Instead of depressing wages or taking jobs, generative AI chatbots like ChatGPT, Claude, and Gemini have had almost no wage or labor impact so far – a finding that calls into question the huge capital expenditures required to create and run AI models.
In a working paper released earlier this month, economists Anders Humlum and Emilie Vestergaard looked at the labor market impact of AI chatbots on 11 occupations, covering 25,000 workers and 7,000 workplaces in Denmark in 2023 and 2024.
Many of these occupations have been described as being vulnerable to AI: accountants, customer support specialists, financial advisors, HR professionals, IT support specialists, journalists, legal professionals, marketing professionals, office clerks, software developers, and teachers.
Yet after Humlum, assistant professor of economics at the Booth School of Business, University of Chicago, and Vestergaard, a PhD student at the University of Copenhagen, analyzed the data, they found the labor and wage impact of chatbots to be minimal.
AI chatbots have had no significant impact on earnings or recorded hours in any occupation
“AI chatbots have had no significant impact on earnings or recorded hours in any occupation,” the authors state in their paper.
The report should concern the tech industry, which has hyped AI’s economic potential while plowing billions into infrastructure meant to support it. Early this year, OpenAI admitted that it loses money per query even on its most expensive enterprise SKU, while companies like Microsoft and Amazon are starting to pull back on their AI infrastructure spending in light of low business adoption past a few pilots.
The problem isn’t that workers are avoiding generative AI chatbots – quite the contrary. But they simply aren’t yet equating to actual economic benefits.
The adoption of these chatbots has been remarkably fast … But then when we look at the economic outcomes, it really has not moved the needle
“The adoption of these chatbots has been remarkably fast,” Humlum told The Register. “Most workers in the exposed occupations have now adopted these chatbots. Employers are also shifting gears and actively encouraging it. But then when we look at the economic outcomes, it really has not moved the needle.”
The researchers looked at the extent to which company investment in AI has contributed to worker adoption of AI tools, and also how chatbot adoption affected workplace processes.
While firm-led investment in AI boosted the adoption of AI tools — saving time for 64 to 90 percent of users across the studied occupations — chatbots had a mixed impact on work quality and satisfaction.
The economists found for example that “AI chatbots have created new job tasks for 8.4 percent of workers, including some who do not use the tools themselves.”
In other words, AI is creating new work that cancels out some potential time savings from using AI in the first place.
“One very stark example that it’s close to home for me is there are a lot of teachers who now say they spend time trying to detect whether their students are using ChatGPT to cheat on their homework,” explained Humlum.
He also observed that a lot of workers now say they’re spending time reviewing the quality of AI output or writing prompts.
Humlum argues that can be spun negatively, as a subtraction from potential productivity gains, or more positively, in the sense that automation tools historically have tended to generate more demand for workers in other tasks.
“These new job tasks create new demand for workers, which may boost their wages, if these are more high value added tasks,” he said.
But overall, the time savings from using AI was less than expected. According to the study, “users report average time savings of just 2.8 percent of work hours” from using AI tools. That’s a bit more than one hour per 40 hour work week.
The authors note that this finding differs from other randomized controlled trials that have found productivity benefits on the order of 15 percent. And they explain this discrepancy by saying that other studies have focused on occupations with high AI productivity potential and that real-world workers don’t operate under the same conditions.
“So I think there are two key reasons why the real economic gains are lower than [the cited studies],” said Humlum, noting that his study relies on actual tax data.
“First, most tasks do not fall into that category where ChatGPT can just automate everything. And then second, we’re in this middle phase where employers are still waking up to the new reality, and we’re trying to figure out how to best really realize the potential in these tools. And just at this stage, it’s just not been that much of a game changer.”
Where there are productivity gains to be had, Humlum and Vestergaard estimate that only a small portion of that benefit – between 3 and 7 percent – gets passed through to workers in the form of higher earnings.
Humlum said while there are gains and time savings to be had, “there’s definitely a question of who they really accrue to. And some of it could be the firms – we cannot directly look at firm profitability. Some of it could also just be that you save some time on existing tasks, but you’re not really able to expand your output and therefore earn more.
“So it’s like it saves you time writing emails. But if you cannot really take on more work or do something else that is really valuable, then that will put a damper on how much we should actually expect those time savings to affect your earning ability, your total hours, your wages.”
Humlum said the impact of using AI chatbots, in the form of productivity, time savings, and work quality, can be improved through company commitment to internal education and evangelism. He pointed in particular to how firm initiatives can reduce the tool-usage gender gap – fewer women use these tools than men.
But doing so at this point doesn’t show much promise of payoff.
“In terms of economic outcomes, when we’re looking at hard metrics – in the administrative labor market data on earnings, wages – these tools have really not made a difference so far,” said Humlum. “So I think that that puts in some sense an upper bound on what return we should expect from these tools, at least in the short run.
“My general conclusion is that any story that you want to tell about these tools being very transformative, needs to contend with the fact that at least two years after [the introduction of AI chatbots], they’ve not made a difference for economic outcomes.” ®
