Human value added is earnings divided by a 55% labor share. AI and robot output is the wage bill of displaced work times a productivity multiplier that rises with adoption.
Sources: US Bureau of Labor Statistics (BLS) Occupational Employment and Wage Statistics (OEWS), May 2025, annual mean wages. * Pay estimated from the published median. Self-employed count from the BLS Current Population Survey (CPS), June 2026. The sex work baselines (0.6 million in person, 0.4 million online creators) split a midpoint of published estimates that range from 70,000 to 2 million people, so treat those rows as highly uncertain. Online creator pay is the OnlyFans average payout of about $2,520 per active creator per year, which hides a very skewed spread. All dollars are real 2025 dollars.
McKinsey Global Institute (MGI) estimates today's AI agents could in theory automate about 57% of US work hours, and robots another 13%. MGI stresses that this is a ceiling on tasks, not a forecast of lost jobs. The "Automatable tasks that become lost jobs" slider is how this model turns that ceiling into job counts.
Goldman Sachs expects AI to displace 6 to 7% of US workers over the adoption period, with unemployment rising about half a point before new work absorbs people. Morgan Stanley's humanoid model sees 40,000 US robots in 2030, 8 million in 2040 and 63 million by 2050. The robot curve here follows that shape.
What is happening now? Stanford's payroll study finds workers aged 22 to 25 in the most AI-exposed jobs are about 19% below where they would otherwise be, while older workers in the same jobs hold steady. The Yale Budget Lab finds no economy-wide disruption yet. Both can be true. Entry-level knowledge work bends first.
Which pattern does the model show? Knowledge jobs built on codified, repeatable tasks fall early. Physical jobs keep growing until robots get cheap in the late 2030s, then fall fast. Hands-on care keeps growing because an aging country needs more of it.
The brief for this model included a claim that most men and women will become entrepreneurs and a significant share of women will become sex workers. Is there evidence for it? Choose "The claim, tested" and watch what happens to income.
There is real evidence for a shift. Census business applications hit a record 3.23 million in the first half of 2026, and one-person filings are up more than 20% since early 2025, with almost no growth in firms that plan to hire. But customers are finite.
No credible study forecasts a large share of women moving into sex work, but two forces pull in opposite directions. Displacement pushes some women in: a 2025 study found a recession in a female-heavy British industry raised local sex work venues by about 20%. AI pushes the online side out: AI-generated creators already earn about 15% of revenue on Fanvue, and they work around the clock at almost no cost. So the model splits the two. Online creator work collapses, and some of those creators move to in-person work to keep their income. In-person work rises.
So which story fits better? A world where displaced people crowd into self-employment and informal work, or one where new kinds of jobs absorb them? The answer depends mostly on one assumption: how many displaced workers find new jobs. That is a policy and business choice, not a law of nature.
Each occupation group gets an AI task share, a robot task share, a baseline demand trend, and a women’s share from the selected country’s baseline data where available. Automation and growth paths are model assumptions informed by published research. Jobs lost each year equal the realization rate times (AI share × AI adoption + robot share × robot adoption). Surviving workers in AI-exposed roles gain a wage premium; heavy displacement pushes wages down. "Not earning" counts displaced people not absorbed by new jobs, new businesses or informal-work pathways. Country-specific limits and source definitions are stated beside the starting data.
BLS OEWS May 2025, national table BLS: heavy truck drivers, May 2025 McKinsey Global Institute: Agents, robots, and us (Nov 2025) Goldman Sachs: How will AI affect the global workforce? Morgan Stanley: humanoid robot forecast Stanford Digital Economy Lab: Canaries in the Coal Mine, Aug 2026 update (summary) Yale Budget Lab: Tracking the impact of AI on the labor market Axios: Nasdaq sees AI behind solo founder boom Self-employment statistics 2026 (BLS CPS Table A-9) ILO and NASK: AI and women's jobsThis is a scenario model, not a forecast. Its starting point is measured data. Its growth and decline rates are assumptions built from published research, and they carry very different levels of confidence. Each layer is shown separately below so anyone can challenge it.
Jobs and average pay for the US occupations come from BLS OEWS, May 2025. This measured baseline is separate from the model’s uncertain projections.
Which jobs face pressure first is consistent across independent studies. Codified office work is most exposed to AI agents (McKinsey, Stanford, ILO). Routine physical work is exposed later, through robots (McKinsey, Morgan Stanley). Hands-on care is least exposed and has rising demand (McKinsey).
The adoption curves, the share of automatable tasks that become lost jobs, each occupation's exact task shares, and how displaced people move into new jobs, businesses or sex work are judgment calls. Small changes here move the 2056 results a lot, which is why they are sliders. The sex work baseline is the least certain number in the model.
Two pieces are anchored to published figures. The robot S-curve follows the shape of Morgan Stanley's US path of about 40,000 humanoids in 2030, 8 million in 2040 and 63 million in 2050: slow before 2035, steep in the 2040s. It runs ahead of that path in 2040, at about 27% of full adoption versus Morgan Stanley's 13%, so even the Moderate case is ahead of Morgan Stanley on robots in the late 2030s. The default Fast adoption case reaches 50% robot adoption in 2040. The realization slider turns McKinsey's 57% "technically automatable hours" ceiling into job losses. Task concentration then raises the loss rate for jobs built around one task, such as driving or checkout, toward 100%, because automating that one task leaves little job behind. In the Fast adoption case about 41% of all wage jobs are displaced by 2056, and 45% of those people find new work. Fast adoption is deliberately at the high end of the research, built for keynotes; Moderate is the central estimate.
Pareto check: the top 20 occupations hold about 50.1 million of 155.5 million wage jobs, or 32%. That is not an 80/20 split, because 20 occupations are only 2.4% of the roughly 830 that BLS tracks. The 80/20 rule would compare the top 20% of occupations, about 166 of them, with all jobs.
The same model equations and scenario assumptions are restarted from BLS May 2022 employment counts, with adoption reset to zero in 2022 and the original adoption timing retained. No parameters are fitted to the observed changes. Observed values are BLS OEWS estimates; rideshare is excluded because it is self-employed and outside OEWS. The assembler row follows the comparable “Miscellaneous Assemblers and Fabricators” SOC category.
“Error” is model change minus observed change, in percentage points. Large mismatches are flagged at an absolute error of 10 points or a wrong direction when the observed change is at least 5%. Values are employment counts, not unique people. May 2022 and May 2025 are survey estimates, not censuses; this two-endpoint check does not show when changes happened. BLS May 2025 uses six panels collected from November 2022 through May 2025, so intermediate annual estimates overlap.
Repeat the automated scenario checks with pnpm --filter @workspace/jobshift test:backtest.
AI task share and robot task share are the model's estimates of how much of each job's work agents and robots can do once adoption is complete. Demand trend is the yearly change in demand before any automation.
The 2022–2025 check is an initial hindcast, not proof that the model predicts future employment accurately. The BLS May 2025 estimates use six survey panels from November 2022 through May 2025; annual releases in between share panels, so they are not independent. This two-endpoint comparison does not identify when changes happened. The large mismatches below are visible, not calibrated away. No occupation-level May 2026 OEWS release is available yet; this comparison cannot score 2026.
Next steps for stronger evidence are to replace judgment-based task shares and demand trends with published occupation-level exposure and BLS employment projections, and to report ranges across the assumption sliders rather than single scenario paths.