Labor Demand in the Age of Generative AI: Early Evidence from the U.S. Job Posting Data (Policy Research Working Paper 11263)
A · Strong evidence World Bank Policy Research Working Paper by Liu, Wang and Yu, a background paper for the World Development Report 2026, using 285 million Lightcast postings with difference-in-differences and a verified reproducibility package.
Summary based on the full text·Original: about 50 minutes to read
In short
A November 2025 World Bank working paper that treats ChatGPT's release as a shock and compares US job postings for occupations more and less likely to be substituted by AI, holding generative AI exposure similar. Postings for high-substitution occupations fell 12% on average, and the gap widened each year. Entry-level roles, administrative support and professional services took the largest hits.
Key numbers
- 285 millionUS job postings analysed, 2018Q1 to 2025Q2
- 12 percentaverage relative drop in postings for occupations with above-median AI substitution scores
- 6 percent to 18 percentgrowth of that effect from the first to the third year after ChatGPT's launch
- 40 percentrelative drop in postings for administrative support occupations
- 20 percentrelative drop for postings that require little experience
As reported by the source.
English·Web page·opens openknowledge.worldbank.org
What it reports
- 12% average relative decline for above-median substitution occupations, growing from 6% in year one to 14% in year two and 18% by June 2025.
- Larger effects for postings requiring neither an advanced degree (18%) nor much experience (20%).
- By sector: administrative support down 40%, professional services 30%.
- Within similar exposure, demand diverged sharply: dentist postings rose by nearly 50% while secretary postings fell by roughly 50%.
- Alternative specifications give larger effects (for example a 23% average reduction, rising to 31% in year three).
- Results hold in placebo tests and with other exposure indices as controls.
How they got it
285 million Lightcast postings, which capture about 60% of US vacancies recorded by JOLTS, aggregated into 6.8 million state-industry-occupation-quarter cells. Exposure comes from the ILO index of Gmyrek, Berg and Bescond (2023); substitution is the inverse of the Pizzinelli et al. (2023) AI-complementarity index, based on work context (communication, responsibility, physical conditions, criticality, routine) and skills. Difference-in-differences and event-study models, with a reproducibility package verified by the World Bank.
Limits
- Postings are demand signals, not hires or employment; Lightcast misses about 40% of vacancies.
- The substitution index was not built for generative AI, and the treatment is a national date, not firm adoption.
- Other post-2022 shocks (rate rises, tech hiring correction) could hit high-substitution occupations differently.
- US only; working paper, not peer reviewed.
Why it matters for companies
The paper adds a second, larger dataset to the evidence that entry-level and administrative hiring is contracting where AI can substitute for the work. Its use of work context rather than task lists is a practical lens for role reviews: roles with high communication, responsibility and criticality held up better. Firms should check whether their own cuts in junior and admin hiring are deliberate, and plan how those skills will be built in-house if the entry routes shrink.
Summary based on the full text. Read the original for the full argument and data.Read the original on openknowledge.worldbank.org
- Tags
- Job postings, Entry level