At a glance
- Source
- Humlum and Vestergaard (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777 (revised March 2026). doi:10.3386/w33777 · Not peer-reviewed: other researchers have not yet checked it for a journal.
- How it was done
- The authors linked large surveys on chatbot use to official records of pay and hours. They compared workers and workplaces that adopted chatbots with those that did not, before and after. Field study. It follows real work inside a company and compares staff who had the tool with staff who did not yet have it.
- Who took part
- Workers and workplaces in occupations where AI chatbots are relevant.
- Where
- Denmark.
“Technological change reshapes work well before it surfaces in earnings or hours.”
What it found
- Most employers in the occupations studied had adopted chatbots. Workers reported productivity benefits.
- Earnings and recorded hours did not change, for workers or for workplaces. The estimates are precise enough to rule out effects larger than 2%.
- What changed was the make-up of work. Employers reorganised tasks and created new ones: generating content, checking AI output, and fitting AI into existing systems.
- Some users moved to better-paid occupations where chatbots are more useful. They were too few to move the average.
An earlier version of the paper, from May 2025, was titled “Large Language Models, Small Labor Market Effects”. It reported that users saved 2.8% of their working hours on average (Becker Friedman Institute Working Paper 2025-56).
When the result holds
The study looks at a whole labour market, not one well-chosen task. It includes firms that simply made chatbots available as well as firms that redesigned work around them. That is why its numbers are much smaller than those from single-task experiments.
Limits
- A working paper. Figures changed between versions.
- Pay and recorded hours can miss gains that show up as better quality or as new tasks.
Practices that use this study
This study supports:
- Measure time saved per task against a baseline Record how long a task takes and how good the result is before you add AI. Without that starting figure you cannot tell a real gain from a hopeful one. Do first · Effort: low (days, one person)
- Start with one costly process, not a company-wide rollout Pick one process that visibly costs time or money. Apply AI there, with a person checking the output, and expand only when the numbers move. Do first · Effort: medium (weeks, a few people)
- Redesign the workflow instead of adding a tool to it Giving people an AI tool inside an unchanged process changes little. Decide again who does each step, in what order, and where a person checks. Do next · Effort: high (months, several teams)