The AI Jobs Report: What Stanford Data Says About Hiring in 2026

September 18, 2026

The Stanford Institute for Human-Centered AI released its ninth annual AI Index Report on April 14, 2026, and for the first time the most cited labor-market numbers in the document came from outside Stanford itself: a revised August 2026 working paper from the Stanford Digital Economy Lab by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, using ADP payroll data covering millions of US workers through June 2026. Together with a Goldman Sachs Research note that put a 16,000-jobs-per-month number on AI substitution, the data tells a story the marketing decks won’t: AI hiring is up, AI salaries are wider than the rest of the labor market, and the people getting hit are precisely the people who used to enter those careers.

If you want to know who is getting paid in 2026, the AI Index says you should stop reading job boards and start reading payroll data. The numbers are sharper, the methodology is cleaner, and the conclusion is uncomfortable: the industry is hiring more experienced AI workers than ever, paying them more than ever, and quietly closing the door on the entry-level rung that produced them.

This is what the Stanford data, plus Goldman Sachs and a handful of other primary sources, actually says about who is getting hired, who is getting paid, and who is being squeezed out of the AI economy.

1. The Stanford data set the new floor

The headline number from the revised Canaries in the Coal Mine? paper (Stanford Digital Economy Lab, August 12, 2026): employment among workers aged 22 to 25 in highly AI-exposed occupations now stands 19% below where it would have been had it kept pace with similarly aged workers in less-exposed occupations. The same gap was 15% in July 2025, when the authors first documented it. The number did not appear in a vacuum. The first paper, dated October 2025, used ADP data through September 2025; the August 2026 revision extends through June 2026 and sharpens the conclusion. Brynjolfsson’s team describes the finding as “canaries in the coal mine” rather than causal proof: the data is descriptive, not experimental, but it is now drawn from a payroll panel covering tens of millions of workers at over 25,000 firms using ADP for payroll services (Canaries Dashboard, live since July 2026).

Three structural findings anchor the rest of the report.

No widespread displacement. The authors are explicit: there is no evidence of economy-wide job destruction from AI. Aggregate employment growth has slowed at the margin, but the unemployment rate has not broken out of its post-pandemic range.
The adjustment is hiring, not firing. The gap between young workers in AI-exposed jobs and their less-exposed peers is driven by reduced hiring of new graduates, not by mass separations of those already employed. The industry is not laying off 22-year-olds; it is failing to hire them.
Concentration in substitution-heavy work. The gap is concentrated in occupations where AI primarily substitutes for human tasks: software development, junior analyst roles, basic customer service, and entry-level marketing and writing. Where AI primarily complements human workers (experienced nurses, skilled trades, mid-career engineers doing design work), employment is flat or rising.

The Stanford Digital Economy Lab now publishes a live Canaries Dashboard updated monthly. This is the only place to watch the 22-to-25 gap in close to real time.

2. Hiring is up. Just not for new graduates.

The Stanford HAI 2026 AI Index reports that 2.5% of all US job postings now mention AI skills, up 55% from 2024 and nearly 300% over the decade, according to Lightcast’s analysis of billions of postings. The aggregate AI hiring market is growing. What is missing is the bottom of the pyramid.

Three data points frame the divergence.

Agentic AI skills grew from 0.06% of job postings in 2024 to 0.23% in 2025 — a more than 280% increase in a single year, representing roughly 90,000 US job postings in 2025 alone, per the Stanford HAI / Lightcast analysis.
Python, the most in-demand specialized skill, appeared in 258,674 postings — up 391% from the 2013 to 2015 baseline and nearly 30% from 2024.
Skills tied to deployment, like Amazon Web Services, scalability, and workflow management, are now the fastest-growing segments in AI job postings — meaning the industry is hiring for production execution, not for experimentation.

The AI Index conclusion is unambiguous: “AI is no longer just a frontier technology.” That is the same conclusion a junior developer reading a 1,000-application job posting cycle reaches, only from a different direction. The postings are real. The pipeline is not.

3. Goldman Sachs puts a number on it

Goldman Sachs Research, in an April-May 2026 note from analyst Elsie Peng and follow-up exchanges with Joseph Briggs, estimated that AI has slowed US monthly payroll growth by roughly 16,000 jobs over the prior year. The number is the net of two flows. AI substitution eliminates about 25,000 jobs per month in highly exposed occupations; AI augmentation adds back about 9,000 through new roles such as prompt engineers, AI trainers, and compliance-quality reviewers. The ratio — roughly 2.8 eliminations for every 1 creation — is what makes the aggregate figure a net negative for total jobs, even as it creates genuinely new categories of work.

The 16,000 figure has to be read carefully. Goldman itself notes that the estimate likely overstates AI’s total labor-market effect because it does not fully capture hiring for data-center construction, AI-enabled output growth, and productivity-led demand that may not yet appear in payroll data. The model is also narrow: concentrated in tech, management consulting, and graphic design, where AI tools have already been deployed at scale.

Across more than 800 occupations, Goldman’s regression analysis finds that the negative effect on job creation falls disproportionately on younger, less-experienced workers. A one-standard-deviation increase in substitution exposure widens the entry-level-to-experienced wage gap by roughly 3.3 percentage points. Call-center employment is now 39% below trend in the US, 33% below trend in Canada, and 27% below trend in Germany, per Goldman’s analysis. The pattern is not unique to AI-exposed work. It is concentrated in entry-level positions inside AI-exposed work.

In an August 2026 Goldman Sachs Exchanges discussion, MIT’s Daron Acemoglu put the longer-run ceiling lower than Goldman’s headline: “limited net job losses within the next five years” in the 2-to-4% range, with the most vulnerable tasks being cognitive routine work such as customer service and back-office administration — a combined US cohort of roughly 8 to 9 million workers. Goldman’s base case is bigger: around 9% of all US workers reallocated during the AI transition, roughly 15 million workers, spread over a decade so the unemployment-rate impact in any single year stays under one percentage point.

4. Productivity gains are real. They are also uneven.

The single most cited finding in the AI productivity literature is from Brynjolfsson, Li, and Raymond’s 2023 study of 5,179 customer-support agents at a Fortune 500 software firm: a 14% average increase in issues resolved per hour, and a 34% gain for novices and low-skilled workers. AI compressed the within-occupation skill premium. That is a real and well-replicated finding; it also happens to explain why call-center employment is now 39% below trend. Productivity gains and entry-level collapse are the same event viewed from two angles.

For software development, the best current evidence comes from Demirer, Musolff, and Yang’s NBER working paper (May 2026, revised September 2026), which used telemetry from more than 500,000 GitHub developers. The headline: AI coding tools raise commit activity by 30% (autocomplete), 180% (adding interactive agents), and 240% (adding autonomous agents). Those gains attenuate sharply across the production hierarchy: the 240% commit-level gain falls to 30% in actual software releases. The bottleneck is no longer writing code. It is reviewing, integrating, testing, and shipping code. The elasticity of substitution between AI and human effort is 0.23, which signals strong complementarity.

Stanford HAI’s Chapter 4 synthesis reports productivity gains of 14% to 15% in customer support, 26% in software development, and 50% in marketing output, mirroring the underlying primary studies. The aggregate numbers are softer. Acemoglu’s conservative macro estimate is +0.07 percentage points per year of total-factor productivity over the next decade. Optimistic scenarios from Aghion-Bunel and the OECD high-exposure case reach +1.3 percentage points per year. The 18x gap between the two bounds is the honest representation of the macro uncertainty.

5. Salaries at the frontier are extreme

For the cohort that does get hired, the pay is widening rather than compressing. The data points worth bookmarking.

OpenAI discloses base pay through H-1B filings. Member of Technical Staff (Research Scientist): $245,000 to $685,000 base. Member of Technical Staff (AI Systems Engineer): $245,000 to $460,000 base. Member of Intelligence & Investigations Staff: $320,000 to $382,500 base. The Wall Street Journal reports OpenAI’s 4,000+ employees average $1.5 million in stock-based compensation, per Business Insider’s analysis.
Anthropic, per Levels.fyi data last updated August 31, 2026: median total comp $402,350. Software Engineer: $367K base. Senior SWE: $591K. Lead SWE: $779K. Staff SWE: $1.25 million total comp. The staff-level band alone exceeds what most software engineers earn at the director level at non-AI companies.
Meta’s superintelligence operation reportedly paid Andrew Tulloch a compensation package worth as much as $1.5 billion when he joined from Thinking Machines Lab, per coverage in Ynet News. He left less than a year later. The Frontier-AI lab pay market has not stabilized.

For the rest of the market — the AI engineer not at a frontier lab — KORE1’s 2026 salary guide synthesizes the four major sources (Glassdoor, Built In, ZipRecruiter, Levels.fyi) into a single statement: mid-level AI engineers land $155,000 to $200,000 base, senior AI engineers land $220,000 to $300,000 base, and total comp routinely clears $300,000 at the senior level once equity is counted. Glassdoor‘s median AI engineer pay sits at $146K total with a 25th-to-75th band of $117K to $184K, consistent with the KORE1 mid-level range. The narrow reading: there are two AI labor markets, and they do not share a comp band.

Pay variation between sources is structural, not noise. Glassdoor captures a broad self-reported pool (955 salaries) skewing toward full-time employees at non-tech firms. Levels.fyi captures equity-heavy compensation at well-funded employers. They measure different parts of the same market.

6. The talent wars are real and the market is small

Per an August 2026 Axios survey of frontier-lab moves, the top AI researchers can command extraordinary compensation while choosing among companies with different cultures, missions, and technical resources. Specific moves worth noting:

Lilian Weng, co-founder of Mira Murati’s Thinking Machines Lab, rejoined OpenAI to work on recursive self-improvement in August 2026.
Noam Shazeer left Google for OpenAI in June 2026. John Jumper, the 2024 Nobel Prize winner in Chemistry for AlphaFold work, left Google DeepMind for Anthropic the same month.
Andrej Karpathy, OpenAI co-founder and former Tesla AI head, joined Anthropic’s pretraining team on May 19, 2026.
Meta lost Andrew Tulloch within a year of his reported $1.5B hire. Anthropic’s Jacob Coxon resigned and walked away from unvested stock and options two months before they were due to vest, in a public statement criticizing the industry’s approach to oversight and testing.

Axios’s framing is the most useful: “Frontier AI now resembles a single, tightly connected ecosystem rather than a collection of isolated rivals.” Companies compete fiercely for talent and customers while simultaneously investing in one another, buying one another’s services, relying on the same cloud providers, and hiring from the same small pool of researchers. The implication for pay and pipeline: the cohort that can command a $1.5B package is countable on one hand, and the bottleneck on the industry’s growth is now human, not compute.

7. What the AI Index says the next twelve months look like

The Stanford HAI 2026 AI Index‘s Executive Summary ends with a sentence that reads as a forecast: “The big story in 2026 in labor will be AI.” Three of the underlying dynamics will dominate the next twelve months regardless of which way the macro numbers swing.

Entry-level hiring will keep tightening. Brynjolfsson’s panel now reports 22-to-25-year-old AI-exposed employment 19% below less-exposed peers. Goldman’s regression finds the gap widening with substitution intensity. There is no current mechanism in either dataset that would reverse this in 2026. The pipeline is failing at the entry rung, not at the senior band.
Frontier-lab pay will keep widening. OpenAI’s stock-based average is $1.5M per worker. Anthropic’s staff SWE total comp is $1.25M. Meta paid $1.5B for a single researcher, then lost him. The dispersion between frontier-lab pay and the rest of the market is now wider than at any point in the prior decade. The dispersion itself becomes a recruiting story.
Productivity macro numbers will lag. Acemoglu’s +0.07pp/year TFP estimate is the conservative anchor. Aghion-Bunel and OECD’s high-exposure case is +1.3pp/year. Real productivity growth will land somewhere in that 18x range. The aggregate number will not be visible in national accounts until 2027 at the earliest.

8. What an AI practitioner should do with this data

Four concrete moves worth making in the next quarter.

Watch the Canaries Dashboard monthly. The Stanford Digital Economy Lab’s Canaries Dashboard is the only close-to-real-time tracker of the 22-to-25 gap. Bookmark the page. Set a calendar reminder. Read the data before reading the next headline about “AI is taking all the jobs” or “AI is creating all the jobs.”
Negotiate from data, not vibes. The right reference set for compensation conversations is Levels.fyi (equity-included) and Glassdoor (broad market) — read both. The right reference set for industry demand is Lightcast’s AI-skills posting breakdown, not generic BLS job counts. Both are pulled into the Stanford HAI report, so citing them in a salary conversation carries weight.
Build a complementary skill, not a substitution one. The Demirer-Musolff-Yang paper’s headline finding is the most actionable in the entire dataset: 240% in commits, 30% in releases. The bottleneck in AI-assisted work is no longer production. It is review, integration, testing, and shipping. The skills that close that gap are the ones whose value goes up, not down, as AI capability scales.
Treat the apprentice layer as the new system to build. The most underdiscussed policy and engineering question of 2026 is: if entry-level roles no longer exist in their historical form, what does the new apprenticeship pipeline look like? Companies that solve that question early will own the senior talent market in 2030.

The honest bottom line

The Stanford AI Index 2026 and the revised Brynjolfsson paper, read together, are the first clean view of AI’s effect on the labor market. The data does not say “AI is taking all the jobs.” It says something more specific: AI is paying the people who can already do the job more than ever, hiring more experienced AI practitioners than ever, and quietly closing the entry-level rung that historically produced them. That is not a story of mass displacement. It is a story of pipeline failure. The two are not the same, and confusing them leads to bad policy, bad hiring, and bad individual decisions.

The right individual reflex in 2026 is to be ruthless about which rung you are climbing and to be deliberate about whether the companies you are considering are rebuilding the rung or removing it. The right organizational reflex is to look at the 22-to-25 gap in your own hiring funnel, not at the aggregate productivity numbers, when you decide whether your apprenticeship layer is producing the senior talent you will need in three years.

The data is published. The dashboard is live. The next twelve months will tell us whether the gap continues to widen, stabilizes, or reverses. Watch the data, not the headlines.

Frequently asked questions

What did the Stanford HAI 2026 AI Index Report actually conclude about jobs?

The Stanford HAI 2026 AI Index, released April 14, 2026 (arXiv:2606.15708), reports that AI hiring is up across the labor market — 2.5% of all US job postings now mention AI skills — but does not by itself isolate the entry-level collapse. The clearest jobs data comes from the companion revised Stanford Digital Economy Lab paper by Brynjolfsson, Chandar, and Chen, which uses ADP payroll data through June 2026 to show that young workers (22 to 25) in AI-exposed occupations are now 19% below less-exposed peers. That gap widened from 15% in July 2025.

Are AI jobs actually disappearing, or just shifting?

Aggregate AI hiring is rising, but the entry-level rung is collapsing. The Brynjolfsson paper explicitly finds “no evidence of widespread, economy-wide job displacement.” The adjustment operates through reduced hiring of new graduates, not mass separations of those already in role. Goldman Sachs estimates AI slows US monthly payroll growth by about 16,000 jobs (net of 25,000 substitutions minus 9,000 augmentations). The pattern is entry-level pipeline failure inside a growing market, not economy-wide collapse.

What does an AI engineer actually earn in 2026?

Two separate markets. At the frontier labs (OpenAI, Anthropic, Meta superintelligence), research scientist base pay runs $245,000 to $685,000 and total comp routinely exceeds $1 million once stock is counted. OpenAI’s 4,000+ employees average $1.5M in stock-based compensation. For the broader market, mid-level AI engineers land $155,000 to $200,000 base, senior engineers $220,000 to $300,000 base, and total comp routinely clears $300,000 once equity is included, per KORE1’s 2026 salary synthesis. Glassdoor’s median sits at $146K total pay for AI engineer roles — narrower than Levels.fyi because Levels.fyi captures equity-heavy frontier-lab data.

Is the AI productivity boom real?

Yes, but unevenly. The Brynjolfsson-Li-Raymond study of 5,179 customer-support agents reports a 14% gain in issues resolved per hour (34% for novices). Cui et al.’s study of 5,000+ developers reports a 26% gain in completed tasks. But the Demirer-Musolff-Yang NBER paper shows 240% in commits falls to 30% in actual software releases. At the macro level, Acemoglu estimates +0.07pp/year of TFP gain over the next decade; Aghion-Bunel and OECD’s high-exposure case reach +1.3pp/year. The 18x range is the honest representation of the macro uncertainty.

What is the Canaries in the Coal Mine paper?

A Stanford Digital Economy Lab working paper by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, first published October 2025 and revised August 12, 2026. It uses high-frequency administrative payroll data from ADP covering millions of US workers to document six facts about AI’s labor-market effects. The headline fact, in the August 2026 revision, is that employment among 22-to-25-year-olds in AI-exposed occupations now sits 19% below less-exposed peers, up from 15% in the prior vintage. The paper’s live dashboard updates the result monthly in collaboration with ADP Research.