Artificial Intelligence is going to destroy the world, or humanity at least. In the meantime, it’s solving unsolvable math problems, possibly with dirty tricks; spying on activists; and stealing our jobs. At least, that’s what recent headlines tell us. But the actual research on the real-world adoption of the myriad tools we lump under the AI label, especially large language models like Claude or ChatGPT, has been slim. That’s slowly beginning to change.
The Federal Reserve Bank of St. Louis, with a joint grant from the Washington Center for Equitable Growth and the Russell Sage Foundation, has been digging deeper into who is using AI, in what fields, and for what tasks, and released the results from their latest survey this month. What they found complicates the picture of how AI is being integrated into the workplace—rapidly, yes, but not yet deeply. How this impacts workers going forward is just a guess at this point. But what’s clear is that there’s a yawning gap between Big Tech’s promise of a technological revolution and regular people’s interest in using that technology.
The St. Louis Fed’s research surveyed workers in a range of occupations about their work lives, and found that the use of AI programs—ChatGPT is the most common—is surprisingly widespread for such a new technology. The survey found that in 75 percent of occupations, at least a fifth of the workforce used AI. “That’s kind of just a wider distribution than we were expecting,” Adam Blandin, a St. Louis Fed research fellow and an assistant professor of economics at Vanderbilt University, told me. “Everybody expects computer programmers or executives or something like that to use it, but it’s just more widely spread across occupations than we realized.”
But even though adoption is fairly widespread, most people don’t use it for huge chunks of their jobs. They used it for just a few tasks. And there was almost no task defined by the survey that had been overwhelmingly taken over by AI. Even the tasks for which it’s most used for—reading documents to gather technical information, preparing research reports, and analyzing data to identify trends—had adoption rates of about 60 percent. Different workers use it for different parts of their jobs. Even people in the same jobs might not both use AI, or use it at the same rates.
There were some patterns in the use of AI that made intuitive sense. Computer programs use AI often, for example, and people who use AI often for their job are more likely to use it outside the workplace, in everyday life. Some of the differences in uptake could come down to individual preferences, worksite practices, or something else completely unknown. “Understanding those personal differences, I think, is underexplored,” Blandin said.
Some past predictions about how AI would be used were a bit off the mark. Many experts had thought that health care professionals might use AI to help them diagnose conditions and come up with treatment plans. But that adoption hasn’t been widespread, which could be because health care workers don’t trust AI with sensitive information, or don’t trust the work it does. At the same time, some very hands-on jobs, like testing materials for environmental hazards or repairing machines, have seen higher uptake than expected. These jobs tend to rely on manual labor, but workers are using AI for research on what tools to use or how best to make a repair.
The researchers asked workers who said they use AI how much longer their tasks would have taken without it, and estimated that use of these tools increased productivity by about 5 percent. There is a huge caveat to that number, of course, because it’s self-reported. But it suggests that workers are using AI because it makes them marginally more productive. That doesn’t necessarily mean AI can do their entire job and replace them. In fact, in the recent past, when workers adopted computers, jobs that relied on computers became more in demand and the people who used computers at their jobs started to earn more than those who didn’t. If AI proves to be similarly useful for workers, but not as useful as workers, it could do the same.
One big difference, though, is that there was a huge gap between jobs that required workers to learn to use computers and those that didn’t, but today, more workers across more fields are using AI, so there isn’t much of a divide. That could potentially mean that everyone’s job becomes a little more in demand and maybe makes everyone a little more productive in countless ways. “There’s a bunch of occupations where adoption is kind of in the middle,” Blandin said. “That again just suggests it may not be as concentrated as we saw with computers, and so maybe more people will benefit from AI. I should just be clear: That’s speculation. But that’s a possibility.”
Of course, as computers were adopted through the 1980s and ’90s, even many of the knowledge workers and workers in other fields who adopted them to become more productive didn’t always see their increased productivity turn into more pay. How much people make in their jobs, and how protected their jobs are, isn’t something that happens without intervention like union contracts and new employment laws to help workplaces and people adapt to changes. An AI industry that is forcing its technology—and data centers—on society without our input, while also warning of a potential apocalypse caused by that technology, certainly can’t be trusted to act in workers’ best interest.