Danish Records Show No Pay Effect From Two Years of AI Chatbots

Workers in Denmark's most AI-exposed occupations have seen no measurable change in earnings or working hours in the two years since ChatGPT launched. That is the central finding of research by Anders Humlum of the University of Chicago and Emilie Vestergaard of the University of Copenhagen, published by the Brookings Institution on 6 October 2026, which links surveys of 25,000 Danish workers to monthly government administrative records running through December 2024.

The result is not an absence of evidence. The authors report what economists call a precise null: their difference-in-differences estimates rule out any differential change in earnings larger than 2 per cent, at both the worker and the workplace level. The underlying study is National Bureau of Economic Research working paper 33777, first issued in May 2025 and revised in March 2026. It circulated earlier under the blunter title "Large Language Models, Small Labor Market Effects".

Eleven occupations were surveyed: software developers, IT support specialists, customer support specialists, office clerks, accountants, financial advisors, HR professionals, legal professionals, marketing professionals, journalists and teachers. Statistics Denmark issued the invitations through the official digital mailbox that every Danish resident holds, 115,000 of them in each of two rounds in late 2023 and late 2024. The 2024 round, which carries the main findings, returned about 25,000 valid responses across 7,000 workplaces.

Adoption that did not wait for permission

43 per cent of workers in those 11 occupations have employers that explicitly encourage chatbot use, against 6 per cent who are prohibited from using them, the 2024 survey found. Bans survive mainly where the data is confidential or the output is legally load-bearing, among financial advisors and legal professionals. The more striking number is 41 per cent: that share of workers had used chatbots at work even in workplaces that took no position at all.

Employer support still changes the picture. Where encouragement arrives with enterprise tools and training attached, 93 per cent of workers have used chatbots at work, 28 per cent use them daily and 19 per cent report saving more than an hour a day. Among workers at encouraging employers, 61 per cent have access to enterprise chatbots and 39 per cent have been trained on them.

The work changed shape rather than price

Workers described their new tasks in free text, which the authors sorted into six categories. Roughly four in ten of the new tasks involve generating something with AI. About a third involve supervising it, splitting between checking outputs and compliance work. The largest single category, about a quarter, is integration: tuning assistants, writing usage policies and wiring chatbots into existing workflows. Marketing professionals reported "prompting and iterating with AI to produce marketing copy, social media posts, and product descriptions". A large share of teachers reported new work related to "detecting AI-generated homework".

The new work reaches people who never touched the tools. About 4 per cent of non-users reported new workloads created by chatbots, rising to 10 per cent among teachers who have never used one. Of the workers who do use them, 85 per cent spend the time they save on other job tasks rather than on more of the same work or on more leisure.

The one number that moves

Occupational mobility is the exception. Measured by hours worked in the occupation each person held in December 2024, chatbot adopters put in about 4 per cent of a full-time equivalent more than comparable non-adopters, meaning they are more likely to have switched occupations since late 2022. Unlike the raw earnings gap between adopters and non-adopters, which was already there before ChatGPT existed, this one shows no pre-trend. The switchers did well: their earnings growth ran 12 percentage points ahead of other Danish workers. The association triples among daily users.

Where they went is as informative as the fact that they went. Switchers moved mostly into IT support and clerical roles, occupations where workers pick their own tools and no credential bars entry. The authors found no excess transitions into licensed occupations such as teaching and accounting, where years of prior education gate the door regardless of what software a newcomer commands.

What it does to the entry-level argument

Denmark shows the same fall in early-career employment in AI-exposed occupations that Stanford's Digital Economy Lab documented in the United States in 2025, including in software development, legal work and marketing. The Danish registers allow a test the aggregate figures do not: splitting those trends by whether the employer actually adopted the technology. The share of early-career workers evolved no differently at workplaces that encourage chatbot use, with the estimates tight enough to exclude differences larger than a third of a percentage point. Whatever is shrinking entry-level hiring in exposed occupations, the firms using chatbots are not the ones doing it.

That sits against the Dallas Fed research Parallax Nexus reported on 22 September, which found a 1.7-point fall in the chance of employment within a year of graduating for every 10-percentage-point rise in a degree subject's automatable task share. Both can hold at once. Exposure measured by occupation can predict worse graduate outcomes while adoption measured at the firm explains none of it, and that gap is the most useful thing in either paper.

The honest reading is that this is a measurement result before it is an economic one. Denmark's registers record earnings, hours and occupations, and none of those moved; they do not record task content, which did. The limits are real. The data stop in December 2024, which leaves out nearly two years of model releases and the whole agentic turn. Denmark is one small, flexible, unusually high-adoption labour market. And the authors cannot say why pay did not follow the time savings workers report, offering two candidates they are unable to separate: that chatbots act as a job amenity taken in easier work rather than higher pay, or that workers overstate how much the tools help them.

One line in the paper reads as a marker of how fast the ground is moving underneath it. Its acknowledgements credit research assistance to a human, Caspar Ringhof, and to Claude Code.