Why AI Job Losses Show Up Before New Jobs: A Two-Margin Study
Through early 2026, US industries more exposed to language models lost no more jobs than their peers. A new paper explains why payroll data see losses first.
Across 72 US industries, those more exposed to large language models have lost no more jobs at existing establishments since ChatGPT's release than their peers, and have gained no more at newly opened ones. [Fact] A November 2026 working paper by Mark Esposito and Liu Zhang of micro1 argues that payroll statistics record job losses at established employers promptly and job creation at new firms years later.
New firms supplied the net job growth of the computer era
From 1982 to 1998, newly formed US firms added an average of 2.78 million payroll jobs a year. Firms aged one to five lost a net 0.44 million a year and older firms 0.52 million. [Fact] Employment at firms that entered in 1978 or later rose from 2.55 million in 1978 to 45.68 million in 1998, while employment at firms already operating before 1978 fell by 5.69 million. By 1998 the later entrants held 42.4% of payroll employment. Entering cohorts kept about 80% of their initial payroll after five years. [Fact]
The timing differed by industry. By 1982, industries one standard deviation more exposed to automation patents had 4.95 percentage points less employment at their pre-1977 firms. The offsetting gain at later entrants built up over the following years, and by 1998 its standard error exceeds the estimate. [Fact]
How fast AI arrives sets the depth of the dip
In the authors' model, one AI capability lowers the cost of existing output and makes additional products worth developing. Founders' own time is counted separately from hired hours, and existing products take time to expand sales after their costs fall.
If compute efficiency rose 50% at once, hired hours would fall 4.58% on impact and recover in about three years. If the same gain diffused over seven years, hired hours would reach a low point 0.23% below baseline near year 3.8, recover around year 6.8 and stand 0.32% above baseline at year 15. [Estimate] In the benchmark without the scaling delay, real wages are 4.41% higher and consumption 6.51% higher at year 15. With stronger income effects, households choose fewer hours, and hired hours remain about 0.5% below baseline at year 15 while consumption rises. [Estimate]
A recording delay enlarges the measured loss. If hours in new activities enter the statistics one to three years late, the 0.23% low point is measured as 0.6% to 1.0%. [Estimate]
Payroll through the first quarter of 2026 shows no exposure gap
The authors use Bureau of Labor Statistics establishment job flows. Between the fourth quarter of 2022 and the first quarter of 2026, establishments that existed before ChatGPT lost 13.6% of 2019 employment through contractions and closings net of expansions, and openings added 16.2%. Industries one standard deviation more exposed to language models lost 14.7% and industries one standard deviation less exposed lost 12.5%. The gradient is 1.1 points with a standard error of 1.9. [Fact]
Two studies cited in the paper find firm formation up by more than 10% and by about 20% in industries where AI can assist the tasks of starting a firm. [Fact] The number of US businesses without employees rose 5.1% in 2024, or 1.56 million, against a net private payroll gain of 0.89 million. [Fact]
What this means for workers, and how to prepare
The paper expects the first measurable changes inside occupations and in the hiring of new entrants to exposed jobs, ahead of any change in industry payroll totals. [Claim] Its examples are closest to software developers, customer service representatives and general office clerks. [Estimate]
- If you are looking for a first job in one of these occupations, include young firms in your search. In the computer era, new employers accounted for all net job creation while established firms shrank.
- If your employer automates part of your work, learn to review and correct AI output. In the model every AI-assisted task still requires paid human hours for review and correction.
- If you work as a contractor or run a business without employees, plan for income outside payroll. That category grew by more than payroll jobs in 2024.
Limits
The model results are calibrated simulations, and the 1998 entrant estimates are imprecise. The authors state that the 2026 null result cannot separate slow diffusion from no AI effect on payroll, and that the growth of businesses without employees cannot be attributed to AI. The paper is a preprint that has not been peer reviewed, both authors are affiliated with the company micro1, and the occupation links to our job pages are our match.
Sources
- Mark Esposito and Liu Zhang, "AI and the Two Margins of Work: Why Job Losses Show Up Before Job Creation," arXiv:2611.17530, November 12, 2026: https://arxiv.org/abs/2611.17530
About this analysis
This article was written with AI assistance from the full text of the paper. Figures marked [Fact] are reported in the paper, figures marked [Estimate] in the model section are the authors' simulation results, and the [Estimate] in the last section is our occupation match.
Aktualisierungsverlauf
- Erstmals veröffentlicht am 17. November 2026.
- Keine wesentlichen Aktualisierungen seit der Erstveröffentlichung.