Will AI Create More Jobs Than It Destroys?

Neither — and that is the finding, not a dodge.

As of August 2026 there is no measurable net job destruction from AI in any major economy's data, and no measurable net creation attributable to it either. Four independent bodies of evidence — US administrative payroll records,¹ Federal Reserve job-postings research,³ Danish administrative records,⁴ and US population survey data⁸ — converge on the same conclusion, and the AI industry's own analysis agrees with them.² Aggregate employment has not moved in a way anyone can pin on AI.

What has moved is composition. Firms are hiring proportionally fewer junior people and more senior ones,¹²⁶ and the measured adjustment is landing on the hiring margin rather than on pay.¹⁴ Whether AI caused that shift is, on the published record, not established — the researchers who found the sharpest version of it say so in their own abstract.¹³ The question in the title is a forecast, and measurement cannot answer it yet. What measurement can do is kill the two loudest claims in the room: the bloodbath has not arrived, and neither has the boom. [ ok ]

The displacement did not show up in the data.

The strongest available evidence is administrative rather than survey-based, and it is unusually consistent. Stanford's analysis of ADP payroll records covering millions of US workers through June 2026 opens with a flat negative: “We find no evidence of widespread, economy-wide job displacement.”¹

The Federal Reserve Bank of New York, working from job postings rather than payroll, reached the same place: the data provide “little indication of a distinct AI-driven decline in labor demand.”³ The Budget Lab at Yale, working from population survey data, described a picture that “largely reflects stability, not major disruption at an economy-wide level.”⁸ A study linking adoption surveys to Danish administrative records estimated null effects on earnings and hours tight enough to rule out changes above 1% — and concluded that the findings “challenge narratives of imminent disruption from Generative AI.”⁴

The most interesting agreement comes from the least disinterested party. Anthropic — a company whose commercial case rests on AI displacing work — published its own labour-market measure and reported that it found “no systematic increase in unemployment for highly exposed workers since late 2022”, alongside “suggestive evidence that hiring of younger workers has slowed in exposed occupations.”²

Bias note: source [2] is an AI developer researching its own market and is not independent. It is included because it converges with four independent sources rather than against them, and because its finding runs contrary to the vendor's commercial interest in appearing transformative.

A four-point gap, not a collapse.

Occupations most exposed to AI did underperform, but they still grew. In the ADP sample, average employment rose by about 6% between November 2022 and June 2026, while employment in the most-exposed quintile grew by about 4%.¹ That is a slowdown, not a contraction, and the researchers argue the absence is itself the result.

“This aggregate stability is itself informative: claims of economy-wide AI-driven job losses are not visible in payroll data through June 2026.”

— Brynjolfsson, Chandar & Chen, Stanford Digital Economy Lab [1]

Job-postings data tells a two-phase version of the same story. Indeed's Hiring Lab found that across 2022–2026 “the more exposed to AI an occupation is, the more it declined” — but that between 2025 and 2026 “the more exposed to AI an occupation is, on average, the more it rebounded.”⁵ Whatever the mechanism, the sign of the relationship has not been stable long enough to extrapolate from.

The market is tilting toward experience.

Two independent data origins agree on one compositional shift, and a third — non-independent — points the same way. Indeed's postings data shows that as of May 2026 “senior-level job postings were up 14.7% year-over-year” while “entry-level job postings have been trending downwards since peaking in 2022, declining 7.5% year-over-year.”⁶ Stanford's payroll analysis finds a comparable divergence by age and exposure, and specifies its mechanism: it “operates primarily through reduced hiring of young workers rather than increased separations.”¹ Anthropic's independent measure reports the same directional signal for younger workers in exposed occupations.²

The scale matters as much as the direction. Even in the study that found the sharpest effect, total employment for 22-to-25-year-olds came out roughly flat over the period — a 1.9% decline — because growth in less-exposed occupations absorbed most of the shortfall.¹ Reallocation is doing real work here. It is not doing all of it.

Every research team studying this says the same thing.

The temptation is to read section 03 as AI displacing juniors. The researchers who produced that finding explicitly decline to draw it.

“We caution that this work does not estimate a causal impact of AI: these are descriptive facts, and ongoing work is needed to determine how much they are caused by the spread of generative AI rather than merely correlated with it.”

— Brynjolfsson, Chandar & Chen [1]

Their own robustness checks cut against the simple reading: the patterns “attenuate when controlling for education, show some divergent trends predating generative AI, and are more pronounced in the ADP analysis sample than in national survey benchmarks.”¹

The New York Fed goes further on the timing. Its postings analysis found the divergence between high- and low-exposure occupations “began before 2022 and does not show a clear additional break in trajectory after 2022”, that the slowdown is “not concentrated specifically in entry-level highly exposed jobs”, and concluded that “while AI may be contributing to recent labor market developments, it is not the main driver of the slowdown in hiring.”³ Indeed's analysts attach the same caution — “Correlation does not imply causation”⁵ — and note that competing explanations including remote work and post-2022 over-hiring “may be at least partially at play.”⁶

In hiring, not in pay.

Two independent datasets agree on which margin is moving. Stanford's sixth documented fact is that “adjustment is occurring through employment rather than base compensation.”¹ The Danish study reached the same conclusion from the other end, finding “precise null effects at both the individual and workplace levels, with confidence intervals ruling out effects above 1%” on earnings and recorded hours.⁴

This is the practical shape of AI's labour-market footprint so far: people already holding jobs are not seeing their pay cut or their hours reduced in any measurable way.¹⁴ The pressure, such as it is, sits on the door rather than on the people already through it.¹⁶

The split that predicts the direction.

The single most useful finding in this literature is not a number — it is a fork. Employment outcomes diverge sharply depending on whether AI is used to replace a task or to extend the person doing it. Stanford's fifth fact states it directly: “Declines are concentrated in occupations where AI usage primarily substitutes for human tasks; where usage primarily complements workers, employment is flat or rising, especially for experienced workers.”¹

That is not a novel property of generative AI. Peer-reviewed work spanning eight decades of US employment establishes the same asymmetry causally: “although flows of augmentation and automation innovations are positively correlated across occupations, the former boosts occupational labor demand while the latter depresses it.”⁷ Augmenting technologies generate new work; automating ones suppress its emergence.⁷

The historical record carries a sting that the optimistic version of this argument usually omits. The same peer-reviewed analysis concludes that “the demand-eroding effects of automation innovations have intensified in the last four decades while the demand-increasing effects of augmentation innovations have not.”⁷ The historical base rate is not a guarantee that new work arrives on time or in the same places. On the measured trend, the erosive half has been getting stronger and the creative half has not.

Real, measurable, and weighted to the top.

The creation side is visible in the data, if smaller than the headlines suggest. The Danish study found that workplace AI initiatives “boost adoption, narrow gender gaps, and create new tasks”, and attributed the muted net effect partly to “the emergence of integration and oversight tasks.”⁴ Work is not simply vanishing; some of it is being converted into the work of running the machine.

Postings data shows the same thing at market scale. Since February 2025, US software development postings on Indeed have risen almost 15% while postings overall declined by 7%.⁵ Indeed's analysts summarise it as a reversal: “The relationship between AI exposure and job postings appears to be flipping, from job destruction to job creation.”⁵

The distribution is the catch. Of that software-postings increase between May 2025 and May 2026, 71% came from senior roles and 37% from jobs that mention AI.⁵ New work is being created in the occupations most exposed to AI — and it is landing disproportionately on people who already have experience,⁵⁶ which is the same compositional tilt documented in section 03.

What this brief deliberately does not claim.

This brief was built verified-only: every claim above is corroborated by at least two independent sources, or is explicitly attributed to the single study that produced it. A substantial amount of the most-quoted material on this topic did not clear that bar and was excluded.

The headline question is unanswerable by measurement. “More jobs than it destroys” is a claim about a counterfactual future. No dataset in existence settles it. Anyone offering a confident number is forecasting, not measuring.

Excluded — contested magnitude. The most widely circulated figure in this area is Stanford's estimate of how far young workers in AI-exposed occupations have fallen behind their less-exposed peers. The finding's direction is corroborated and appears in section 03; its specific magnitude rests on one dataset, is disputed on attribution by the New York Fed,³ and is described by its own authors as descriptive rather than causal.¹ It is not asserted here as a settled quantity.

Excluded — forecasts. The World Economic Forum's widely-quoted net-jobs projection, MIT productivity estimates for AI's macroeconomic effect, and public statements by AI-company executives are all forecasts or single-origin claims. None are measurements, and none appear in the body.

Excluded — could not be fetched. Frequently-cited exposure statistics from the IMF, the OECD, the ILO and S&P Global were excluded because their primary documents could not be retrieved for direct quotation. Search snippets existed for all of them. Under this method a snippet is a lead, not a citation.

Known limits of the evidence that is here. The Danish null results cover 11 occupations in one small, high-income economy.⁴ The Stanford payroll findings are more pronounced in the ADP sample than in national survey benchmarks, by the authors' own admission.¹ Source [8] reports the Yale Budget Lab's findings second-hand; the primary publication could not be retrieved directly. And all of it measures roughly three and a half years of a technology whose capabilities are still moving. Absence of a measured effect in 2026 is not evidence of absence in 2030.

The register.

1. Brynjolfsson, Chandar & Chen — Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (August 2026) Primary · independent · ADP administrative payroll data through June 2026
2. Anthropic — Labor market impacts of AI: A new measure and early evidence (5 March 2026) Primary · not independent · AI developer researching its own market; converges with [1][3][4][8]
3. Federal Reserve Bank of New York, Liberty Street Economics — Do Job Postings Show Early Labor-Market Effects of AI? (14 May 2026) Primary · independent · central bank research; disputes AI attribution in [1]
4. Humlum & Vestergaard — Large Language Models, Small Labor Market Effects (15 July 2025) Primary · independent · Danish administrative records, 11 occupations
5. Indeed Hiring Lab — AI and Job Postings: From Destruction to Creation? (8 July 2026) Primary · independent · job-postings data
6. Indeed Hiring Lab — The Labor Market Is Tilting Toward Seniority (23 July 2026) Primary · same origin as [5] · counted as one source of independence
7. Autor, Chin, Salomons & Seegmiller — New Frontiers: The Origins and Content of New Work, 1940–2018 Primary · peer-reviewed (QJE, 2024) · independent · historical base rate
8. Fortune — reporting The Budget Lab at Yale on AI and the labour market (2 February 2026) Secondary · derives from The Budget Lab at Yale · primary publication not retrievable
9. World Economic Forum — Future of Jobs Report 2025 (January 2025) Secondary · employer-expectation survey · excluded from the body; cited only for the over-association guard