Observed · U.S. monthly data
The labour market is cooling.
That is not the same as an AI collapse.
Job openings have fallen since ChatGPT arrived, while unemployment is modestly higher. But openings had already peaked in March 2022—eight months before ChatGPT. These lines describe the market; they do not assign a cause.
Sources: BLS/FRED JOLTS, monthly openings and hires ↗ + BLS/FRED unemployment ↗ · Scope & caveats ↓
Live indicator
U.S. job openings
Loading official monthly observations…
Focus the chart and use the left and right arrow keys to inspect monthly observations. The complete values are also available in the data table below.
Source: U.S. BLS via FRED, JOLTS job openings, monthly ↗ · Scope & caveats ↓
View the chart as data
Exposure estimate · Global
One in four workers is exposed.
Only a small share is highly exposed.
The ILO asks whether generative AI can perform tasks inside an occupation. It does not count redundancies. That distinction changes the story.
Some GenAI exposure≈25%
Highest exposure category3.3%
Highest exposure Other exposure Lower / no measured exposure
Source: ILO, Generative AI and Jobs exposure index, 2025 ↗ · Scope & caveats ↓
The likely mechanism
Tasks change before occupations disappear.
Most occupations combine automatable tasks with judgement, physical work, relationships, accountability, or context that still needs people.
The uneven distribution
Exposure is higher for women in the top category.
Women4.7%
Men2.4%
The ILO links the gap to the concentration of women in clerical work, especially in high-income countries.
Source: ILO exposure index, sex comparison, 2025 ↗ · Scope & caveats ↓
Observed association · U.S.
The first effects may appear
at the entry ramp, not the exit door.
Two rigorous analyses can be true at the same time because they look through different lenses: one at a vulnerable subgroup, the other at the entire labour market.
Concentrated signal
−16%Relative employment decline for ages 22–25 in the most AI-exposed occupations.
Stanford’s U.S. payroll-data working paper finds the decline after firm-level controls, while older workers in the same occupations and workers in less-exposed fields were steadier.
Economy-wide counterpoint
No discernible broad disruption in the first 33 months after ChatGPT.
Yale’s Budget Lab found exposure, automation, and augmentation measures were not related to employment or unemployment changes across the whole market.
Interpretation
There is a canary, but not yet a mine-wide alarm. Entry-level hiring can weaken before aggregate employment moves. It can also be affected by interest rates, post-pandemic tech corrections, education choices and remote-work shifts. Watch the subgroup; do not inflate it into an economy-wide count.
What would change our mind?
Signals worth watching next
- 01
PersistenceDo early-career declines continue across revisions and datasets?
- 02
DiffusionDo effects spread from highly exposed occupations to the wider labour market?
- 03
MechanismDo employer records tie fewer hires or separations explicitly to AI deployment?
- 04
CompensationDo wages, hours and job quality move before headcount does?
Forecast · Global to 2030
The scary number is rarely
the whole number.
The World Economic Forum’s headline is not “92 million jobs vanish.” Its employer-survey scenario pairs displacement with more creation—and covers every structural macrotrend, not AI alone.
Projected creation+170M
Projected displacement−92M
Projected net+78M
Scope check: 1.2 billion formal jobs, 2025–2030, across technology, demographics, geoeconomic fragmentation, the green transition and other forces.
Source: World Economic Forum, Future of Jobs Report 2025, forecast to 2030 ↗ · Scope & caveats ↓
22%
Projected structural churn—jobs created plus jobs displaced as a share of the formal jobs studied. It describes movement, not inevitable unemployment.
Workplace experiment · One company
Before AI replaces work,
it can redistribute expertise.
In a study of 5,179 customer-support agents, an AI assistant increased issues resolved per hour—especially for novice and lower-skilled workers. This is strong evidence about a task setting, not a universal employment forecast.
Productivity gain
Issues resolved per hour
The largest gain accrued to workers with less experience. Experienced, highly skilled workers saw minimal impact.
Source: Brynjolfsson, Li & Raymond, Generative AI at Work, 5,179 agents ↗ · Scope & caveats ↓
Adoption context · 2025
88%of surveyed organizations reported using AI.
Adoption is widespread; measured workforce effects are not. That gap is exactly why outcomes must be tracked separately from usage.
Source: Stanford HAI, 2026 AI Index Economy chapter, 2025 survey ↗ · Scope & caveats ↓
Decision discipline · Evidence + conditional judgment
A Factfulness discipline
for uncertainty.
A disorderly AI transition is possible. It is not established as the present economy-wide outcome. Adapted from Gapminder’s Factfulness framework, this chapter checks dramatic narratives against base rates, definitions, rival causes and evidence that could reverse the advice. It implies no Gapminder affiliation and offers neither clinical nor universal career authority.
Start with the whole market.
Subgroup warnings matter, but must be sized against aggregate employment, unemployment, hiring and ordinary churn.
Ask what moved before AI.
Openings peaked before ChatGPT. Timing can fit several explanations.
Potential is not outcome.
Task overlap flags where change could occur; it does not count redundancies.
Jobs bundle unlike work.
Automation of one task can remove, redesign, intensify or complement other tasks.
Local effects stay local.
A customer-support productivity gain shows a mechanism in one setting, not future national employment.
Keep rivals alive.
Rates, demand, post-pandemic normalization, sector cycles, demographics and policy can move alongside AI.
Observed basis
What current evidence permits
- Broad U.S. disruption is not discernible in the cited first 33 months.
- A concerning early-career association appears in highly exposed occupations.
- AI can raise task productivity in a controlled workplace setting.
- Exposure and employer forecasts identify possible pressure, not realized loss.
Prudent judgment · not universal prescription
Robust under several futures
- Judgment: preserve learning capacity and domain knowledge because augmentation and substitution both change tasks.
- Judgment: test tools on real work with quality checks before reorganizing roles around forecasts.
- Judgment: keep financial and organizational slack where feasible because transition speed is uncertain.
- Judgment: protect junior pathways where evidence shows concentrated pressure; do not infer one career answer for everyone.
Reversal indicators
What would change the advice?
- 01
SeparationsPersistent AI-attributed layoffs across independent employer and administrative data.
- 02
HiringBroad, sustained entry-level contraction beyond exposed occupations and ordinary cycles.
- 03
DiffusionWage, hours and employment effects spreading across sectors and countries.
- 04
CounterweightNew tasks, occupations, firms and demand offsetting displacement—or failing to do so.
Framework source: Gapminder, Factfulness ↗. Principles paraphrased and adapted for labour-market evidence.
Current answer · July 2026
So, is AI transforming
the job market?
Yes—but transformation is not a synonym for net job loss.
What is established
- AI exposure reaches a meaningful share of occupations.
- AI can materially change productivity and task performance.
- Early-career workers in highly exposed U.S. occupations show a concerning employment signal.
What is not established
- That today’s broad market cooling was caused mainly by AI.
- That exposed jobs will disappear one-for-one.
- That economy-wide AI mass unemployment is already visible in the data.
Our working prediction
- Near-term disruption will remain uneven and hiring-led.
- Junior pipelines and routine cognitive tasks deserve the closest watch.
- Job redesign, skill shifts and wage effects may arrive before aggregate headcount losses.
Method & sources
How this briefing stays factual
Label the evidence
Observed outcomes, exposure models, experiments and forecasts never share an unlabeled axis.
Match the scope
A subgroup result stays a subgroup result. A one-company experiment stays a one-company experiment.
Separate timing from cause
“After ChatGPT” is not automatically “because of ChatGPT.”
Keep uncertainty visible
Working papers, survey scenarios, revisions and source limitations travel with the number.
Evidence ledger
Primary and institution-grade sources
- 01Job Openings and Labor Turnover Survey: total nonfarm job openings and hires ↗
Observed U.S. labour-market indicators. They cannot identify whether a change was caused by AI. Values can be revised.
- 02Civilian unemployment rate ↗
Observed U.S. macro indicator; not an AI attribution series.
- 03Generative AI and Jobs: A Refined Global Index of Occupational Exposure ↗
Working Paper 140, DOI 10.54394/HETP0387. Exposure estimates are not counts of jobs lost.
- 04Evaluating the Impact of AI on the Labor Market: Current State of Affairs ↗
Economy-wide observational analysis; explicitly not predictive of future effects.
- 05Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence ↗
Working paper using U.S. payroll-provider data. Strong subgroup signal, but not a peer-reviewed economy-wide causal estimate.
- 06Future of Jobs Report 2025 ↗
Forecast based on the Future of Jobs Survey 2024 and ILOSTAT; covers all structural macrotrends, not AI alone.
- 07Generative AI at Work, Working Paper 31161 ↗
Workplace study of 5,179 customer-support agents; scope limits generalization.
- 082026 AI Index Report — Economy ↗
Synthesis of investment, adoption, labour-market and productivity evidence; underlying sources vary by indicator.
- 09Factfulness ↗
Critical-thinking framework for checking dramatic interpretations. Principles are paraphrased and adapted here; no affiliation or professional authority is implied.
Also in Signals · Brief 004
Is science getting faster—or merely producing more?
Separate investment and output from reliability, translation and demonstrated AI-assisted acceleration.
InputsVolumeReliabilityTranslation
Open Brief 004 →