🏛️ Issues and Advocacy Training
Courses › AI Issues: Technical, Legal, Social and Logistical

Lesson 6 of 9 · about 16 minutes

Jobs and the Economy: What the Research Shows So Far

You'll be able to

  • Tell exposure estimates apart from measured job effects
  • Summarize the main studies on AI and employment through 2026 and their limits
  • Explain why productivity studies disagree
  • Compare macroeconomic forecasts and the assumptions that drive them
  • Check a jobs claim for its method and data source

The jobs question gets more confident predictions than any other AI topic and has the thinnest evidence base. As of October 2026 the honest summary is this: there is no sign of economy-wide job losses from AI, there are early and disputed signs of weaker hiring for young workers in exposed fields, and productivity results run from big gains to measurable slowdowns depending on the task and the study. Long-range forecasts disagree by a factor of several, mostly because they start from different guesses.

Three numbers that get confused

Kind of numberWhat it measuresTypical sourceExample
ExposureHow many tasks in a job overlap with what AI can doTask lists matched against AI capabilitiesIMF: nearly 40 percent of global jobs exposed, about 60 percent in advanced economies (Jan 2024)
ProductivityHow much faster or better a worker does a task with AIExperiments and field studiesCustomer support agents about 14 percent more productive (Brynjolfsson, Li, Raymond)
EmploymentWhether jobs, hiring or wages actually changedPayroll and survey data over timeYoung workers in exposed jobs about 19 percent below trend (Stanford, Aug 2026)

Exposure is not loss. The IMF's own analysis estimated that about half of exposed jobs in advanced economies could gain from AI rather than shrink. A radiologist is "exposed" to image-reading AI. So far that has not meant fewer radiologists.

What has been measured

Employment

The most watched study is "Canaries in the Coal Mine?" by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab, built on ADP payroll records for millions of US workers. Their August 2026 revision, with data through June 2026, finds employment of workers aged 22 to 25 in AI-exposed occupations is about 19 percent below where it would be if it had tracked less-exposed peers. Experienced workers in the same jobs show no such gap. The decline comes mainly from fewer hires, not more firings. Where AI mostly helps workers rather than replacing tasks, employment is flat or rising.

The authors are careful about limits. They call the findings descriptive, not proof that AI caused the gap. The gap holds when tech companies are excluded and when they control for interest rates and remote work, but it shrinks when they control for education, some trends started before ChatGPT, and the gap looks larger in ADP data than in national surveys.

The Budget Lab at Yale took a different approach in October 2025, measuring how fast the overall mix of occupations has changed since ChatGPT's release compared with the PC era, the internet era and a quiet control period. Change was only slightly faster than in past periods, and the shifts in information, finance and professional services started before ChatGPT. Exposure and actual usage data showed no clear link to job or unemployment changes. They concluded there was no discernible AI-driven disruption yet, while saying it is too early to know long-run effects.

These two findings fit together better than the loudest claims suggest: no broad disruption, plus a possible early squeeze on entry-level hiring in specific fields. Watch for new data on recent graduates.

Productivity

Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents given an AI assistant. Issues resolved per hour rose about 14 percent on average and about 34 percent for novices, with little change for the best agents. The paper was published in The Quarterly Journal of Economics in 2025. The AI effectively passed the top agents' habits to new hires.

METR ran a randomized trial in early 2025 with 16 experienced open source developers working 246 real tasks in code they knew well. With AI allowed, they took 19 percent longer. Before starting they predicted AI would speed them up 24 percent, and afterward they still believed it had sped them up about 20 percent. METR warns the sample is small, the developers were experts on familiar code, and the tools were early 2025. A February 2026 follow-up with newer tools hinted at a speedup, but METR called its own data very weak evidence because developers expecting big gains often declined to take part.

Put those side by side. AI helps most where a worker lacks the knowledge the AI has. It can slow down experts in work they already do well. And people are poor judges of their own AI speedup, which matters every time a survey asks workers how much time AI saves them.

Forecasts and why they disagree

ForecastClaimKey assumption
Goldman Sachs, April 2023Global GDP up about 7 percent (near $7 trillion) over a decade; productivity growth up 1.5 points a year; about 300 million jobs exposedBroad adoption; 25 to 50 percent of workload in exposed US jobs replaceable
Daron Acemoglu, MIT, 2024US total factor productivity up no more than about 0.7 percent over 10 years; GDP up roughly 1 percentAbout 20 percent of tasks exposed, about a quarter of those profitably automated in 10 years, so about 4.6 percent of tasks affected
Dario Amodei, Anthropic CEO, May 2025 (interview with Axios)AI could eliminate about half of entry-level white-collar jobs and push US unemployment to 10 to 20 percent within one to five yearsRapid capability gains and fast adoption by employers; a warning from an industry leader, with no published model behind the numbers
IMF staff, January 2024About 40 percent of global jobs exposed; about half of exposed jobs in rich countries may benefitTask exposure plus complementarity with human skills

Example: seeing the assumption do the work

Acemoglu's arithmetic is simple enough to follow. About 19.9 percent of US labor tasks are exposed. He assumes 23 percent of those can be profitably automated within 10 years: 0.199 times 0.23 is about 0.046, so 4.6 percent of tasks. He estimates cost savings on those tasks of about 15.4 percent. Multiply 4.6 percent of the economy by about 15 percent savings and you get roughly 0.7 percent, his productivity ceiling.

Amodei's warning sits at the other end. He runs a company building these systems and says he sees capability trends early; skeptics note that alarming forecasts also signal how powerful a product is. Neither side has measured data for the 2030s. Now go back to Acemoglu and change one input. If 60 percent of exposed tasks could be profitably automated instead of 23 percent, the affected share becomes 0.199 times 0.6, about 12 percent, and the productivity gain nearly triples. Neither number is measured. Both are judgment calls. When two forecasts disagree, find the adoption assumption first.

Common mix-ups

  • Claim: AI will eliminate 300 million jobs. What the evidence shows: Goldman Sachs said 300 million full-time jobs could be exposed to automation, which is different from eliminated.
  • Claim: AI is already causing mass unemployment. What the evidence shows: Yale's Budget Lab found no discernible economy-wide disruption through late 2025, and the 2026 International AI Safety Report saw no visible effect on overall employment. The Stanford team's 2026 revision finds a specific hiring gap for young workers in exposed jobs.
  • Claim: Workers say AI saves them hours a week, so productivity is up. What the evidence shows: METR's developers believed they were 20 percent faster while measuring 19 percent slower. Self-reports need checking against output.
  • Claim: Layoff announcements that mention AI prove AI caused the cuts. What the evidence shows: Companies cite AI for many reasons, including to signal efficiency to investors. Look at what roles were cut and whether hiring resumed.

Reporting angle

  • When a local employer announces AI-related cuts or a hiring freeze, ask which job titles, how many positions, whether the work moved to AI, to contractors or nowhere, and what the headcount was a year earlier.
  • Ask your state labor department for job postings and unemployment claims by occupation over time. Entry-level postings in customer service, bookkeeping and junior coding are good ones to watch.
  • Ask local colleges about job placement rates for recent graduates by program, and whether employers told them hiring plans changed.
  • When a government agency says AI will save staff time, ask for the baseline, how the savings will be measured, and whether the agency plans to cut positions or redeploy them.

Research angle

Check yourself

1. Goldman Sachs estimated in 2023 that about 300 million full-time jobs could be 'exposed' to automation by generative AI. What does 'exposed' mean here?

2. The Stanford Digital Economy Lab's 'Canaries in the Coal Mine' paper, revised August 2026, found what?

3. Yale's Budget Lab (October 2025) compared how fast the mix of jobs has changed since ChatGPT with earlier tech eras. What did it find?

4. In the Brynjolfsson, Li and Raymond customer support study, who gained the most from an AI assistant?

5. METR's 2025 randomized study of experienced open source developers found that allowing AI tools made them how much faster or slower?

6. Why do Goldman Sachs (about 7 percent global GDP over a decade) and Daron Acemoglu (about 1 percent US GDP over a decade) differ so much?

Terms in this lesson

Exposure
How much of a job's task list overlaps with what AI can do. Not a forecast of job loss.
Augmentation
AI helping a worker do a task faster or better, as opposed to doing it instead of the worker.
Total factor productivity
Output growth not explained by more workers or more machines. The economist's measure of doing more with the same.
Randomized controlled trial
A study that randomly assigns who gets a treatment, such as AI tools, so differences can be credited to it.
Payroll data
Employment records from payroll processors like ADP, used to track hiring and jobs month by month.

Full glossary

Sources