Does AI chatbot adoption change worker pay and hours?

Humlum and Vestergaard link large-scale adoption surveys to administrative labor market records in Denmark and find that the early effects of AI chatbots show up in the structure of work rather than in pay or hours. Most employers in exposed occupations have adopted chatbot initiatives, workers report productivity benefits, and new AI-related tasks are widespread. The authors call this the "rapid currents," and say these "have not broken the surface": difference-in-differences estimates give "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."

The mechanism the excerpt gives is task reorganization. Employers absorb AI through new tasks in content generation, AI oversight and AI integration, and workers who adopt it "transition into higher-paying occupations where AI chatbots are more relevant." The authors are explicit that this movement is "still too few to move average earnings," so the pay effect is present at the margin but absent from the aggregate. The closing sentence makes the timing claim: "Technological change reshapes work well before it surfaces in earnings or hours." The two-year window is what makes that a claim about sequence. The reorganization is observed inside the window, and the matching movement in earnings and hours is not.

This sits against the Upwork study in the library, which ties ChatGPT's release to lower employment and earnings for freelancers in highly affected occupations, with monthly earnings down 5.2% Did ChatGPT's release reduce freelance writing work and pay?. The two studies do not measure the same population. Upwork covers freelancers on one platform, while the Danish records cover workers in exposed occupations generally, and the outcomes differ in places. The contrast is still informative, because a 5.2% earnings fall lies outside the range the Danish estimates rule out. Scope, not a verdict on either study, decides which description applies to a given market. The change in task structure parallels the trace data showing heavy generative AI users shifting toward documentation-focused work Does generative AI shift knowledge workers away from communication?. The Danish excerpt reaches a similar place through employer surveys rather than application traces, so the two agree on where the first change happens without sharing a measurement. The occupational sorting the Danish authors describe is the same kind of movement that the skill-demand data show as a divergence Is AI creating common skills across jobs or deepening divisions?.

The excerpt is an abstract, so it does not establish the sample size, the exposure measure that defines "exposed occupations," which occupations drive the transitions to higher pay, or whether the nulls hold beyond two years. The 2% figure is an upper bound on the effect, not a finding of zero, and it covers earnings and recorded hours rather than employment counts. The excerpt also does not say whether the new tasks pay as well as the work they replace. The defensible reading is narrow: in Denmark, in the first two years after ChatGPT, earnings and hours for workers in exposed occupations did not move by more than 2%, and the reorganization of tasks came first. Whether that lag widens into a pay effect is open, and the abstract cannot settle it.