The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence
A · Strong evidence ILO research brief by Merola, Ernst and Samaan (ILO) with del Rio-Chanona (UCL) and Teutloff (Oxford) that grades experimental, firm and representative-survey evidence by design and effect size; a careful, checkable synthesis, though it adds no new data of its own.
Summary based on the full text·Original: about 20 minutes to read
In short
A May 2026 ILO research brief summarising empirical evidence on GenAI's effects on productivity, employment and work organisation, building on del Rio-Chanona et al. (2025) and new representative surveys from Australia, Denmark, Germany, Korea, Kuwait, the UK and the US. Its three messages: task-level productivity gains are real but uneven and often unverified; large-scale job loss is not yet visible; and the time workers save is not showing up as more output, pay or employment, which the authors call an "aggregation paradox".
Key numbers
- 51.8%Korean workers using GenAI for their job, saving 3.8% of active working time (Suh and Oh 2026)
- 1.4%share of total working hours saved economy-wide, found in both Korean and US representative surveys
- more than 80%AI-using firms reporting no impact on employment or productivity over three years (Yotzov et al. 2026, nearly 6,000 executives)
- 19 percentincrease in task completion time for experienced open-source developers using AI (METR, Becker et al. 2025)
- around 12 percentlarge US businesses using AI for production by late 2025 (US Census Bureau)
As reported by the source.
English·Web page·opens ilo.org
What it reports
- Simple-task RCTs (writing, coding) show consistent time savings of 20-55%, but some of the speed reflects passive pasting of AI output.
- Complex-task RCTs (legal, consulting, academic writing) are 12-65% faster with mostly positive quality; results depend on whether users can tell when the model is unreliable ("jagged frontier").
- Software: enterprise developer experiments show more completed tasks, mostly for less-experienced staff; experienced open-source developers were 19% slower in METR's trial; GitHub data show 2-4% average gains concentrated among experienced developers.
- Time savings go to on-the-job leisure and task reallocation: in Korea, a 1.3-point rise in on-the-job leisure; Danish registers show about 3% hours saved and no effect on earnings or hours.
- Firm survey (US, UK, Germany, Australia): about 70% of firms use AI, more than 80% of them report no employment or productivity impact; executives average about 1.5 hours of AI use a week. Executives expect -0.7% employment over three years, staff expect +0.5%.
- Entry-level effects are contested: the 16% decline for US workers aged 22-25 in exposed occupations (Brynjolfsson et al.) faces critiques on timing and interest-rate exposure; UK evidence on junior roles is stronger but limited.
- Work organisation: AI can flatten skill hierarchies, raise team coordination costs, reduce diversity of ideas, and mostly reorganises tasks within jobs rather than removing occupations.
How they got it
Literature review classifying studies by design (simple and complex RCTs, field experiments, natural experiments) and by direction of findings on time, quality and inequality, with explicit agreement thresholds. Supplemented with recent representative worker and firm surveys and administrative data.
Limits
- A short brief; detailed methods sit in the underlying review.
- Many cited studies are working papers and preprints.
- Evidence comes mainly from advanced economies plus Kuwait and China; little from low-income countries.
- Most time-savings figures are self-reported.
Why it matters for companies
The brief is the best short answer to a board asking "where are the AI gains?". Saved minutes rarely become output unless work is redesigned and measured, and most firms lack the KPIs to see either. Pilots should therefore define in advance what freed time is for and how output will be counted. The evidence that gains favour less-experienced staff on simple tasks but can widen gaps on judgment-heavy work is directly relevant to who gets trained first and how AI use is supervised.
Summary based on the full text. Read the original for the full argument and data.Read the original on ilo.org