- AI chatbots have been rolled out across hundreds of white-collar workplaces, but on average, their effect on hours and pay has been negligible, according to a National Bureau of Economic Research working paper linking AI use to corporate records in Denmark. On average, employees saved 3% of their time, while just 3%-7% of their productivity gains came back to them in the form of higher pay.
Since OpenAI rolled out ChatGPT just over two years ago, AI chatbots have become the fastest-adopted technologies in history, rivaling the PC three decades ago. Their popularity has created and destroyed entire job descriptions and sent company valuations into the stratosphere—then back down to earth.
And yet, one of the first studies to look at AI use in conjunction with employment data finds the technology’s effect on time and money to be negligible.
“AI chatbots have had no significant impact on earnings or recorded hours in any occupation,” economists Anders Humlum and Emilie Vestergaard wrote in a National Bureau of Economic Research working paper released this week.
Humlum, an assistant professor of economics at the University of Chicago’s Booth School of Business, and Emilie Vestergaard, an economics PhD student at the University of Copenhagen, looked at 25,000 workers across 7,000 workspaces, focusing on occupations believed to be susceptible to disruption by AI: accountants, customer support specialists, financial advisors, HR professionals, IT support specialists, journalists, legal professionals, marketing professionals, office clerks, software developers, and teachers.
They pulled records from Denmark, a country whose rates of AI adoption as well as hiring and firing practices are similar to those in the U.S. But where record-keeping is far more detailed, allowing the study to anonymously match survey responses to records of actual hours and pay.
On average, users of AI at work had a time savings of 3%, the researchers found. Some saved more time, but didn’t see better pay, with just 3%-7% of productivity gains being passed on to paychecks.
In other words, while they found no mass displacement of human workers, neither did they see transformed productivity or hefty raises for AI-wielding superworkers.
“While adoption has been rapid, with firms now heavily invested in unlocking the technological potential, the economic impacts remain small,” the authors write.
Productivity, interrupted
The findings might be a surprise against the backdrop of aggressive corporate adoption of AI: from Duolingo replacing its contract workers with AI to Shopify decreeing it will only hire humans as a second choice to AI. Meanwhile, investors have been bidding up shares of companies involved in AI.
But the NBER paper doesn’t mean that earlier findings of AI’s productivity boost have been wrong, said Humlum—just incomplete.
Most of the earlier research has focused “exactly on the occupations where the time savings are largest,” Humlum told Coins2Day.
“Software, writing code, writing marketing tasks, writing job posts for HR professionals—these are the tasks the AI can speed up. But in a broader occupational survey, where AI can still be helpful, we see much smaller savings,” he said.
Other factors that explain AI’s overall ho-hum impact include employer buy-in and employees’ own time management.
“I might save time drafting an email using a large language model, so I save some time there, but the important question is, what do I use that time savings for?” He said. “Is the marginal task I’m shifting my work toward a productive task?”
Workers in the study allocated more than 80% of their saved time to other work tasks (less than 10% said they took more breaks or leisure time), including new tasks created by the use of AI, such as editing AI-generated copy, or, in Humlum’s own case, adjusting exams to make sure that students aren’t using AI to cheat.
There’s also the fact that real workplaces are much messier than structured experiments.
“In the real world, many workers are using these tools without even the endorsement of the boss. Some don’t even know if they’re allowed to use it; some are allowed but not really encouraged to use it,” Humlum said. “In a workplace where it’s not explicitly encouraged, there’s limited space to go to your boss and say, ‘I’d like to take on more work because AI has made me more productive,’” let alone negotiate for higher pay based on higher productivity.
And of course, employees might not want to advertise how much more productive AI has made them, especially considering the well-trod adage that the reward for efficient workers is more work.
Some of the findings around hours and pay in workplaces where AI isn’t used “suggest that workers are not exactly knocking on the boss’s door asking for more work,” Humlum said.
Great expectations, mid results
The NBER paper comes on the heels of other indications suggesting that AI’s potential, while tremendous, has been vastly overstated in the media and the market.
Payment processor Klarna, which made waves last year when it revealed it stopped hiring humans in favor of a super-productive AI, recently tempered its rhetoric.
An IBM survey of 2,000 CEOs revealed that just 25% of AI projects deliver on their promised return on investment. The main driver of adoption, it seems, is corporate FOMO, with nearly two-thirds of CEOs agreeing that “the risk of falling behind drives them to invest in some technologies before they have a clear understanding of the value they bring to the organization,” according to the study.
Nobel laureate Daron Acemoglu, who has extensively researched automation and labor, estimates AI’s productivity boost at approximately 1.1% to 1.6% of GDP in the next decade—a sizable boost for an advanced economy like the U.S., but far from the doubling of GDP some technologists have predicted.
The danger with AI is that “the hype will likely go on for a while and do much more damage in the process than experts are anticipating,” he wrote for Coins2Day last year. In fact, “getting productivity gains from any technology requires organizational adjustment, a range of complementary investments, and improvements in worker skills, via training and on-the-job learning,” he said.
That’s a finding backed up by Humlum and Vestegaard, whose paper showed greater productivity gains when employers encouraged AI use and trained workers in it.
It could also be just a matter of time. After all, the Industrial Revolution went on for a century, transforming how people lived and worked long after the invention of the steam engine.
“It took a couple decades to see that we can have an assembly line powered by electricity instead of having everything run centrally via a steam engine,” Humlum said.
