AI chatbots and earnings: Denmark study of 25,000 workers finds about 3% time savings and no earnings effect

Verification status: IN CHECKING. Not publish-ready, not pending, not verified.

The problem

AI chatbots spread through white-collar workplaces faster than almost any earlier technology, but whether that adoption shows up in pay and hours had barely been measured against administrative records [source]. Humlum and Vestergaard set out to study “the early labor market impacts of AI chatbots by linking large-scale adoption surveys to administrative labor market records in Denmark” ([source]). For a COO budgeting a company-wide chatbot rollout, the question is the gap between what workers say they save and what registers record [source].

What was built

This is a research finding, not a product deployment: NBER Working Paper 33777 by Anders Humlum of the University of Chicago Booth School of Business and Emilie Vestergaard of the University of Copenhagen, first issued in May 2025 and revised in March 2026 under the new title “Still Waters, Rapid Currents” [source]. It was first circulated as “Large Language Models, Small Labor Market Effects” ([source]). The surveys cover 11 exposed occupations: “accountants, customer support specialists, financial advisors, HR professionals, IT support specialists, journalists, legal professionals, marketing professionals, office clerks, software developers, and teachers” [source]. The latest round, “conducted in late 2024, includes responses from 25,000 workers across 7,000 workplaces” [source]. Responses are linked to Statistics Denmark’s Employment Statistics of Employees, “which records earnings, hours, occupation, and industry” ([source]). Ars Technica reported the same scale, “25,000 workers and 7,000 workplaces in Denmark” [source].

The outcome

Adoption is broad: “about 43% explicitly encourage their use, another 21% allow it, while only about 6% explicitly prohibit it” [source]. Workers say they save time, but modestly: “On average, adopters in our sample report savings of about 3% of their work hours” [source]. The May 2025 version put the same figure at “2.8% of the total work hours of users” ([source]), which Ars Technica reported as “just 2.8 percent of work hours (about an hour per week)” [source]. That number is a survey self-report, computed by the authors from how often workers use chatbots and the minutes they say they save, not a register measurement ([source]).

The registers show no movement. The authors “estimate precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2% two years after the launch of ChatGPT” [source]. The May 2025 version was tighter, “ruling out effects larger than 1%” ([source]); this page leads with the current revision and prints both rather than merging them. Fortune quoted the paper’s original conclusion that “AI chatbots have had no significant impact on earnings or recorded hours in any occupation” [source]. Where the saved time goes: “most chatbot users (85%) report reallocating time savings from AI chatbots to other job tasks” in the revision [source], against “the vast majority (80%)” in the May 2025 version ([source]).

The weakest points are stated plainly here: this is an NBER working paper that, by NBER’s own notice, has “not been peer-reviewed”, and the press coverage that independently corroborates it reported the May 2025 numbers, not the March 2026 revision [source]. The employer-policy shares and the reallocation shares come only from the authors, with no independent outlet carrying them ([source]).

Related case files

• The Bank of Korea’s finding that AI cut work time 3.8% with near-zero productivity gain, the same hours-saved versus output gap on a national survey
• The MIT ChatGPT writing experiment that found 40% faster work, the kind of task-level gain this register study does not see in pay
• Three field experiments on software developers with 26% more tasks completed, a large controlled gain in one of the occupations surveyed here
• The BCG and Harvard jagged-frontier study of GPT-4 consultants, a task-level result to read against an economy-wide null