AI & the Labor Market: What Happens to Work

Stanford Digital Economy Lab's August 2026 update (ADP payroll data through June 2026) finds employment of 22-25 year olds in highly AI-exposed occupations about 19% below the less-exposed trend, up from 15% at the July 2025 vintage. From November 2022 to June 2026 that group fell about 11% in the two most exposed quintiles and grew about 10% in the three least exposed. The authors stress the results are descriptive, not causal, find no economy-wide displacement, and see the adjustment in reduced hiring rather than separations.

Census BTOS data for 14 Dec 2025 to 3 May 2026 show AI use between 17% and 20% of firms (19.8% at the latest reading), with 20-23% expecting use within six months. Use is 37% at firms with 250+ employees, 39.7% in Information and 33.9% in Finance and Insurance. Census changed the wording in November 2025 from AI use in producing goods or services to use in any business function, so comparison with the earlier sub-10% figure is not exact. The Federal Reserve (3 Apr 2026) reported about 41% of the workforce using generative AI for work in Nov 2025, with daily use near 12%.

Challenger, Gray & Christmas data (as reported by Startup Fortune, Jul 2026) show 112,713 announced US job cuts citing AI through July 2026, more than double the 54,836 for all of 2025, and AI was the leading stated reason for the fifth straight month. Total announced cuts through July were 477,033, down 41% year on year. These are employer-stated reasons, not measured job losses, and the figures are from a secondhand report.

OfficeChai (21 Feb 2026) reported METR measured a 50% time horizon of about 14.5 hours for Claude Opus 4.6, with a fitted doubling time near 123 days on 2023-onward data and a very wide confidence interval (6 to 98 hours). METR's task suite is near saturation and covers well-specified software, ML and cybersecurity tasks, so it does not equal a full professional role.

Census data published 1 Jun 2026 (Bloomberg Government) showed US data-center construction spending above a $50bn annualized rate for the first time in April 2026 (about $50.7bn, versus $39.8bn a year earlier), 2.3% of all US construction spending. The reinstatement side of the ledger keeps growing.

Not refreshed: Goldman 300M jobs estimate, the 2023 task-exposure study (~80%), Bick/Blandin/Deming adult-use curves, MIT pilot and HBR workslop figures, radiologist pay, corporate layoff anecdotes, developer-usage share and the trades shortage forecast.

A doctor diagnoses, orders tests, interprets results and coordinates care. Automate one task and you reshape the role; you rarely erase it.

1. Layer 0, Tasks: every job is a bundle of tasks; AI acts on tasks, not titles.
2. Layer 1, Exposure: which of those tasks current AI could technically do.
3. Layer 2, Adoption: whether firms actually deploy it, usually slowly and unevenly.
4. Layer 3, Response: the net effect: augment, automate, or create new work.

The tasks large language models do best (reading, writing, analysing, coding) are exactly the ones offices pay for. The office, not the job site, is where exposure concentrates.

• Break the job into tasks. Every occupation is a task list; O*NET catalogues them.
• Score each task. Can current AI plausibly do it? (task-to-ability mapping)
• Weight and sum. The result is an AI Occupational Exposure index, 0-100.
• Read it correctly. Exposure is potential, not destiny; adoption decides the rest.

Not because the capability is fake, but because deployment is the bottleneck. AI is a general-purpose technology, and those pay off only after firms reorganise around them, which takes years.

Augmentation keeps the task in human hands and raises the worker's wage. Automation transfers the task, and its value, to whoever owns the machine. The same tool can do either.

A 2x2 of AI exposure (horizontal) against human-judgment premium (vertical). Human-anchored, high judgment and low exposure: surgeon, teacher, home-health aide, electrician (the physically shielded low-exposure side). Augmented, the sweet spot, high exposure and high judgment: radiologist, software developer, financial analyst. Automation-exposed, high exposure and low judgment: paralegal, transcriptionist, data-entry clerk.

1. Autocomplete: code and email drafts (augment, human keeps the task).
2. Expertise multiplier: juniors lifted to expert.
3. Human-in-the-loop: AI proposes, human signs off.
4. Task substitution: one does five workers' output.
5. Full automation: task leaves human hands (automate, machine takes it).

In 2016 a 'godfather of AI' said to stop training radiologists. AI now beats them on narrow benchmarks and holds FDA clearances, yet demand rose. Reading images is a fraction of the job, and cheaper scans (Jevons' paradox) mean more scans ordered, not fewer readers needed.

Stanford's August 2026 update finds employment of 22-to-25-year-olds in the most AI-exposed occupations about 19% below where it would be had it tracked less-exposed jobs (15% a year earlier), while their peers in hands-on work kept hiring. The career ladder is thinning at the bottom rung first.

The slide illustrates displacement as one central node losing its links to surrounding nodes, and reinstatement as a new, denser network of connected nodes.

1. Capability sets exposure.
2. Firms choose to adopt.
3. Hiring, wages and skills shift.
4. Workers reskill, resist, negotiate.
5. Demand steers what AI is built next.

The last computer wave hollowed out the middle. This one reaches higher up the ladder, and could bend the skill premium either way, toward equality or away from it.

• The skill premium. The wage gap education buys: AI could shrink or widen it.
• Polarization. Routine middle-skill work hollows; the tails grow.
• The equaliser effect. AI lifts novices most, compressing some gaps (call centres).
• The amplifier risk. If AI complements top judgment, high earners pull further ahead.

When telephone operators were automated, young women found better clerical jobs within years, but the career operators never recovered their lost earnings. AI's transition will rhyme: society gains on average while specific people, cohorts and towns pay the bill. Averages hide the pain.

Looms, switchboards, ATMs, spreadsheets: the task changed, the headcount often grew, and entirely new work appeared that no one had forecast.

1. 1810sPower looms: Luddites smash frames; textile employment then grows for decades.
2. 1920-40sAutomatic switchboards: operator jobs fall 50-80%; women move to clerical work.
3. 1970sATMs: tellers predicted extinct; teller headcount rises into the 2000s.
4. 1980sSpreadsheets: bookkeeping clerks decline; financial-analyst jobs boom.
5. 2000se-Discovery: software reads documents; paralegal jobs grow +50k.
6. 2020sGenerative AI: cognitive tasks exposed; the experiment now running.

Automation has never simply subtracted jobs from a fixed pile of work. It changes the tasks, moves the people, and, given time, invents work no one could have named beforehand. The danger has always been the transition, not the destination.

• Felten, Raj & Seamans: Occupational exposure to AI / AIOE (2021).
• Eloundou, Manning, Mishkin & Rock: 'GPTs are GPTs' (OpenAI, 2023).
• Goldman Sachs: 'The Potentially Large Effects of AI on Economic Growth' (2023).
• OECD: Employment Outlook: AI and the labour market (2023).
• Brynjolfsson, Chandar & Chen: entry-level jobs and AI (Stanford Digital Economy Lab, 2025).
• Dell'Acqua et al.: 'Navigating the Jagged Technological Frontier' (BCG x Harvard, 2023).
• Acemoglu: 'The Simple Macroeconomics of AI' (NBER, 2024); Autor: 'New Frontiers' (2024).
• Bessen (2015); Bick, Blandin & Deming (2025); METR (2025); US Census BTOS. Figures marked illustrative are composites.