What do we know about the economics of AI?
What new tasks will generative AI bring to humans?” Acemoglu asks. “I don’t think we know that yet, and that’s the question. What applications will really change the way we do things?”
What are the measurable effects of AI?
Since 1947, U.S. GDP has grown by an average of about 3% per year, and productivity has grown by about 2% per year. Some forecasts claim that AI will double that growth, or at least create a higher-than-usual growth trajectory. In contrast, in a paper published in the August issue of Economic Policy, “The Simple Macroeconomics of Artificial Intelligence,” Acemoglu estimates that AI will increase GDP by “modestly” between 1.1% and 1.6% over the next decade, and productivity by about 0.05% per year.
Acemoglu based his assessment on recent estimates of the number of jobs impacted by AI, including a 2023 study by researchers at OpenAI, OpenResearch, and the University of Pennsylvania, which found that about 20% of U.S. jobs could be affected byAI capabilities. A 2024 study by researchers at MIT’s Center for the Future of Technology, the Productivity Institute, and IBM found that about 23% of computer vision tasks that could eventually be automated could be profitable over the next decade. Still more studies have put the average cost savings from AI at about 27%.
When it comes to productivity, “I don’t think we should underestimate a 0.5% increase over 10 years. It’s better than zero,” Acemoglu said. “But it’s disappointing compared to the promises made by people in the industry and in the tech press.”
To be sure, this is just an estimate, and many more AIapplications are likely: As Acemoglu wrote in his paper, his calculations did not include using AI to predict the shapes of proteins, for which other academics subsequently won a Nobel Prize in October.
Other observers think that “reallocation” of workers displaced by AI will generate additional growth and productivity beyond Acemoglu’s estimates, though he thinks it’s not significant. “Starting from the actual distribution we have, reallocation generally yields only small benefits,” Acemoglu says. “The immediate benefits are what matter.”
“I tried to write the paper in a very transparent way about what was included and what was not included,” he adds. “People can object and say that what I excluded was important or that the numbers for what I included were too low, and that’s totally fine.”
As Acemoglu and Johnson make clear, they favor technological innovations that increase worker productivity while keeping people employed, which should do a better job of sustaining economic growth.
But in Acemoglu’s view, the point of generative AI is to mimic humans as a whole. This produces what he has for years called “so-so technology,” applications that perform at best only slightly better than humans but save companies money. Call center automation isn’t always more efficient than humans; it just saves companies less money than workers do. AI applications that supplement workers seem generally to take a backseat to big tech companies.
“I don’t think complementary uses for AI will magically emerge unless industry invests a lot of effort and time,” Acemoglu said.
What does history teach us about AI?
The fact that technology is often designed to replace workers is the focus of another recent paper by Acemoglu and Johnson, “Learning from Ricardo and Thompson: Machinery and Labor in the Early Industrial Revolution—and in the Age of AI,” published in the August issue of the Annual Review of Economics.
The article discusses the current debate over AI, particularly the claim that even if technology replaces workers, the resulting growth will almost inevitably benefit society over time. Britain during the Industrial Revolution is sometimes cited as an example. But Acemoglu and Johnson argue that spreading the benefits of technology is not easy. In 19th-century Britain, they assert, it happened only after decades of social struggle and workers’ action.
What is the optimal pace of innovation?
If technology helps promote economic growth, then rapid innovation would seem ideal because it would bring growth faster. But in another paper, “Regulating Transformative Technologies,” in the September issue of the American Economic Review: Insights, Acemoglu and MIT doctoral student Todd Lensman offer another view. If some technologies have both benefits and disadvantages, then it is better to adopt them at a more measured pace while mitigating those problems.
“If the social harms are large and proportional to the productivity of the new technology, then higher growth rates will lead to slower adoption,” the authors write in the paper. Their model suggests that, ideally, adoption should start out slow and then gradually speed up over time.
“Market fundamentalism and technology fundamentalism might claim that you should always develop technology at the fastest pace,” Acemoglu says. “I don’t think there is such a rule in economics. More thoughtfulness, especially about avoiding harms and pitfalls, is warranted.”
The model is a response to trends over the past decade or so, in which many technologies were hyped as inevitable and welcomed for their disruptive nature. In contrast, Acemoglu and Lensman suggest that we can reasonably judge the trade-offs involved with a particular technology, and aim to stimulate more discussion about this.
How can we get to the right pace forAIadoption?
If the idea is to adopt technology more gradually, how should that be achieved?
First, Acemoglu said, “government regulation has a role to play.” However, it’s not clear what type of long-term guidelines for AI the U.S. or countries around the world might adopt.
Second, he added, if the “hype” cycle around AI abates, then the rush to use AI “will naturally slow down.” This scenario might be more likely than regulation if AI doesn’t soon turn a profit for companies.
“We’re moving so fast because of the hype from venture capitalists and other investors because they think we’re going to get closer to general AI” Acemoglu says. “I think that hype has caused us to invest improperly in the technology, and a lot of companies have been affected prematurely and don’t know what to do with it. We wrote that paper to say, look, if we’re more thoughtful and understanding about our use of this technology, its macroeconomics will benefit us.”
In that sense, Acemoglu emphasizes that hype is a tangible aspect of the economics of AI, because it drives investment in specific AI visions and thus influences the AI tools we’re likely to encounter.
“The faster the speed and the more excitement, the less likely you are to make a course correction,” Acemoglu says. “If you’re going 200 miles an hour, it’s very difficult to make a 180-degree turn.
