---
id: "lib-2026-09-16-artificial-intelligence-and-labor-market"
type: "library"
title: "Artificial Intelligence and Labor Market Reallocation"
lang: "EN"
org: "Federal Reserve Board"
source_type: "Paper"
date: "2026-09-16"
week: "2026-W38"
url: "https://nosov.ai/library/lib-2026-09-16-artificial-intelligence-and-labor-market/"
original_url: "https://conference.nber.org/conf_papers/f247923.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "92-page Federal Reserve Board working paper presented at an NBER conference, combining CPS, JOLTS and Lightcast microdata with instrumental-variable and Bayesian trend-cycle methods to estimate LLM effects on hiring and the natural rate of unemployment."
summary_basis: "full-text"
original_minutes: "120"
tags: ["job finding", "natural rate"]
---

# Artificial Intelligence and Labor Market Reallocation

Tier A (Strong evidence): 92-page Federal Reserve Board working paper presented at an NBER conference, combining CPS, JOLTS and Lightcast microdata with instrumental-variable and Bayesian trend-cycle methods to estimate LLM effects on hiring and the natural rate of unemployment.

Summary based on the full text.

## In short

Hie Joo Ahn and Nicholas A. Carollo, economists at the Federal Reserve Board, measure how large language models have changed the flow of US workers into and out of jobs (paper dated 16 September 2026). Workers in jobs with both high AI exposure and high AI adoption have seen job-finding and job-switching rates fall more than other groups since late 2022, while layoff rates barely changed. The adjustment runs through less hiring and more task reshuffling inside existing jobs. They estimate the natural rate of unemployment has risen by about 0.1 to 0.2 percentage point, with considerable uncertainty.

## Key numbers

- 0.1-0.2 pp: estimated rise in the US natural rate of unemployment since LLMs were introduced, through mid-2026
- 1.2 pp vs 0.4 pp: unemployment rise Dec 2022 to Jul 2026 in the top-decile AI-adoption industries versus the bottom quartile
- ~1.1 pp: rise in the weighted standard deviation of individual hirability, largely predicted by AI factors

## What it reports

- High-exposure, high-adoption workers have lower unemployment than others on average, because they are laid off less; but since LLMs their unemployment rose more.
- In industries in the top decile of AI adoption, unemployment rose 1.2 percentage points between December 2022 and July 2026, versus 0.4 point in the bottom quartile, even though the top decile is normally the least cyclical.
- The high-exposure/high-adoption group's unemployment rose slightly below 1 point from end-2022 to July 2026, versus 0.3 point for the low/low group and about 0.5 point for high-exposure/low-adoption.
- Job-finding rates in the top adoption decile fell from late 2022 to levels comparable to the COVID-era trough by 2023 and stayed low; separation rates show little difference across groups.
- Within-job activity switching rose disproportionately for high-exposure workers, and the pre-LLM wage-growth and hours premium for AI-exposed work weakened markedly.
- Dispersion in individual hirability rose by about 1.1 percentage points (0.6-0.7 under fixed-effects variants), mostly attributable to AI factors; dispersion in separability did not rise.
- An alternative instrument implies a natural-rate increase of about half a percentage point. The authors also note recent signs of recovery in job-finding for the most exposed group.

## How they got it

Current Population Survey microdata (2013 to July 2026) matched to occupation-level exposure scores from Eloundou et al. (OpenAI) and an industry-level adoption measure from Lightcast job postings mentioning AI or machine learning skills, checked against the Census BTOS survey. JOLTS vacancies are allocated to occupations to build market tightness. Hiring effects are identified with industry-by-month fixed effects and a shift-share instrument; the natural rate comes from a Jackman-Roper mismatch model and a Bayesian trend-cycle decomposition of flows.

## Limits

- Adoption is inferred from job-posting text, which can miss use by current staff or pick up boilerplate AI disclaimers.
- The effects sit in the top 5-10% of the exposure and adoption distribution; coarser studies find little, so results depend on how groups are cut.
- Weak hiring may partly reflect general macro uncertainty rather than LLMs alone, as the authors acknowledge.

## Why it matters for companies

The paper locates where AI is showing up in the labour market: in fewer hires and redesigned tasks, rather than layoffs. For employers this means the effect is easy to miss internally, because no one is let go; roles simply are not backfilled or opened, and existing staff absorb changed work. It also implies that candidates from highly exposed occupations face longer searches, which affects both internal mobility and external hiring pools. Tracking hiring and task-change data by role, not only headcount, is how a company would see this in its own numbers.

Original: https://conference.nber.org/conf_papers/f247923.pdf