Danish study finds AI chatbots barely moved wages or hours despite time savings

Economists Anders Humlum and Emilie Vestergaard published a study linking two large-scale adoption surveys — covering roughly 25,000 workers across 7,000 workplaces in 11 occupations exposed to AI chatbots, from call-centre agents to journalists — to Danish administrative payroll records. The design let them compare adopters and non-adopters within the same occupation and workplace over two years, rather than relying on self-reported productivity claims.

The headline finding was a “precise null.” Workers who adopted chatbots such as ChatGPT reported saving time — on average around 3% of their working hours — but this did not show up as higher earnings or fewer recorded hours at either the worker or workplace level. The confidence intervals were tight enough to rule out effects larger than about 1–2% in either direction, which the authors said was informative in itself: it meant the null was not simply a product of noisy data.

The paper did find adoption associated with some reorganisation of work — task-switching and changing job content — without this reorganisation translating into pay. The authors argued the results were best read not as evidence that generative AI had no economic effect, but that early adoption was being absorbed into existing jobs faster than it was showing up in wages, a pattern they later described, in a revised version of the paper, as “still waters, rapid currents.”

The study became one of the most frequently cited empirical counterpoints to predictions of imminent labour-market disruption from large language models, precisely because it used administrative records rather than surveys of expectations, and because Denmark’s high AI-chatbot adoption rates gave it an unusually large sample of exposed workers to study.