---
id: "lib-2025-11-13-canaries-in-the-coal-mine-six-facts-about-the"
type: "library"
title: "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (November 2025 version)"
lang: "EN"
org: "Stanford Digital Economy Lab"
source_type: "Paper"
date: "2025-11-13"
week: "2025-W46"
url: "https://nosov.ai/library/lib-2025-11-13-canaries-in-the-coal-mine-six-facts-about-the/"
original_url: "https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/"
tier: "S"
tier_name: "Shapes the field"
tier_reason: "Brynjolfsson, Chandar and Chen at the Stanford Digital Economy Lab use ADP payroll records for millions of workers; the most cited evidence that generative AI is cutting entry-level employment, with firm-time controls and robustness checks."
summary_basis: "full-text"
original_minutes: "45"
tags: ["early career employment", "ai exposure", "adp data"]
---

# Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (November 2025 version)

Tier S (Shapes the field): Brynjolfsson, Chandar and Chen at the Stanford Digital Economy Lab use ADP payroll records for millions of workers; the most cited evidence that generative AI is cutting entry-level employment, with firm-time controls and robustness checks.

Summary based on the full text.

## In short

The November 2025 version of the Stanford paper that put "canaries in the coal mine" into the AI jobs debate. Using monthly ADP payroll data through September 2025, it documents six facts. Young workers (22-25) in the occupations most exposed to AI lost ground after late 2022 while older colleagues in the same jobs did not. The decline shows up where AI automates work, not where it augments it. Adjustment is through headcount, not pay.

## Key numbers

- 16%: relative employment decline for workers aged 22-25 in AI-exposed occupations, controlling for firm-level shocks
- nearly 20%: fall in employment of software developers aged 22-25 from the late-2022 peak to September 2025
- 6% vs 6-9%: employment change late 2022 to September 2025 for 22-25 year-olds vs older workers in the most exposed occupations
- 15 log-point: relative decline for 22-25 year-olds in the most exposed quintiles vs least exposed, with firm-time effects
- 3.5 and 5 million: workers per month in the main ADP analysis sample

## What it reports

- Workers aged 22-25 in AI-exposed occupations saw a 16% relative employment decline after controlling for firm-level shocks; experienced workers were stable.
- Software developers aged 22-25: employment down nearly 20% from the late-2022 peak by September 2025. Customer service shows a similar pattern.
- In the most exposed occupations, 22-25 year-olds fell 6% from late 2022 while older workers grew 6-9%; overall employment kept growing.
- Using Claude usage data to split automation from augmentation, entry-level employment fell in automative uses and grew in the most augmentative ones.
- With firm-time fixed effects, the most exposed quintiles show a 15 log-point relative decline for 22-25 year-olds; other ages show small, insignificant effects.
- Salaries show little difference by age or exposure, suggesting sticky wages.
- Results hold excluding tech firms and remote-capable occupations, and exposure did not predict youth outcomes before LLMs, including during COVID.

## How they got it

Individual monthly payroll records from ADP, the largest US payroll provider, for 3.5 to 5 million workers a month in the main sample, linked to occupation exposure from Eloundou et al. (2024) and to automation/augmentation shares from the Anthropic Economic Index. Event-study regressions by age and exposure quintile with firm-by-month fixed effects absorb shocks common to all workers at a firm.

## Limits

- ADP's client firms do not exactly match the US economy.
- Exposure measures are model-based; the design shows patterns consistent with AI, not proof that AI caused them.
- Age is imputed to month from birth year; occupation coding relies on ADP job titles.
- Working paper; the authors call for reconciling results with other studies that find smaller effects.

## Why it matters for companies

This is the strongest early signal that the entry rung is where AI first shows up in headcount. Firms that have quietly stopped hiring juniors into automatable roles should decide whether that is a policy, and how they will develop future experienced staff. The automation-versus-augmentation split is a design choice inside companies: roles where AI is set up to support people rather than replace their tasks did not show the decline.

Original: https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/