging-productivity effect: −0.27.
  • Long-run migration: +0.10.
  • Long-run fertility: +0.02.

The estimator behind these scores must be specified before they can be interpreted as regression coefficients, correlations or variance contributions. The ranking also depends on how widely each parameter is allowed to vary.

Within the reported exercise, productivity assumptions account for more variation in GDP per person than fertility or migration. That conclusion is specific to this model and its parameter ranges. It does not diminish the importance of demography for aggregate output or retirement financing.

4.4 The payroll paradox

The central accounting relationship is straightforward:

Total labor compensation = labor share × output.

In the reported baseline AI scenario, 2056 GDP is 5.2% above the frozen-adoption counterfactual, while the labor share falls from 52.8% to 47.1%. Using those numbers:

Compensation ratio = 1.052 × (0.471 ÷ 0.528) ≈ 0.938.

Total labor compensation is therefore approximately 6.2% lower, despite higher output. Calling this a wage-bill result requires an additional assumption about how total compensation maps into wages and taxable earnings.

The transformative scenario needs numerical reconciliation. The stated 40% GDP increase and 38.4% labor share imply:

Compensation ratio = 1.40 × (0.384 ÷ 0.528) ≈ 1.018.

Those inputs imply compensation approximately 1.8% higher, not the 7.4% reported in the original analysis. This arithmetic check does not replace a model rerun; it identifies an inconsistency among the published figures.

A threshold can be derived under a restricted version of the model. Let α be the production elasticity of capital, π the task-level cost saving, λ the decline in labor share per unit of additional AI exposure, and s₀ the initial labor share. Holding effective labor and the capital user cost fixed, the local condition for compensation to rise is:

π ÷ (1 − α) > λ ÷ s₀.

With α = 0.35, λ = 0.25 and s₀ = 0.528, the local threshold is approximately 30.8%.

For a finite exposure increase Δτ, with the TFP level ratio specified as 1 + πΔτ, the corresponding compensation ratio is:

W_AI ÷ W_noAI = [(s₀ − λΔτ) ÷ s₀] × (1 + πΔτ)^(1 ÷ (1 − α)).

Compensation rises when this expression exceeds one. The finite-change threshold generally differs from the local 30.8% approximation. It also depends on the distributional assumptions described in Section 3.1. There is no universal 31% cutoff established by the evidence.

The literature does not justify saying all central cost-saving estimates fall well below that threshold. PWBM assumes average labor-cost savings of roughly 25% initially, increasing toward 40% over subsequent decades. Translating those estimates into V1-DAI’s π requires consistent task and cost definitions. [15]

The reported fiscal results vary much less than GDP across AI scenarios. Under the model’s CBO-informed demographics, the 75-year shortfall measure ranges from 5.44% to 5.57% of taxable payroll, while the 2075 cost rate rounds to 21.1% in all four cases.

That similarity partly reflects the benefit rule. Higher earnings increase payroll revenue but also feed into later benefit payments. Lower earnings can reduce both. A nearly unchanged cost-to-payroll ratio does not mean AI has no effect on revenue, benefits or recipients’ living standards.

The defensible conclusion is that stronger output growth need not repair a payroll-financed system. Its effect depends on where income accrues, which earnings are taxable, and how benefits respond.

4.5 Demographic assumptions and the financing gap

The initial analysis reports these 75-year model shortfall measures:

  • CBO-informed demographics: 5.57% of taxable payroll.
  • Restrictive migration: 6.52%.
  • Zero migration: 7.21%.

It also reports a 2075 cost rate of 21.1% under CBO-informed demographics and 24.5% under zero migration. The official 2026 Trustees’ intermediate cost rate for that year is 19.90%. [16]

These comparisons suggest demographic assumptions deserve close attention. They do not establish that the Trustees understate the shortfall by a particular amount. The comparison must first reconcile report vintages, population definitions, economic assumptions, benefit rules and actuarial accounting.

The reported Monte Carlo median shortfall is 5.80%, and 99.6% of draws exceed the official 4.42% benchmark. That percentage is a frequency within the chosen simulation, not a 99.6% probability that the Trustees are wrong. The current disagreement concerns fertility paths converging toward 1.53 and 1.75, not 1.53 and 1.90. [2, 3]

4.6 Attribution of the 2075 outlook

The analysis reports an exact Shapley decomposition over all 16 combinations of four mechanisms: lower fertility, a migration reset, the aging-productivity effect and baseline AI.

Its stated contributions are:

  • Lower fertility: GDP per person in 2075 increases by 0.3%; the 75-year shortfall rises by 0.51 percentage points of payroll.
  • Migration reset: GDP per person falls by 2.3%; the shortfall rises by 0.88 points.
  • Aging-productivity effect: GDP per person falls by 8.6%; the shortfall rises by 0.06 points.
  • Baseline AI: GDP per person rises by 5.1%; the shortfall rises by 0.06 points.

These are allocations of modeled differences across specified counterfactuals. Shapley values do not independently establish causation. The fertility comparison here is 1.70 versus 1.53, which differs from the Trustees comparison discussed elsewhere.

The decomposition suggests that productivity and demographics affect different outcomes with different strengths. The mechanisms still interact, and the result does not show that growth policies and financing policies are substitutes for one another.

4.7 Policy search

The reported grid search evaluates 720 combinations across five levers:

  • Long-run net migration of 500,000 to 1.7 million people annually.
  • An assumed migrant skill premium of zero to 30%.
  • Participation and claiming adjustments of up to nine percentage points.
  • Older-worker relative AI uptake of 0.5 to 0.8.
  • Inclusion of zero to 30% of capital income in an additional financing base.

The initial policy case has a reported shortfall of 6.53% of payroll. The stated results are:

  • Raising migration to 1.7 million reduces the shortfall by 1.39 percentage points and raises 2075 GDP per person by 6.2%.
  • Skill selection raises GDP per person by 8%, but improves the financing measure by only 0.06 points.
  • The maximum later-work and later-claiming adjustment reduces the shortfall by 1.32 points.
  • Higher older-worker AI uptake raises output slightly, with little fiscal effect.
  • Combining the growth and participation levers leaves a reported shortfall of 3.76%.

These effects cannot be assumed to add across arbitrary policy sequences. The precise comparison case and near-term migration path must be documented for each run.

The initial analysis also reports the following capital-income inclusion shares as sufficient to eliminate its summarized shortfall:

  • Status quo with baseline AI: 29.7%.
  • CBO-informed demographics without additional AI: 30.0%.
  • CBO-informed demographics with baseline AI: 25.2%.
  • CBO-informed demographics with transformative AI: 17.7%.
  • All growth levers with baseline AI: 16.6%.
  • All growth levers with transformative AI: 11.1%.

These remain provisional model outputs, not tax recommendations. An inclusion share is different from a tax rate. If a fraction c of a defined capital-income base K_income is taxed at rate t, annual revenue before behavioral responses is:

Additional revenue = t × c × K_income.

For illustration, applying the current combined 12.4% OASDI payroll rate to 20% of a capital-income base would impose an effective levy of 2.48% on that entire base before exemptions and responses. The manuscript must specify whether that is the intended rule. [11]

One numerical discrepancy also remains. The stated policy waterfall reduces the shortfall from 3.76% to 1.27% when 10% of capital income is included. With linear revenue and an unchanged denominator, that implies break-even inclusion of about 15.1%, rather than 16.6%. A different scenario, nonlinear response or calculation convention could explain the difference, but the outputs are needed to determine which.

The policy result is consequently narrower than a claim that capital taxation solves Social Security. A broader revenue base improves the model’s accounting balance. Whether it achieves durable solvency depends on the definition of taxable capital income, collection, investment responses, international ownership, annual cash flows and benefit obligations. Closing a 75-year summarized gap does not guarantee that reserves stay positive in every year.

5. Forecasts that can be evaluated

The initial analysis reports event frequencies from 1,200 parameter draws:

  • 2030: AI task exposure exceeds 5%. Reported frequency: 74%.
  • 2030: Nonfarm business labor share falls below 51%. Reported frequency: 2%.
  • 2035: Nonfarm business labor share falls below 50%. Reported frequency: 57%.
  • 2040: OASDI costs exceed 16.5% of taxable payroll. Reported frequency: 62%.
  • 2056: Social Security area population is below 364 million. Reported frequency: 59%.
  • Long-run model comparison: The summarized shortfall exceeds 4.42% of payroll. Reported frequency: 99.6%.

The first claim cannot be scored against RPS hours saved without a defined measurement bridge. It must either forecast the survey statistic directly or identify an observable task-exposure series. Changing the outcome definition would require recalculating its probability.

The final claim is a comparison between model outputs, not yet a dated, observable event. It needs a resolution date and an agreed outcome before it can be scored as a forecast.

For the remaining claims, a forecasting record should specify the exact series, annual or quarterly frequency, data vintage and revision policy. Brier scores can then evaluate the probability assignments as outcomes become observable. A failed forecast should prompt investigation of the relevant mechanisms rather than automatically being attributed to AI adoption alone.

6. Limitations

The initial age distribution is stylized rather than a complete official cohort table. Migrants largely share native-born participation and fertility assumptions apart from the skill lever. Those choices can affect births, labor supply and the timing of benefit claims.

The economy’s capital-return path is imposed. The model does not solve the global saving, investment and ownership relationships that determine how additional domestic output translates into residents’ income. GDP per person is therefore an incomplete measure of living standards.

AI exposure, time savings and substitution are only approximately connected. The labor-share rule omits an explicit mechanism for new labor-intensive tasks, changing bargaining power and sectoral reallocation. The national aging-productivity effect is also uncertain.

The fiscal block simplifies benefit indexing, delayed claiming, earnings inequality, disability and survivor transitions, and the taxable maximum. It must avoid double counting recipients across benefit categories. Higher benefits following delayed claiming could offset part of the apparent saving from postponement. Medicare and the wider fiscal effects of migration are outside the model.

Capital-income coverage is an accounting experiment until the taxable base and behavioral responses are specified. Gross capital income, net income after depreciation, corporate profits and household investment income are not interchangeable bases.

Finally, independent parameter draws omit dependence among assumptions. Introducing correlations could widen or narrow the outcome distribution. The direction cannot be known without specifying those relationships.

7. Conclusions

The payroll paradox is a distributional and fiscal possibility worth examining. Output can rise while total labor compensation falls relative to a counterfactual, and a retirement system financed mainly through taxable earnings can capture only part of a productivity gain.

The reported simulations suggest that immigration affects the scale and age structure of the economy, while productivity assumptions strongly influence output per person. They also illustrate how broader revenue coverage can connect retirement financing to income generated outside payroll.

The quantitative strength of those conclusions remains conditional. V1-DAI’s calibration must be updated to the 2026 Trustees’ demographic assumptions, its labor-share mechanism reconciled with production, and its reported outputs reproduced. The policy search establishes only what happens within its tested choices and simplified fiscal rules. It cannot show that every feasible combination of immigration, retirement or training policies would fail.

The question remains useful even before those estimates are settled: if AI changes who receives the income from production, should retirement financing continue to rely on the same income base?

8. Methods and reproducibility

The initial manuscript describes a self-contained NumPy implementation inspired by temporal structural causal models and a grid search over policy combinations. It refers to pycausalsim.futures and Papilon as design influences. Compatibility with their APIs or file formats has not been established in this review.

The manuscript also identifies code/run_analysis.py and random seed 20260923 as its intended reproduction entry point. No executable code, data outputs or figure files accompanied the supplied article, so this revision does not claim that the results have been rerun or that those materials are attached.

A reproducible release should provide the executable model, dependency versions, dated input data, parameter distributions, complete scenario definitions, output files and all eight figures. It should document the actuarial shortfall formula and show the full benchmark paths, including discrepancies.

Figure captions should distinguish official data from simulations and identify the scenario and run version. Quantitative figures affected by the corrections above must be regenerated from the reconciled model.

About the Author

Brian Curry is a Kansas City-based machine learning and AI strategist, marketing economist, and applied researcher. He leads Vector1, an independent research studio working on cognitive AI, causal inference, marketing economics, and knowledge engineering, with a focus on simulation-driven intelligence systems.

Over more than twenty years, Brian has led data science and analytics initiatives at enterprises ranging from $4B to $125B in revenue, including Koch Industries, Tractor Supply, Vail Resorts, Hallmark, Garmin, and AT&T. His work applies economics and causal inference to connect marketing investment to structural outcomes, operational decisions to economic value, and AI adoption to durable competitive advantage.

He is also the founder of KC AI Lab, a 1,500-member AI and ML education community in the Midwest.

Research and Open-Source Projects

The modeling approach in this paper draws on PyCausalSim and Papilon.

Causal Inference and Economics

  • PyCausalSim: Simulation-driven causal discovery and evaluation, using counterfactual scenarios to isolate structural drivers.
  • Papilon: Complex-system simulation and optimization, with applications in media mix modeling and adaptive architectures.
  • Daedalus: Causal analytics for money-flow mapping and macroeconomic tracking.
  • The Autonomy Economy: A macroeconomic framework for the combined effects of rapid productivity growth and structural price deflation in an AI-automated economy.

Forecasting and Applied Modeling

  • Regimecast: Time-series regime detection, structural breaks, and economic forecasting.
  • CarryNet: Financial modeling for commodity carry trade analysis.
  • Foodmesh-ml: Machine learning, optimization, and simulation for distributed food systems.

Cognitive AI and Semantic Architecture

  • Memory-Node Encapsulation (MNE): A cognitive AI architecture and data structure for artificial episodic memory (NeoCortex-M).
  • MeaningFlow: Semantic content modeling, embeddings, clustering, and coverage gap analysis.

Contact: brian@vector1.ai | GitHub @Bodhi8 | Substack

The text of this article may be shared with attribution under Creative Commons Attribution 4.0 International (CC BY 4.0). Code and models referenced here are available separately to paid subscribers.

References

[1] Benzell, Seth G., Laurence J. Kotlikoff, and Victor Yifan Ye. 2026. “The Global Transition: The Impact of Demographics and AI on Economic Power.” NBER Working Paper 35618, August. The working-paper abstract supports the global figures summarized here.
https://www.nber.org/papers/w35618

[2] Congressional Budget Office. 2026. “The Demographic Outlook: 2026 to 2056.” January 7. Its comparison with the Trustees uses the earlier Trustees report, not the June 2026 update.
https://www.cbo.gov/publication/61994

[3] Social Security Administration. 2026. “Demographic Assumptions and Methods,” Section V.A of the 2026 OASDI Trustees Report. Intermediate fertility reaches 1.75 in 2050.
https://www.ssa.gov/oact/TR/2026/V_A_demo.html

[4] Edelberg, Wendy, Stan Veuger, and Tara Watson. 2026. “Macroeconomic Implications of Immigration Flows in 2025 and 2026: January 2026 Update.” Brookings Institution.
https://www.brookings.edu/articles/macroeconomic-implications-of-immigration-flows-in-2025-and-2026-january-2026-update/

[5] U.S. Census Bureau. 2026. “New Population Estimates Show Historic Decline in Net International Migration.” January 27.
https://www.census.gov/newsroom/blogs/random-samplings/2026/01/historic-decline-in-net-international-migration.html

[6] Murray, Seth, and Ivan Vidangos. 2026. “Labor Force Growth, Breakeven Employment, and Potential GDP Growth.” FEDS Notes, April 2. Board of Governors of the Federal Reserve System.
https://www.federalreserve.gov/econres/notes/feds-notes/labor-force-growth-breakeven-employment-and-potential-gdp-growth-20260402.html

[7] U.S. Census Bureau. 2026. “Large Firms With at Least 20 Employees Biggest AI Users.” May. Business Trends and Outlook Survey.
https://www.census.gov/library/stories/2026/05/ai-use-businesses.html

[8] Federal Reserve Bank of St. Louis. 2026. “Does Generative AI Save Time at Work?” FRED Blog, August. Based on Real-Time Population Survey measures associated with Alexander Bick, Adam Blandin and David Deming.
https://fredblog.stlouisfed.org/2026/08/does-generative-ai-save-time-at-work/

[9] Sanzenbacher, Geoffrey T. 2026. “Are the Careers of Older Workers Being Cut Short by AI?” Issue in Brief 26–13. Center for Retirement Research at Boston College, June 30. The age-specific adoption statistic is attributed there to Bick, Blandin and Deming (2025).
https://crr.bc.edu/are-the-careers-of-older-workers-being-cut-short-by-ai/

[10] Bureau of Labor Statistics. 2026. “Productivity and Costs, Second Quarter 2026, Revised.” September 3.
https://www.dol.gov/newsroom/economicdata/prod2_09032026.pdf

[11] Social Security and Medicare Boards of Trustees. 2026. “A Summary of the 2026 Annual Reports.” June 9. Includes depletion dates, payable-benefit percentages and current payroll-tax rates.
https://www.ssa.gov/oact/trsum/

[12] Maestas, Nicole, Kathleen J. Mullen, and David Powell. 2023. “The Effect of Population Aging on Economic Growth, the Labor Force and Productivity.” American Economic Journal: Macroeconomics 15(2). Working-paper version: NBER Working Paper 22452.
https://www.nber.org/papers/w22452

[13] Acemoglu, Daron, and Pascual Restrepo. 2022. “Demographics and Automation.” Review of Economic Studies 89(1). Working-paper version: NBER Working Paper 24421.
https://www.nber.org/papers/w24421

[14] Acemoglu, Daron. 2024. “The Simple Macroeconomics of AI.” NBER Working Paper 32487.
https://www.nber.org/papers/w32487

[15] Arnon, Alexander. 2025. “The Projected Impact of Generative AI on Future Productivity Growth.” Penn Wharton Budget Model, September 8.
https://budgetmodel.wharton.upenn.edu/p/2025-09-08-the-projected-impact-of-generative-ai-on-future-productivity-growth/

[16] Social Security Administration. 2026. “Annual Income Rates, Cost Rates, and Balances.” Single-year data, 2026 OASDI Trustees Report, intermediate assumptions.
https://www.ssa.gov/oact/TR/2026/lr4b1.html

[17] Social Security Administration. “Social Security Retirement Benefit Calculation.” Office of the Chief Actuary.
https://www.ssa.gov/oact/ProgData/retirebenefit1.html

[18] Social Security Administration. 2026. “Long-Range Estimates,” Section IV.B of the 2026 OASDI Trustees Report. Includes definitions and calculations for summarized income, cost and actuarial balance.
https://www.ssa.gov/OACT/TR/2026/IV_B_LRest.html

Related research