The Price of Superintelligence

The ten labor-market regions above are drawn from the same 2024 IMPLAN occupation data, broken out one level further, to the county. Two neighbors inside the Finger Lakes region show why that finer view matters: Monroe County, anchored by Rochester, and Yates County, a small rural county of vineyards and farms an hour to the south. Averaged into one region, their very different economies, and their very different exposure to AI, mostly disappear.

Monroe County counts about 400,000 jobs, forty-nine times Yates County's roughly 8,100 — a reminder that "the Finger Lakes region" is really one big city county and eight small rural ones stacked together. Under the same moderate scenario used throughout this tool, an estimated 6.2 percent of Monroe's jobs are exposed to AI-driven displacement, above the region as a whole and closing in on the statewide rate of 6.3 percent. Yates sits noticeably lower, at 5.4 percent.

The gap traces to what people in each county do for work. Monroe is a metro economy anchored by Rochester's hospitals, insurers, colleges, and remaining imaging and optics firms: about one job in nine is office or administrative support, and computer and mathematical work — a heavily exposed category in this model — claims more than twice the employment share it does in Yates. Yates, home to the vineyards along Keuka Lake, runs the other way: farming claims 5.5 percent of its jobs against Monroe's 0.2 percent, and production work is also a larger share of the county's economy. Both kinds of work are more hands-on, and score lower on the exposure measure behind this model.

Neither county's number is extreme. But an 0.8-point gap between two counties inside the same labor region is exactly the kind of local variation a statewide or regional figure alone cannot show.

The map above shows where AI's reach falls hardest — which occupations, which regions. It does not say what to do about it. No one yet knows how, where, or when the losses will land. But that uncertainty is not a reason to wait. Several moves make sense now, before the picture sharpens, because they help workers whether the impact turns out large or small.

Begin with services for the workers most exposed. The state can help them build basic AI skills now, working through corporate partners, colleges and universities, libraries, and community organizations. It can help displaced workers find alternative careers that build on the education and experience they already have. And it can help both new workers and displaced ones use the AI business tools now widely available to work for themselves or start a small business — the same technology that puts some jobs at risk also lowers the cost of starting one. The clearest displacement signal in the evidence so far falls on the young: one study found employment down about 16 percent among workers aged 22 to 25 in the most-exposed occupations. Training that reaches people before their first job, not only those already in work, has the strongest claim on the state's attention.

Averages also hide what happens beneath them. Statewide and regional numbers mask wide differences at the county and local level. A region with a small share of the state's jobs can still carry a large share of its own jobs at risk. This tool can be run at the county and sub-county level to find those local concentrations, so that help can be aimed where it is needed most.

AI is only one of the forces bearing on New York's economy, and it should be weighed against the others. The state's population is aging. It is drawing fewer immigrant workers and international students, and losing some workers to other states — trends that together point to a shrinking labor force. That scarcity can itself pull automation forward: over the past seventy years, economies short of younger workers have tended to meet the gap with labor-saving technology, leaving overall output broadly intact [3]. Long-standing trade and travel with partners such as Canada are under strain. New tariffs and overseas conflict are pushing costs up. Working the other way, New York may gain from new foreign investment. The costs and benefits of AI adoption will play out inside this larger picture, not apart from it.

The state should also keep a close watch on its unemployment insurance system. Even under optimistic assumptions, several hundred thousand New York workers could face some risk of displacement. New York has recently strengthened the system: it repaid the money borrowed from the federal government during the pandemic, raised the wage level subject to unemployment taxes, and lifted the maximum weekly benefit to $869. Even so, AI-related job loss — added to the other pressures above — could raise the cost of benefits, administration, and services over the next few years. The Department of Labor and the Division of the Budget should keep the fund's health under close review.

Our methodology for calculating the number of exposed jobs in each occupation consists of two steps — an exposure calculation, repeated 20,000 times in a Monte Carlo simulation.

Step 1. Employment Exposure Calculation

We calculate the number of jobs exposed to AI for 845 occupations tracked in a region using the following formula:

jobs exposed = employment × exposure × adoption × displacement × (1 − protection × augmentation)

Below, we discuss highlights from the literature and how they shape inputs and our assumptions for each term.

Employment. The total number of jobs is the 2024 wage-and-salary employment for each occupation in each of New York's ten labor market regions, drawn from IMPLAN, a commercial regional economic modeling system [4]. Occupations follow the Standard Occupational Classification, the federal taxonomy that divides the workforce into 845 detailed occupations, from registered nurses to bus drivers to computer programmers. Payroll counts cover employees only; the self-employed are not included.

Exposure. We multiply each occupation's employment by an exposure score, which estimates the share of that occupation's work that AI could technically perform. Exposure scores are drawn from published research by Yale Budget Lab, which combines published exposure measures from six separate studies into a single index for each occupation [5].

Researchers have measured exposure in several ways, and the differences matter. Frey and Osborne produced the most famous number in the field — 47 percent of United States employment at risk of computerization — by scoring whole occupations at once [6]. Brynjolfsson and Mitchell scored individual work tasks instead, and found that nearly every job mixes tasks machines can do with tasks they cannot, so whole occupations are rarely automated outright [7]. Felten, Raj, and Seamans measured how much AI capabilities overlap with the human abilities each occupation relies on [8]. Webb analyzed historical automation innovations, comparing the language of patents against the language of job descriptions, and compared exposure predictions based on that comparison with what software and industrial robots actually did to employment between 1980 and 2010 [9]. Eloundou and colleagues asked whether a large language model could cut the time needed for each work activity in half, with and without supporting software built around the model [10].

Measures of AI exposure often disagree. Frank, Ahn, and Moro tested thirteen exposure scores against ten years of state unemployment claims and found that no single score reliably predicted actual job loss; only a combination of all thirteen performed reasonably well [11]. A 2026 review argues that exposure scores are merely snapshots of one AI system's abilities at one moment [12]. Both findings argue for using a blended score, and for reading it as a measure of technical feasibility rather than a forecast. A high exposure score means AI can do much of the occupation's work. It does not mean those jobs will disappear.

Adoption. Feasibility matters only where employers actually put AI to work, which is what the first scenario variable tries to capture, by setting the share of employers assumed to deploy AI in earnest. Census Bureau business surveys report that roughly a fifth of United States firms report using AI; weighted by employment, the share rises to about a third, with the information and finance sectors well ahead of the rest [13]. A global survey by McKinsey found that most organizations report using AI somewhere in the business, but only about 7 percent say they have scaled it beyond a pilot [14]. We set the range of possible values for this variable wide, given the distance between these findings — many firms are trying AI, but few are using it at meaningful scale.

Displacement. Displacement measures the share of adopted AI work that ends in job cuts rather than higher output from the same staff. When an employer adopts AI, some of the work the technology absorbs shows up as layoffs; the rest shows up as the same workers producing more. The framework comes from Acemoglu and Restrepo: automation pushes workers out of existing tasks while newly created kinds of work pull them back in, and what matters for employment is the net of the two forces. Over recent decades, new work offset 50 to 80 percent of what automation displaced [15], which sets our allowed range for this variable, an approach used in other recent studies [16]. Acemoglu and Restrepo also showed that the displacement side is real and measurable: each additional industrial robot per thousand workers lowered the local employment rate in one study by roughly a third of a percentage point [17]. A study of millions of online job postings adds a caution about timing. Firms hiring AI specialists cut their hiring for other roles years before any effect appeared in economy-wide statistics, so AI-related displacement can be under way but remain invisible in the aggregate jobs data [18].

Protection and Augmentation. The final term is a correction used to represent how AI is actually used once it is adopted in an occupation, separating the effects of automation (where the AI performs the task in place of a person) and augmentation (where the AI assists a person who is still doing the job), an approach used in recent research [19].

We compute this correction by multiplying two terms: - Augmentation is the share of an occupation's AI use that assists rather than replaces, drawn from the Anthropic Economic Index. Anthropic classifies millions of anonymized conversations from Claude.ai, the company's consumer chat service, by occupation and by which of the two modes each conversation fits [20]. The augmentation share varies widely by occupation family — about 65 percent for legal work, about 25 percent for protective services — and the model applies each family's own share. - Protection is a global variable (same for all occupations) representing how much the assistance of augmentation actually saves jobs in practice. Assistance alone is no guarantee: an augmented worker may keep a job because human judgment still matters, or lose the job anyway because one assisted worker now produces what three once did. No published study yet measures how protective augmentation really is. The Budget Lab at Yale tracks the same usage data against monthly employment changes and has so far found no detectable relationship between exposure, automation, or augmentation and real employment shifts [21], so this slider remains a judgment bounded by evidence rather than a measured rate.

The augmentation shares rest on the narrowest data in the model. They come from two months of conversations on a single company's consumer chat service, Claude.ai, recorded in April and May 2026 [20]. Those conversations capture what the workers who chose to use Claude brought to it, and don't include the tasks that were never conducted with a chatbot, nor those conducted with AI-enabled enterprise software. This matters because the use of AI on consumer and enterprise platforms can point in different directions, and will change when product offerings change. The low augmentation share for computing work is the clearest warning. It may reflect nothing more than programmers being early to adopt advanced agentic coding assistants, the use of which is classified as automation rather than augmentation[22]. Regardless, this uncertainty has a small impact on the result of the overall calculation. If we replace every occupation's unique augmentation score with a single economy-wide average augmentation score, only about 3 percent of the statewide job exposure shifts between occupations — easing the counts for computing, office support, and protective service while raising them for business, sales, and legal work. The upshot: trust the statewide figure; treat the occupation-by-occupation detail with more caution.

Example. Take customer service representatives, a large and highly exposed occupation. Suppose a region employs 10,000 of them, and the exposure index scores 60 percent of their work as something AI could technically do. Run the model once at moderate settings — adoption at 30 percent, displacement at 35 percent, an office-support augmentation share of 30 percent, and protection at 50 percent (0.5) — and the formula reads: 10,000 employed × 0.60 exposure × 0.30 adoption × 0.35 displacement × (1 − 0.50 protection × 0.30 augmentation). The first three terms carry 10,000 jobs down to 1,800; displacement takes that to 630; and the protection-and-augmentation correction, a factor of 0.85, trims it to about 535 jobs exposed to AI displacement. That is one of 845 occupations in one of the model's 20,000 runs.

Step 2. Monte Carlo Simulation

Since these key variables are not yet precisely measured, we use a Monte Carlo simulation to generate 20,000 simulated futures, using randomly selected values from a representative distribution across the ranges drawn from the literature above. This is a widely used approach in forecasting, because it draws attention to the range of probable outcomes versus a single predicted outcome that will either be right or wrong. For instance, the Social Security Administration projects the program's finances by varying a dozen uncertain assumptions at random across 5,000 simulated futures and publishing the percentile bands [23].

Example. The single run above fixes adoption, displacement, and protection at one set of values. The simulation repeats this calculation 20,000 times, each time sampling the key global variables at random from across their plausible ranges — adoption from roughly 15 to 45 percent, displacement from 20 to 50 percent, protection from 20 to 80 percent — while employment, exposure, and the augmentation share stay put. For our 10,000 customer service representatives, the result is no longer a single number but a spread: a median near 500 jobs exposed, with most runs landing between about 200 and 1,100. The model reports that band rather than the midpoint alone, because the honest answer to "how many jobs" is a range, not a point.

Users should keep in mind several key limitations when interpreting the results of these simulations:

1. It is not net job change. The model counts gross exposure to displacement from AI. We do not estimate the jobs AI creates — new roles, demand induced by cheaper services, growth from higher productivity, and the new businesses AI is helping people start. That last channel is no longer hypothetical: industries more exposed to AI are forming about 20 percent more startups than less-exposed ones [24], nearly 60 percent of 2025's new founders used AI to launch [25], and solo businesses are now forming and reaching real revenue faster than employer firms [26] — none of which this payroll base can see. The output is best seen as an upper bound on net loss, not a layoff prediction.
2. It is not something that has already happened. Exposure measures what AI could technically do. Through early 2026, the Budget Lab at Yale finds no measurable AI-driven displacement in U.S. labor data [21].
3. It has no timeline. The model tells you what happens if AI adoption reaches the level you set — not when that level arrives. Fast or slow, the arithmetic is the same. Nothing pins the outcome to a date. Furthermore, our workforce is frozen in 2024. We use 2024 payroll jobs [4] and hold them fixed — no growth, no migration, no retirements, and no one finding a new job.
4. It counts jobs, not the ladders between them. The calculation scores each occupation on its own and is silent on how workers move between them. However, a recent Brookings analysis finds AI exposure concentrated in the clerical, administrative, and customer-service "gateway" roles that workers without four-year degrees rely on to climb toward higher-wage work — nearly 11 million such workers sit in gateway jobs highly exposed to AI [27]. Even where a particular job survives, the path out of it may narrow, a loss this arithmetic cannot see.

These summaries were prepared with the assistance of AI — Claude's Fable, Opus, and Sonnet models — then reviewed and corrected by the project team. The list is a working map of a fast-moving literature, refreshed as the research moves, not a finished survey.

1. 1

   Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence Brynjolfsson, Chandar & Chen, Stanford Digital Economy Lab, 2025

   Using ADP payroll microdata, the authors find employment among early-career workers (ages 22–25) in the most AI-exposed occupations fell about 16 percent relative to less-exposed peers, while experienced workers in the same occupations held steady. Two credible studies, two data sources, two opposite conclusions about who AI is affecting — the unresolved tension this model's ranges are built to hold open rather than settle.

2. 2

   A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment Kharazian, Simon & Stevens, Ramp/Revelio Labs, 2026

   The most important counterweight in this list. Instead of surveys or exposure scores, the study joins corporate-card AI vendor spending at 21,559 U.S. firms to payroll-derived headcount records. Firms in the top spending quartile grew headcount 10.2 percent over the following two years — including entry-level headcount — while low spenders saw no change. The authors flag the selection problem themselves: heavy AI spenders were already larger and faster-growing, so the correlation is not proof that AI causes growth. It is direct evidence that firm-level adoption and occupation-level exposure can point in opposite directions.

3. 3

   Baby Busts and Growth Booms: Demographic Change and the Macroeconomy Acemoglu, Autor, Beirne & Scott, NBER working paper, 2026

   The one source here about the demographic backdrop AI is arriving into rather than about AI itself. Studying seven decades of falling birth rates across countries and U.S. regions, the authors find that aging and slower population growth have raised GDP per working-age adult with no drop in total output or earnings — the opposite of the usual worry that a shrinking workforce must mean a shrinking economy. Their reading is that technology responds to a scarcity of younger workers by saving labor: places with lower birth rates produce more labor-saving patents and faster productivity growth.

4. 4

   IMPLAN regional economic models IMPLAN Group, 2024 data year

   The source of the employment base: 2024 wage-and-salary employment by six-digit SOC occupation for New York State and its ten labor market regions, extracted under Appleseed's license. IMPLAN assembles county-level employment from federal statistical programs and allocates it to detailed occupations through industry staffing patterns. The counts cover employees only, so proprietors and the self-employed sit outside every estimate on this page.

5. 5

   What We Do and Don't Know About How AI Is Affecting the Labor Market The Budget Lab at Yale, 2026

   The direct source of the exposure term. The Budget Lab standardizes six published exposure indices — built from ability overlap, patent text, task ratings, and usage data — and blends them into a single score per occupation. Using a blend rather than any single index follows the validation evidence in [11]. The same research program also monitors U.S. occupational statistics for signs of AI-driven disruption and, through early 2026, finds the occupational mix shifting no faster than in earlier technology transitions.

6. 6

   The Future of Employment: How Susceptible Are Jobs to Computerisation? Frey & Osborne, Technological Forecasting and Social Change, 2017

   The paper behind the famous claim that 47 percent of U.S. employment is at risk of automation. Machine-learning experts hand-labeled a sample of occupations as automatable or not; a Gaussian-process classifier trained on O*NET job characteristics then extrapolated a computerization probability to 702 occupations. Later work (Arntz, Gregory & Zierahn, 2016) re-estimated the exercise at the task level and cut the at-risk share to roughly 9 percent, showing how much the whole-occupation assumption drives the headline. Included as an influential, contested landmark — not as an input to this model.

7. 7

   What Can Machine Learning Do? Workforce Implications Brynjolfsson & Mitchell, Science, 2017

   An early ancestor of the exposure literature that scored tasks, not occupations. Human raters evaluated individual work tasks against a 21-item rubric of "suitability for machine learning" — criteria like well-defined inputs and outputs, tolerance for error, and no need for long chains of reasoning. Its lasting contribution is structural: jobs are bundles of tasks with very different automatability, so wholesale automation of an occupation is rare. That is the reasoning behind this model treating augmentation as a within-occupation, task-level phenomenon rather than a yes/no switch.

8. 8

   AI Occupational Exposure (AIOE) Felten, Raj & Seamans, Strategic Management Journal, 2021

   The field's standard ability-based measure. Crowdworkers rated the relatedness of ten AI application areas (drawn from the EFF's AI Progress Measurement project) to 52 O*NET human abilities, such as inductive reasoning and oral comprehension; an occupation's score aggregates those ratings weighted by each ability's importance and prevalence in the job. The authors are explicit that a high score measures capability overlap, not predicted job loss. One of the published indices blended into the exposure score [5].

9. 9

   The Impact of Artificial Intelligence on the Labor Market Webb, Stanford working paper, 2020

   A patent-based exposure measure with a rare historical validation. Webb quantifies the overlap between verb–noun pairs in patent text and in O*NET task descriptions, on the logic that patents describe what a technology does and task statements describe what a worker does. Applied retrospectively to software and robot patents, the measure correctly identifies occupations that lost employment and wages between 1980 and 2010. Applied to AI patents, it finds exposure concentrated in high-skill occupations — the inverse of the robot pattern, and part of why this model's soft spots skew white-collar.

10. 10

    GPTs are GPTs: Labor Market Impact Potential of LLMs Eloundou, Manning, Mishkin & Rock, Science, 2024

    The closest published precedent for this model's formula. Human raters and GPT-4 itself scored each O*NET work activity on whether a large language model could cut its completion time in half — first with the model alone, then allowing complementary software built around it. That two-tier rubric is an early exposure-times-complementarity structure: roughly 2 percent of jobs are heavily exposed to the bare model, rising to about 46 percent once supporting software is counted. The gap between those tiers is the space this model's sliders explore.

11. 11

    AI Exposure Predicts Unemployment Risk Frank, Ahn & Moro, PNAS Nexus, 2025

    The single most important caution in this list. The authors back-tested thirteen exposure scores — including [6], [8], [9], and [10] — against monthly state unemployment-insurance claims across a decade. No individual score reliably predicted realized job loss; an ensemble combining all thirteen performed moderately well. This is the direct empirical justification for two design choices here: using a blended exposure score, and reporting simulated ranges instead of treating any one index as ground truth.

12. 12

    AI Exposure Scores: What They Measure, What They Miss, and What Comes Next Lund, Euyang, Munyikwa & Fadaee, preprint, 2026

    A methodological critique of the whole exposure genre. Every score is calibrated to one AI system's capabilities at one moment, yet the numbers circulate in policy work long after the frontier has moved, shedding their authors' caveats along the way. Different indices can rank the same occupation in opposite directions. Not yet peer-reviewed. Taken here as an operating instruction: date every source, and refresh the slider ranges as the literature moves.

13. 13

    Business Trends and Outlook Survey, AI supplement U.S. Census Bureau, 2025–2026

    The largest recurring federal measurement of firm-level AI use, and the central anchor for the adoption slider. Roughly a fifth of U.S. firms report using AI to produce goods or services; weighted by employment the share rises to about a third, because larger firms adopt at higher rates. The Information and Finance sectors lead by a wide margin — both heavily represented in New York's employment mix.

14. 14

    The State of AI in 2025 McKinsey & Company, 2025

    A global executive survey rather than a statistical sample: most organizations report using AI somewhere in the business, but only about 7 percent say they have scaled it beyond pilots. Self-reported and respondent-selected, so treated as directional. The recurring gap between "using AI" and "running AI at scale" is the main reason the adoption slider's range is as wide as it is.

15. 15

    Automation and New Tasks: How Technology Displaces and Reinstates Labor Acemoglu & Restrepo, Journal of Economic Perspectives, 2019

    The conceptual backbone of the displacement slider. In the task framework, automation exerts a displacement effect (machines take over existing tasks) while new task creation exerts a reinstatement effect (new kinds of work absorb labor). Over the postwar period, reinstatement offset roughly 50 to 80 percent of displacement; the authors trace recent decades' slower employment growth to automation outpacing new-task creation. This model's displacement slider is defined as the net of the two effects — automation minus reinstatement — and its range is anchored to that historical envelope.

16. 16

    Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis arXiv preprint, 2026

    A deterministic task-exposure model projecting displacement from agentic AI — systems that complete multi-step work without supervision — across several regions. Kept as a citable numeric anchor for the displacement slider, with a correction: an earlier version of this bibliography attributed the 50–80 percent reinstatement range to this paper, but the range is borrowed from [15], and the underlying evidence is Acemoglu and Restrepo's. Unreviewed; treat its own projections as provisional.

17. 17

    Robots and Jobs: Evidence from US Labor Markets Acemoglu & Restrepo, Journal of Political Economy, 2020

    The cleanest causal estimate that automation-driven displacement is real. Using variation in industrial-robot penetration across U.S. commuting zones, the authors find each additional robot per thousand workers lowered the local employment-to-population ratio by roughly a third of a percentage point, with wage declines alongside. Cited for the method and the existence proof — not for the coefficient, since industrial robots and generative AI reach different occupations through different channels.

18. 18

    AI and Jobs: Evidence from Online Vacancies Acemoglu, Autor, Hazell & Restrepo, NBER working paper, 2020

    Revealed adoption rather than theoretical exposure: millions of online job postings, 2010–2018. Establishments hiring for AI skills simultaneously reduced their postings for other roles — within-firm substitution — while no effect was yet detectable in economy-wide employment or wages. The gap between the firm-level and aggregate findings is the best available evidence on how long displacement can run before surfacing in official statistics, and it informs this model's refusal to attach a timeline to its estimates.

19. 19

    Displacement or Complementarity? The Labor Market Impact of Generative AI Chen, Srinivasan & Zakerinia, Harvard Business School working paper, 2024

    Builds displacement and complementarity as two separately measured occupation-level indices rather than one blended number — the same decomposition this model's formula uses. Occupations with heterogeneous task mixes score as complementary; occupations with narrow, repetitive task mixes score as displaceable. Not yet peer-reviewed, so its specific coefficients are treated as provisional; the structural finding is what this model borrows.

20. 20

    Anthropic Economic Index Anthropic, dataset release 2026-06-26 (CC-BY)

    The direct source of the augmentation term: anonymized, aggregated conversations from Claude.ai — the consumer chat application, not the business channels through which companies build Claude into their own software — from April–May 2026, mapped to O*NET tasks and SOC occupation families and classified by interaction pattern as automation (the model performs the task) or augmentation (the model assists a person performing it). This model applies each family's augmentation share directly. Known limits: one AI assistant, one distribution channel, a two-month window, and classification at the occupation-family rather than detailed-occupation level. For the deeper problem — the people and tasks that appear in any platform's logs are not a representative sample of any occupation's work — see [22].

21. 21

    Evaluating the Impact of AI on the Labor Market The Budget Lab at Yale, rolling update

    Pairs the usage data in [20] with monthly shifts in the U.S. occupational employment mix. The headline result so far is a null: neither exposure scores nor the automation/augmentation split shows a detectable relationship with realized employment changes. That null grounds two things here — the protection slider's status as a judgment bounded by evidence, and the caveat that this model describes what could happen, not what has.

22. 22

    Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure Yin & Ogut, arXiv working paper, May 2026

    The strongest published critique of the data class behind the augmentation term [20]. The authors show that exposure measures built from AI platform conversation logs largely reflect who a platform's customers are, not which occupations AI affects: computer occupations supply about a third of Claude's consumer conversations but three percent of U.S. employment, the same company's consumer and business data yield opposite-signed employment estimates, and four in ten occupations change exposure quartile between data releases as the product line shifts. Reweighting the logs to match the real workforce shrinks published employment-effect estimates by 42 to 93 percent. Two findings matter directly here. First, the damage concentrates in measures built on conversation volume, which this model never uses — it uses only the assists-versus-replaces ratio within each occupation's own conversations, a quantity volume cannot touch. Second, the bias that remains — the people and tasks that reach a chatbot are not a representative slice of any occupation's work — cannot be corrected with any data that exists, only disclosed and bounded, which the sensitivity check in the methodology's limits section does.

23. 23

    A Stochastic Model of the Long-Range Financial Status of the OASDI Program Social Security Administration, Office of the Chief Actuary

    The strongest government precedent for this project's approach. Social Security's actuaries model a dozen uncertain demographic and economic assumptions — fertility, mortality, wage growth, interest rates — as random processes, run 5,000 simulated 75-year futures, and report the program's finances as percentile bands rather than a single forecast. The structure here is the same, applied to three assumptions instead of twelve.

24. 24

    Prompted to Start: How Generative AI Is Transforming Entrepreneurship Bena, Bian & Giannetti, NBER working paper, 2026

    The first of three sources on a channel this model leaves out by design: the businesses AI helps people start. Using the same task-level exposure logic as the displacement literature, the authors find that industries more exposed to generative AI saw about 20 percent more startups form after ChatGPT's release than less-exposed industries, and that the new entry spread beyond the traditional venture-capital hubs into regions with thin specialized labor markets. Individual startups enter smaller, but the higher entry rate more than offsets their size, producing net gains — roughly 7 percent more employment and 5 percent higher earnings — in new firms in the most exposed industries. Where this model's payroll base reads AI exposure only as a threat to existing jobs, the same exposure here drives job creation through new firms. Not yet peer-reviewed; taken as evidence that the sign of AI's employment effect depends on where you look.

25. 25

    New Business Formation: The AI Advantage Gusto, 2026

    A survey of 1,051 people who started a business in 2025, run by the payroll-software firm Gusto. Nearly 60 percent said they used AI to help launch the business and half said it made starting faster or cheaper, though only 3 percent said they would not have started without it — AI as an accelerator more than a cause. The finding most relevant here cuts against the displacement story from the other side: new businesses using AI in their operations were more likely to plan hiring in 2026 than those that did not, 49 percent against 41 percent. Self-reported and drawn from one platform's customers, so read as directional. Included, with the Stripe and Bena analyses, because the people it counts — founders and the employees they mean to hire — sit almost entirely outside this model's 2024 payroll base.

26. 26

    The Age of the Solopreneur Tedeschi, Rama & Cruickshank, Stripe Economics, 2026

    The clearest read on the part of the workforce this model omits entirely: the self-employed. Combining Census business-formation records, international registration data, and Stripe's own payments records, the authors show that solo businesses are forming faster than employer businesses and, increasingly, reaching real revenue — more than twice as many solo operators cleared a million dollars in 2025 as in 2023. Nonemployer business growth since early 2025 tracks industry AI adoption in the Census Bureau's business survey, which the authors read as AI filling the skill gaps that once forced a founder to hire. The lead author is a former chief economist at the White House Council of Economic Advisers; the piece is an analyst post, not peer-reviewed research. It bounds the same blind spot from the income side: this model counts payroll employees only, and solo work is where some of the displaced — and some of the newly productive — may be landing instead.

27. 27

    How AI May Reshape Career Pathways to Better Jobs Heck, Muro, Methkupally & Siegmund, Brookings, 2026

    The one source here about the connections between occupations rather than the occupations themselves. Brookings maps the career pathways that workers without four-year degrees — about 70 million people it calls STARs, skilled through alternative routes — climb from entry-level "origin" jobs through "gateway" jobs to higher-wage "destination" work, and scores each rung for AI exposure. The exposure lands hardest on the gateway rungs: the clerical, administrative, and customer-service roles, often held by women, that nearly 11 million of these workers rely on to move up. The warning is that AI can break a mobility ladder even where it eliminates no single job on it — a loss this model, which scores each occupation in isolation, is built to miss. Uses an occupation-level exposure measure in the same family this model draws on; a research report, not peer-reviewed.

Version 0.1 — July 22, 2026. Draft release, prepared by Appleseed and Cornell Tech's Urban Tech Hub.