Danish data found wide chatbot use but no pay effect yet

Linking a national survey to tax and payroll records

Economists Anders Humlum and Emilie Vestergaard's NBER working paper, issued in May 2025, takes a different approach from a workplace case study: it links large-scale, representative surveys of AI chatbot adoption, run with Statistics Denmark, to that country's administrative labour-market records. The full paper covers eleven occupations judged highly exposed to chatbots, including accountants, customer support specialists, IT support staff, journalists and software developers, with their latest survey round, from late 2024, drawing responses from 25,000 workers across 7,000 workplaces. Because adoption varied even among similar workplaces in the same occupation, the authors could compare otherwise-alike adopters and non-adopters using a difference-in-differences design.

Widespread use, reorganised tasks, and a precise null

Adoption is high: about 40% of workers in occupations with no employer AI policy had already used a chatbot at work, rising to 93% where employers combined encouragement, enterprise tools and training, with 19% of that group reporting saving more than an hour a day. Most chatbot users say they reallocated saved time to other tasks rather than working less, and a meaningful share, up to 17% in the most supportive workplaces, report taking on entirely new tasks such as reviewing AI output or setting usage policy. Despite that visible reorganisation of work, the paper's difference-in-differences estimates find no detectable effect on earnings or recorded hours at either the worker or workplace level, precise enough to rule out effects larger than 2%, two years after ChatGPT's public launch.

Why a null result needs its own caution

A null finding this precise is still bounded by what the data can see: Danish administrative records capture earnings and hours, not the within-job reallocation of tasks the survey documents directly, so the paper's own framing is that transformation is happening “beneath the surface” before it reaches pay. Two years is also a short window against typical wage-adjustment timelines, and the occupations studied are the ones judged most exposed already, not a random cross-section of the labour market. The result should be read alongside, not instead of, task-level experimental findings such as the NBER paper on generative AI in customer support, which measures a real productivity gain inside one workflow without saying anything about whether it shows up in that worker's pay.

Questions to ask before you adopt it

• Is a claimed AI effect measured in pay and hours, or in task-level output that has not yet reached compensation?
• How long after adoption was the outcome measured, relative to how quickly wages typically adjust in your setting?
• Does “no effect on earnings” mean “no effect,” or that any effect is still too new or too small for this data to detect?

Large adoption with a precisely measured null effect on pay is a genuinely different kind of evidence from a task-level experiment, and both can be true at once: work can be changing task by task while the paycheque has not moved yet.