Research · No. 3 · September 2026
Bill Gates's job essay, checked against the data
The door didn't close. It narrowed. Bill Gates says we have crossed the job-market threshold. I read the essay, then spent a week checking it against the employment data he never cites. The diagnosis holds. The shape of the damage is not the one most people picture.
A forecast worth taking apart
I don't follow Bill Gates. I read his August essay because someone put it in front of me, and it stopped me — not because it was alarming, but because it wasn't. It reads like a person describing a system he has looked at carefully, listing what he thinks breaks and admitting which parts he cannot fix.
That combination is rare enough to be worth an afternoon. What follows is his argument, section by section, with the numbers he leaves out. Every figure here links to its source.
Where his essay and his same-day interviews say different things, I mark which is which — that distinction turns out to matter, because the most-quoted numbers live in the interviews, not the essay.
The interesting question was never whether he is right. It is what the data does to the shape of his claim.
Why the usual comparison fails
Most people reach for farm-to-office when they think about this transition. Gates argues the analogy breaks in two places, and both are worth stating plainly.
First, timing. That shift ran across generations. This one, on his estimate, lands on law, customer service, medicine, software and manufacturing inside a decade.
Second, and more important: the farm-to-office move created work that required human cognition. This technology substitutes for the cognition itself. There is no obvious next rung of the same kind.
He adds a detail that explains the speed. The personal computer needed twenty years to change how people worked, because someone had to write the software, drive the price down, and train the workforce. This runs on hardware people already own and speaks plain language. It adapts to us rather than the other way round.
What he claims has already been crossed
In interviews the same day, Gates names five thresholds he believes are behind us: biological capability, cyber capability, psychosocial dependence, job-market disruption, and early signs of losing control6.
His point is not that any one has arrived. It is that the industry spent years promising to slow down as these came into view, and then walked past them without a public conversation.
A note on sourcing. The five-threshold list is not in the essay. The essay names three risks — vanishing work, empowered bad actors, and harm to children and human relationships — and says plainly that no plan exists for the transition5. The sharper framing came in the interviews. Most coverage merges the two.
The word threshold is not rhetorical. The leading labs each publish a safety framework that names capability levels in advance — assistance with biological weapons, autonomous cyber operations, self-propagation — and commits to pausing at them. That is the promise Gates is measuring against.
Where the effect actually shows up
Here is where his essay goes quiet. He writes that employment fell among young workers in vulnerable occupations while holding steady for their older colleagues, and gives no figure at all. The number exists, and it is more interesting than the sentence.
Look at the whole market first, and there is almost nothing to see. Payroll data covering millions of workers, grouped by how exposed the occupation is, shows every group moving within a percentage point of every other3. This is why several research teams report finding no AI footprint in the labour market at all.
Stanford Digital Economy Lab, balanced ADP payroll panel3.
Now add one variable. Same data, same month, same occupations — split by age.
Same panel, same month3. The gap between the youngest and the 41–49 group is 4.9 percentage points — inside the same occupations.
The effect is not between occupations. It is inside them, sorted by age. Over the longer window the split is starker still: since late 2022, employment for 22-to-25-year-olds in the two most exposed groups fell 11 per cent, while the same age group in the three least exposed rose 10 per cent3.
−11%
In the two most exposed groups
+10%
Same age, in the three least exposed
The dashed outline is the level at the end of 2022; both openings are measured from it3.
Nobody is being fired. Fewer people are being let in.
That distinction is the finding. Separation rates for exposed young workers did not rise — hiring slowed.
The door did not close behind the people already inside; it narrowed for the people trying to enter. Gates describes exactly this when he writes about entry-level roles, and it is the part of his essay that survives contact with the data best.
What the researchers themselves flag. The effect halves once you control for education, part of the decline predates generative AI, and the gap is smaller in national surveys than in this payroll panel.
The panel is not representative of the US labour market, and the paper is not peer-reviewed. Its authors call the results descriptive, not causal3.
Check your own job
Gates names occupations by observation: sales and support, software, paralegal work, then loan assessment, data analysis, patient triage. There is now an official measure to check him against.
In August, the Bureau of Labor Statistics published its first classification of occupations by AI exposure, built from five independent measures — three theoretical, two from observed real-world use — alongside employment projections to 20351.
Enter a job below. Everything returned comes from the sources listed at the end, and every number is footnoted to the one it came from.
Interactive · 825 occupations
The occupation
And you in it
This is not a prediction about you. It shows where your occupation sits among others, and what happened to employment for people your age in occupations with similar exposure. Exposure means a model can be applied to the tasks — the Bureau states explicitly that it implies neither job loss nor automation nor wage effects1.
One fifth of the market, thoroughly
Put all 825 occupations in one table and the picture is calmer than the discourse. Sort by any column, filter by name, and read the exposure category off the marker beside it.
How to read the table
Bureau of Labor Statistics, AI exposure categories and employment projections 2025–351.
The share of employment in each category, across all 831 occupations the Bureau projects:
| Exposure | Occupations | Employed, 2025 | Share | Projected 2025–35 |
|---|---|---|---|---|
| Very high | 206 | 53.2m | 31.3% | +2.0% |
| High | 206 | 42.1m | 24.8% | +3.1% |
| Moderate | 206 | 44.7m | 26.2% | +6.2% |
| Low | 213 | 30.3m | 17.8% | +2.5% |
Bureau of Labor Statistics, AI exposure categories and employment projections 2025–351.
The most exposed third of the market grows slowest — and still grows. That is the official projection, not an opinion. And the spread inside the top category is wider than the spread between categories:
| Very high exposure | Employed | Projected to 2035 |
|---|---|---|
| Office clerks, general | 2.60m | −6.0% |
| Secretaries and administrative assistants | 1.89m | −6.0% |
| Bookkeeping and auditing clerks | 1.53m | −5.6% |
| Customer service representatives | 2.67m | −5.3% |
| Computer and information systems managers | 0.69m | +15.8% |
| Logisticians | 0.26m | +17.6% |
| Data scientists | 0.28m | +34.6% |
Same exposure category, opposite trajectories. Gates makes this point once, in a clause most summaries dropped: in software, falling costs will generate new demand, so the net loss there will be smaller than elsewhere. The projections agree with him.
The shift arrives when the checking stops
The most underrated line in the essay is not about any occupation. Gates writes that the biggest shift for workers comes when the technology is nearly error-free — because that is the moment it can run without a human reviewing it, and every economic incentive points at letting it.
There is a benchmark that puts numbers on this.
OpenAI's GDPval evaluates models against experienced professionals on 1,320 real work products across 44 occupations, averaging about seven hours of work each7. Its headline is that a frontier model matches or beats the human expert on roughly half the tasks. Buried in the same paper is a table almost nobody quotes.
The cost of checking
Model advantage over the human expert
Everything between those two numbers is a question of how much checking you keep7.
Measured naively, the model is around 90 times faster and several hundred times cheaper. Once expert verification is priced in, the advantage collapses to roughly 1.1–1.4× on time and 1.2–1.6× on cost7.
So the entire economic case sits in the review step. While a person still checks, the gain is marginal. The whole prize appears the moment checking is removed — which is precisely why it will be removed, and why Gates is right to call this the real threshold rather than any particular capability.
The part the essay leaves out
Gates frames the whole piece around fairness: this technology becomes either the greatest equaliser or the worst source of injustice. Then he never asks who is actually in the exposed jobs.
I matched the exposure classification against the employment survey. Weighted by how many people work in them, occupations in the highest exposure category are 57.5 per cent female.
In the lowest category they are 36.9 per cent female14. A twenty-point gap, and the reason is not subtle: clerical, administrative and support work is both the most exposed and, nearly everywhere, the most female.
The International Labour Organization finds the same thing globally, with a method built on international classifications rather than American ones: in the highest exposure tier sit 4.7 per cent of women's employment against 2.4 per cent of men's, and in high-income countries 9.6 against 3.58.
That same ILO work is also a useful check on scale. It puts global exposure at 24 per cent of employment, where the IMF puts it near 40 per cent9. Two credible institutions, one phenomenon, a factor of nearly two between them. Anyone quoting a single global figure with confidence has picked one and hidden the other.
Reserves, tokens, and four unanswered questions
Human Reserved
Gates proposes setting certain work aside for people — not because machines couldn't do it, but because society decides the loss would be too great.
He borrows the language of nature reserves. The personal origin is his father's final years, and the caregivers who understood the man when he could no longer make himself understood.
He offers two grounds. One economic: a 55-year-old who has spent a career in construction cannot simply be redirected into elder care. One human: a robot delivering a terminal diagnosis is technically possible and should not happen.
Then he lists, in his own text, the four questions he cannot answer. Who decides what gets reserved. By what criteria. How you stop companies from cheating. What happens to trade when one country lets the robots build it and another doesn't.
Worth noting against the data: elder care, his central example, is already short of workers. Reserving an occupation that cannot fill its vacancies protects it from a pressure it isn't yet under. He anticipates part of this — he expects Japan, with too few young people, to welcome the caregiving robot.
Taxing tokens and robots
The asymmetry he wants to remove is real and rarely stated this plainly. Hire a person and the employer pays payroll tax. Buy a machine and the company writes it off. The tax code currently subsidises the substitution.
His fix is a targeted levy on AI tokens and robots, funding retraining and a stronger safety net, designed to spare the uses that make medicine and education cheaper.
He concedes it is economically inefficient and considers that an acceptable price. No rate appears in the essay; the widely quoted figure comes from an interview and is floated as an option, not a calculation6.
The strongest objection is mechanical rather than ideological: a token tax measures computation, not displacement. Token prices have fallen by orders of magnitude, so the base shrinks as the problem grows, and inference migrates to local devices and other jurisdictions.
What honest people still argue about
It would be easy to end by declaring the case proved. It isn't, and the disagreement is worth more than a verdict.
| Finding | Source |
|---|---|
| Software employment growth slowed by ~3 points annually after late 2022; roughly 500,000 jobs not created | Federal Reserve10 |
| No connection between AI use and employment changes, through July 2026 | Budget Lab at Yale11 |
| 2.8–3% of hours saved, precisely zero effect on earnings or hours | Danish administrative data12 |
| Decline in exposed postings began March 2022 — before ChatGPT — tracking the rate cycle | 238m job postings13 |
| Unemployment rose more in the least exposed quintile than the most | Economic Innovation Group14 |
Three things every one of these teams agrees on, and they are the safest conclusions available today: there is no mass displacement at the macro level; the mechanism is slowed hiring rather than layoffs; and something is happening to young workers in exposed occupations.
One more limit belongs here. You cannot yet test any 2026 forecast against actual employment by occupation: the official occupational series runs only through May 2025, and the next release covering this period is not due until 2027. Anyone showing you that check today is showing you noise.
What the numbers change
Gates's diagnosis survives. The mechanism is right, the sectors are right, and the emphasis on entry-level work is the part the data supports most directly.
What the numbers change is the shape. This is not a wave rolling across occupations, taking them one by one. It is a filter operating inside occupations, sorting by how long you have been there. The people it selects against are the ones with the least to fall back on and the most years ahead of them.
He ends his essay with four questions and no answers.
The one I would add is narrower and more answerable than any of them: if the mechanism is a hiring slowdown at the entry level, then the policy target is not which professions to fence off, but who gets the first rung. Reserves protect the people already standing on the ladder.
Sources
- US Bureau of Labor Statistics. AI exposure categories and employment projections, 2025–35, published 27 August 2026. Categories built from five independent measures. bls.gov/emp/publications/ai-exposure-categories.htm
- Anthropic Economic Index, labor market impacts dataset — measured share of an occupation's tasks brought to the model. huggingface.co/datasets/Anthropic/EconomicIndex
- Brynjolfsson, Chandar, Chen. Canaries in the Coal Mine? Stanford Digital Economy Lab, August 2026 revision; ADP payroll panel through July 2026. digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine
- US Bureau of Labor Statistics, Current Population Survey, table 11: employed persons by detailed occupation and sex, 2025 annual averages. bls.gov/cps/cpsaat11.htm
- Bill Gates. The turbulent AI era is here. The choices we make now are critical. GatesNotes, 26 August 2026. gatesnotes.com
- MIT Technology Review interview with Bill Gates, 26 August 2026 — source of the five thresholds and the quoted rates. technologyreview.com
- GDPval, OpenAI — 1,320 real work products across 44 occupations; see the cost table for verification-adjusted figures. arxiv.org/abs/2510.04374
- International Labour Organization. Generative AI and Jobs: A Refined Global Index of Occupational Exposure, Working Paper 140, May 2025. ilo.org
- International Monetary Fund. Gen-AI: Artificial Intelligence and the Future of Work, SDN/2024/001, January 2024. imf.org
- Crane and Soto, Federal Reserve Board, FEDS 2026-018, March 2026.
- Budget Lab at Yale, Tracking the impact of AI on the labor market, updated 19 August 2026. budgetlab.yale.edu
- Humlum and Vestergaard, NBER Working Paper 33777, March 2026 revision.
- Iscenko and Curto Millet, January 2026 — 238 million job postings.
- Eckhardt and Goldschlag, Economic Innovation Group, AI and Jobs, August 2025. eig.org
Merged dataset behind the interactives: 825 occupations with employment, projections, median wage, official exposure category, measured task exposure and gender composition. Assembled September 2026 from sources 1, 2, 4 and the Occupational Employment and Wage Statistics release of May 2025.