The BaseResearchNBER
Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI
Humlum and Vestergaard link Statistics Denmark surveys of chatbot use in 11 exposed occupations to administrative earnings and hours records through December 2024. They report that most employers have adopted chatbot initiatives and workers take on new tasks, yet find precise null effects on earnings and hours, ruling out changes above 2%, with occupation switching the one measurable shift.
Open the source ↗Top story of the week
Why it matters
The same thing, for different readers.
What it means for you
Danish workers who use chatbots, even daily, show no clear gap in pay or hours growth from non-users so far; some report new tasks, and some switch occupations.
For lawmakers
The authors say tracking this needs linked data on how work is reorganized inside firms as well as on pay and hours; conventional statistics do not capture the reorganization.
For researchers
Splitting early-career employment trends by firm adoption, the authors find the declines are not driven by adopters; firms acting in anticipation would bias estimates toward zero.
The question it raises
More than a third of the new AI-related tasks workers report involve reviewing AI output or compliance, which raises who is responsible for what the tools produce.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“What does not move is the surface—recorded hours and earnings. Using difference-in-differences, we estimate precise null effects at both the worker and workplace levels, ruling out effects larger than 2% two years after ChatGPT’s launch.”
Found word for word in the source; punctuation differs
“While we replicate the pattern of declining early-career employment in exposed occupations in Denmark, our difference-in-differences analysis reveals that the declines are not driven by firms adopting AI chatbots.”
Found word for word in the source
“Occupational mobility is the one margin where these rapid currents break through to administrative labor market outcomes: workers who use AI chatbots are more likely to have transitioned into new occupations, particularly into roles where workers have greater discretion over their toolsets.”
Found word for word in the source
“If so, the absence of measurable labor market effects is not evidence that nothing is happening—but rather that transformation is occurring on margins that conventional economic statistics do not capture.”
Found word for word in the source
“For instance, adopting firms might expand by poaching workers from non-adopting competitors, biasing our workplace employment estimates upward. Alternatively, if firms adjust labor demand in anticipation of AI chatbots—rather than based on their actual adoption—the estimates would be biased toward zero.”
Found word for word in the source
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Statistics Denmark surveys in November-December 2023 and 2024 of workers in 11 AI-exposed occupations (accountants, customer support, financial advisors, HR, IT support, journalists, legal, marketing, office clerks, software developers, teachers); the late-2024 round has 25,000 workers in 7,000 workplaces. Responses are linked to administrative monthly earnings, hours and occupation records through December 2024. Difference-in-differences indexed to ChatGPT's launch in November 2022 compare adopters with non-adopters, and workplaces that encourage chatbot use with those that do not.
- What it found
- About 43% of employers encourage chatbot use, 21% allow it and 6% prohibit it; even where employers have no initiative, about 41% of workers have used chatbots at work. Earnings, hours and wages: estimates center on zero, with confidence intervals ruling out average effects larger than 2% (3% in the dynamic estimates); the nulls also hold among daily users, and no occupation shows a significant effect. Workplaces encouraging chatbots show no differential change in total hours (effects above 2% ruled out), job churn, or the early-career employment share (effects above a third of a percentage point ruled out). By December 2024 adopters' hours in their most recent occupation were about 4% of a full-time equivalent higher than non-adopters', consistent with more occupation switching, and switchers' earnings grew 12 percentage points faster than other workers'. 97.7% of adopters and 99.5% of non-adopters say chatbots have had no impact on their earnings.
- What it does not show
- The design identifies differences between adopters and non-adopters only; any level effect on non-adopters is unidentified, though the authors argue with several checks that it is small. Firms adjusting in anticipation of AI would bias estimates toward zero. The survey records no exact adoption date, employer initiatives are as workers perceive them, outcomes end in December 2024, only 11 highly exposed occupations in Denmark are covered, and occupation-level estimates are wide. The authors conjecture, rather than show, that stable pay and hours coexist with task reorganization that statistics do not record.
- Status
- A working paper: not yet peer reviewed
The record
- Questions
- What happens to work, income and the economy?
- People
- Anders Humlum (Wrote it, University of Chicago Booth School of Business and NBER) · Emilie Vestergaard (Wrote it, University of Copenhagen)
- What we read
- We read the whole source.
- Found by
- A research pass: work, income and the economy, gathered October 13, 2026, via
research-work(tier 1: the source’s own channel or an official record) - Cite this
- NBER, “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI”, March 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/nber-2026-03-ai-chatbots-denmark-earnings-hours
Our notesText saved by the research pass: the NBER page gives only the abstract, so the full March 2026 revision text was used. Main text read in full; the online appendix was skimmed. Previously circulated as "Large Language Models, Small Labor Market Effects". Quotations are attributed to the first author on behalf of both. The paper itself says working papers are not peer-reviewed. Its acknowledgements credit Claude Code (Opus 4.6) with research assistance, and a footnote cites a forecast by Anthropic's CEO as one pole of the debate; the entry is not about Anthropic or Claude, so conflict is left false for the editor to decide. Found no earlier Base entry that this clearly bears on.