AI, Output, and Employment (CESifo Working Paper 12579)
A · Strong evidence Johnston (UT Austin, NBER) and Makridis (ASU, Stanford) estimate AI's sector-level effects with near-universal US administrative data, a difference-in-differences design, placebo tests and robustness checks; method shown, results checkable.
Summary based on the full text·Original: about 60 minutes to read
In short
CESifo Working Paper 12579 (25 March 2026), previously circulated as "The Labor Market Effects of Generative AI". Andrew C. Johnston and Christos A. Makridis find that US industry-state sectors with more AI exposure saw output rise about 7% per standard deviation from 2021, when enterprise AI tools such as the GPT-3 API and GitHub Copilot entered workflows. Employment also rose, but only where AI augments workers. Because output grew faster than pay, labour's share of income fell.
Key numbers
- 7%rise in sector output per one standard deviation of AI exposure, emerging from 2021
- 4%rise in employment where AI exposure is augmentative (needs human collaboration); no significant effect where it automates
- about 29%share of AI-induced output gains that accrue to labour
- roughly 5 percentage pointsfall in labour share of income per standard deviation of AI exposure
- over 95%share of US wage and salary jobs covered by the QCEW employment data used
As reported by the source.
English·PDF·opens ifo.de
What it reports
- One standard deviation more AI exposure raises sector output by 7%, employment by about 4% and wage bills by 5% by the end of the sample.
- Splitting exposure: where AI needs human collaboration, employment rises 4%; where AI can do tasks alone, effects are small and insignificant.
- Only about 29% of the output gain goes to labour. The labour share falls roughly 5 percentage points per standard deviation.
- Wage gains concentrate among younger and more educated people, groups with higher AI adoption.
- Effects start in 2021, earlier than ChatGPT, which the authors argue is the right treatment date for enterprise AI.
- Event-study coefficients for 2017-2019 are near zero (no pre-trends). A placebo using an unrelated occupational trait predicts nothing.
How they got it
State by 3-digit NAICS panel, 2017-2024. Exposure is built from Eloundou et al. (2024) occupation scores weighted by pre-2020 occupation shares. Output from BEA state GDP by industry, employment from the QCEW (over 95% of US wage and salary jobs, including firm entry and exit), wages from the American Community Survey. Identification uses the same industry's different occupation mix across states, with state-by-year and industry-by-year fixed effects and controls for teleworkability and interest-rate exposure.
Limits
- Exposure is a fixed technical measure, not observed AI use.
- Working paper, not yet peer reviewed. Sector-level estimates cannot show which firms or staff gained or lost.
- The 2021 timing differs from studies dated to ChatGPT's November 2022 launch, so results are not directly comparable.
Why it matters for companies
The paper gives large-sample support for designing AI around augmentation: in exposed sectors that used AI with people, output and jobs both grew, while automation-type exposure did not add jobs. The labour-share result is the warning: most gains went to capital, which will shape pay negotiations, works council discussions and how firms explain AI benefits to staff.
Summary based on the full text. Read the original for the full argument and data.Open PDF on ifo.de
- Tags
- AI adoption