Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Resume and Job Posting Data
A · Strong evidence Harvard economists Hosseini and Lichtinger use linked resume and posting data for 285,000 US firms with difference-in-differences, triple-difference and event-study designs; strong, transparent method, still an SSRN working paper.
Summary based on the full text·Original: about 55 minutes to read
In short
The paper asks whether generative AI hits junior staff harder than senior staff. Using US LinkedIn-based resume and job-posting data, it marks a firm as an adopter once it advertises a dedicated "GenAI integrator" role. From 2023Q1, junior employment at adopters fell relative to non-adopters while senior employment kept rising. The decline comes from slower hiring of juniors, not more exits or fewer promotions, and is concentrated in high-exposure occupations.
Key numbers
- nearly 62 millionworkers covered by the Revelio Labs resume and posting data, 2015-2025
- 10,599firms classed as GenAI adopters, about 3.7 percent of firms but 17.3 percent of positions
- 7.7 percentrelative decline in junior employment at adopting firms six quarters after 2023Q1 (difference-in-differences)
- roughly a 10 percentjunior-versus-senior drop within the same firm after six quarters (triple-difference)
As reported by the source.
English·Web page·opens papers.ssrn.com
What it reports
- Data: about 285,000 firms, over 150 million employment spells, nearly 62 million workers and almost 200 million postings (Revelio Labs, 2015-2025).
- 10,599 firms adopted GenAI by the paper's definition: 3.7 percent of firms, 17.3 percent of positions. Adoption was flat before 2023 and surged after.
- Junior employment at adopters fell 7.7 percent after six quarters relative to controls; triple-difference estimates show roughly a 10 percent drop.
- Event study: junior decline starts about three quarters after adoption and reaches 8 percent after eight quarters.
- Highly exposed junior roles fell from 2022Q4; low-exposure junior roles did not change significantly.
- Mechanism: hiring of juniors slowed sharply; junior separations fell slightly; promotions were unchanged.
- U-shaped by school tier: steepest declines for graduates of mid-tier (Tier 3) and less selective (Tier 4) schools.
How they got it
Adoption was detected in two steps: keyword flags in postings, then an LLM check that the post is a genuine integrator role. Junior and senior headcount within firms was compared with difference-in-differences, a triple-difference design with firm-by-time and industry-by-seniority-by-time fixed effects, and a staggered event study. Exposure came from Eloundou et al. (2024) via O*NET codes.
Limits
- Adoption is not random: adopters are larger, more technical and more educated, and the authors cannot rule out other explanations.
- The proxy misses "silent" adoption by staff using tools without an integrator hire; pooled event-study estimates show pre-trends, removed when the information sector is excluded.
- Firm-level effects do not translate directly to the aggregate economy (the "missing intercept" problem).
- The hiring cuts may reflect firms' expectations of future automation rather than tasks already automated.
Why it matters for companies
If firms are cutting junior hiring ahead of actual automation, they risk thinning the pipeline that produces future seniors. Workforce plans should track entry-level intake separately from total headcount and decide deliberately how juniors will learn work that AI now does. The finding that promotions did not change suggests the senior layer is not yet being refilled from below any faster or slower, which makes the gap visible only years later.
Summary based on the full text. Read the original for the full argument and data.Read the original on papers.ssrn.com
- Tags
- Job postings