AI, Productivity, and Labor Markets: A Review of the Empirical Evidence
B · Useful, limited International Center for Law & Economics issue brief by Eric Fruits and Kristian Stout; a competent, well-referenced literature review with an annotated bibliography, but secondary and written toward a deregulatory policy conclusion.
Summary based on the full text·Original: about 45 minutes to read
In short
A February 2026 review of the empirical work on AI, productivity and jobs. Its reading: experiments show large task-level gains, especially for less experienced staff; aggregate labour data through 2024-2025 show little disruption, with effects concentrated in entry-level roles in exposed occupations; macro forecasts disagree because of assumptions, not data. It concludes the evidence gives little support for regulation premised on mass displacement or data monopolies.
Key numbers
- 15% to more than 50%reductions in task-completion time reported by experiments across writing, support, coding, accounting, law and translation
- less than 0.66%Acemoglu's (2025) estimate of AI's total factor productivity gain over 10 years
- 35.9%US workers who used generative AI by December 2025 (Hartley et al., 2026)
- roughly 16%employment decline for ages 22-25 in highly exposed occupations (Brynjolfsson, Chandar and Chen, 2025)
As reported by the source.
English·Web page·opens laweconcenter.org
What it reports
- Macro forecasts range from Acemoglu's under 0.66% TFP gain over 10 years to Goldman Sachs' 7% rise in global GDP (about $7 trillion). Penn Wharton estimates about 10% of US GDP could be affected short term, 15% over two decades.
- Labour data show null or small aggregate effects: Yale Budget Lab finds no link between exposure and unemployment through August 2025; Humlum and Vestergaard find near-zero earnings effects in Denmark.
- Entry-level effects: about 16% decline for ages 22-25 in exposed US occupations; UK exposed firms cut junior hiring (Klein Teeselink, 2025).
- Micro evidence: Noy and Zhang, about 40% less time and 18% higher quality; customer support, 15% average and 36% for the bottom quintile; Copilot, 55.8% faster; field experiments with nearly 5,000 developers, 26.08% more weekly tasks; law students, 50% to 130% faster.
- Gains are largest for lower performers, compressing skill gaps within occupations.
- On competition: AI lowers minimum viable scale downstream; open-source and contracting weaken the case for mandated data access.
How they got it
Narrative review of published papers, working papers and official reports from 2021 to early 2026, with an annotated bibliography. No new data or meta-analysis.
Limits
- Selective rather than systematic: no stated search or inclusion criteria.
- ICLE's policy conclusions (against new regulation) shape the framing, for example reading entry-level effects as "adjustment".
- Many cited studies are working papers on early, short windows.
Why it matters for companies
A compact map of the evidence for anyone building an AI business case: task-level gains are real and well measured, firm-level and economy-wide gains are still uncertain, and junior roles are where change shows first. The skill-compression finding argues for targeting AI tools and training at newer and lower-performing staff, while planning how juniors still get the practice that used to come from routine work.
Summary based on the full text. Read the original for the full argument and data.Read the original on laweconcenter.org