Looking for the Ladder: Is AI Impacting Entry-Level Jobs?
B · Useful, limited Short EIG essay by Google's chief economist and AI and economy lead rebutting the 'Canaries' paper with 238 million job postings; a credible and influential argument, but descriptive charts only and authored by an AI developer's economists.
Summary based on the full text·Original: about 15 minutes to read
In short
Zanna Iscenko and Fabien Curto Millet argue that the entry-level slump in AI-exposed jobs is mostly a monetary-policy shock, not AI. Job postings in the most exposed occupations peaked in March to April 2022, when the Federal Reserve began raising rates and months before ChatGPT. Exposed occupations sit in rate-sensitive sectors, so "AI exposure" and "interest-rate exposure" are hard to tell apart.
Key numbers
- 16 percentrelative employment decline for ages 22-25 in AI-exposed occupations reported by the 'Canaries' paper, the claim under test
- 238 millionUS job postings (Lightcast, Sept 2019 to Aug 2025) analysed
- 38 percentworkers in the most AI-exposed quintile employed in information, finance and professional services, against less than 2 percent in the least exposed
- 12 percentlarge US businesses using AI to produce goods or services by Q3 2025 (Census BTOS)
As reported by the source.
English·PDF·opens eig.org
What it reports
- The 'Canaries' paper (Brynjolfsson, Chandar and Chen, 2025) found a 16% relative decline for early-career staff in exposed occupations since November 2022. Nearly half of the decline for software engineers and customer service workers had appeared by June 2023.
- Postings for the top exposure quintile peaked in spring 2022 and fell sharply through the year. Junior and senior postings fell roughly in parallel.
- A hiring freeze shrinks the 22-25 cohort mechanically: with no hires and no layoffs for a year, that age group would fall 25% while others stay level.
- 38% of workers in the most exposed quintile are in information, finance and professional services, against less than 2% in the least exposed. These sectors also fell more in the 2020 slowdown, before generative AI.
- Large-firm AI use for production was under 10% planned in Q4 2023 and 12% by Q3 2025. Recent graduates' unemployment is 4.8% (New York Fed).
- Five monitoring priorities: wages as well as headcount, postings as well as employment, several exposure measures, mechanisms, and task change within occupations.
How they got it
Lightcast postings data, more than 238 million US postings from September 2019 to August 2025 across 767 occupations, with exposure quintiles rebuilt from the Eloundou et al. (2024) replication code, the same basis as 'Canaries'. Compared with the Federal Funds rate and Census adoption data.
Limits
- Charts and comparisons only: no regression separates the rate effect from the AI effect.
- Written about the August 2025 version of 'Canaries', which its authors have since updated.
- Both authors are Google economists. The conclusion favours AI developers, which does not make it wrong but is worth weighing.
Why it matters for companies
Before cutting graduate intake on the theory that AI does junior work, companies should check whether their own hiring fell with the rate cycle. The monitoring list is practical for internal workforce analytics: track postings, pay and task content by seniority, and use more than one exposure measure. The essay also warns that a wrong diagnosis leads to wrong remedies, such as AI policy where the problem is the business cycle.
Summary based on the full text. Read the original for the full argument and data.Open PDF on eig.org