The idea that AI will dramatically expand productivity over the next 10 years is a foregone conclusion for many tech leaders and analysts.
It isn’t for Daron Acemoglu. A professor of economics at Massachusetts Institute of Technology, Acemoglu studies everything from why nations fail to the impact of robots and AI on jobs. He estimates that AI will have a measurable impact on productivity and GDP over the next decade, but it will be significantly smaller than many have predicted.
We spoke with Acemoglu about what his forecasts mean for businesses looking to adopt the technology, and his predictions for how AI will impact jobs by 2030. Here are highlights from that conversation, edited for length and clarity:
Given your prediction that AI will have a smaller impact on productivity and GDP over the next 10 years, what do you think businesses should be doing today?
Wait and see. I think right now business leaders, in the United States at least, are under tremendous pressure to do something on AI just to show it. They are being told by tech leaders, by tech journalists, by all sorts of advertisements, by all management consulting companies—many of which have reinvented themselves as AI integrators—that if you don’t make a big AI investment now, you’re falling behind your competitors.
So the default for many of them is going to be to do something just so that they’re not the odd one out. It’s not going to be conducive to deliberative, intelligent investment. Often they’re not going to know how to use AI. Often they’re not organizationally ready to do that. Often they’re not going to invest enough in training their employees to work with AI. Of course, there are some companies that can benefit from AI and they are already using it—for example, social media companies, for better or worse. But for most businesses, the case for huge investments in AI is not clear. I think the case for tracking it, understanding it, and planning for later investments is much clearer.
What do you think the impact of AI will be on the quality and quantity of jobs?
Hard to tell. First of all, I don’t think AI is like the weather. It’s not something we should just forecast, because it’s partly determined by our own choices—for example, whether business managers are going to excessively invest right away or take a more deliberative approach. It is determined by the choices of tech companies of what type of AI to develop, unions’ choices of how to work with the tech industry, and the government’s choices on regulation.
But in either case, I don’t think we’re going to see a massive effect on the quantity of jobs within the next five years or so. So if I were to make a prediction—which is, as all predictions are, subject to uncertainty, especially in this fast-changing area—I expect the 2030 job market will not be massively different from today in terms of the occupations, things that people do, and the types of jobs that they’re looking for.
I don’t think that you’re going to see any occupation that we have today disappear by 2030. There might be some job losses. IT security is one area where I think we’ll see much less employment in five, six years time, because those are actually simple tasks that AI can do.
The quality of jobs is the key. If AI is used right, in the way that calculators are, or in the way that we use internet-based platforms for creating better linkages between businesses, we could create higher-quality jobs. If, on the other hand, we go crazy on automation and we displace a lot of workers and then they have to go and all work in the fast food industry, that would of course lead to massively lower-quality jobs.
Do you believe that no occupations will disappear in five years due to limitations of the technology itself, or because of how long it will take companies to implement it?
Both. I don’t think that current AI can replace financial analysts or paralegals. Certainly not doctors, certainly not nurses, certainly not journalists. But it is possible that some of these occupations will be thinned in 15 years, even if there are no breakthroughs. But when you’re talking about beyond the next seven, eight, 10 years, many different models, many different architectures, many different types of AI products might come up. I think there is no hard limit to cognition by AI. I just think there’s a hard limit to cognition by AI based on the current approach, which is not super good.
Are there occupations you think will grow over the coming years?
That’s exactly the kind of thing which will really depend on how we use AI. There’s a way to use AI to complement educators, and we need better education. So I could see a good scenario where we have a much higher-quality education system and we have more educators. Do I think that’s likely? Absolutely not. In fact, AI companies are doing exactly the opposite. They’re trying to replace educators and they’re trying to put products in the hands of parents or students, bypassing teachers. So yes, it is possible that education could be a growth sector and with very good social benefits—or the opposite is possible. Nurses could be a growth occupation. But all of those will depend on how we use AI.
You’ve written about what you call ‘so-so technologies’…
Yeah, my concerns just a second ago were all about so-so technologies, I just didn’t use the term. Imagine that we keep on telling business leaders that you’ll be fired because you’re falling behind your competitors if you don’t make a multimillion dollar investment in AI in the next year. They will invest in AI. And what will they do with it? Well, the easiest thing to do with AI is to automate. You can automate your customer service, you can automate your financial analysts. You can automate your accountants, auditors, IT people.
In all of these cases, it will be so-so automation. That means it will more or less work—the company won’t collapse—but it won’t get productivity improvements. In some cases, it will get deterioration.
One of the simplest occupations in some sense is accounting. But if you think about it, an accountant does so many different things, including a lot of social interaction with people. If instead of company accountants, you put in a souped-up version of TurboTax, that just wouldn’t work. That company won’t go bankrupt, but it will create more problems for the higher-level managers because things will get stuck. You’ll get hallucination-like behavior because you’re asking too much of these models. That’s the so-so automation I’m talking about.
Erik [Brynjolfsson’s] study shows how you can slightly improve customer service with AI, but most companies are actually going to use AI to eliminate customer-service agents. That’s going to be terrible for many customers, and some customers may walk away from your company because the service quality has declined so much.
Read a full transcript of the interview.
Read more of Charter’s coverage of how organizations should be approaching AI.
Read our case studies of how companies and workers are getting AI right.