How artificial intelligence can help American workers

Labor economist David Autor’s research shows how historically, technological advances hurt the incomes of middle- and working-class Americans.

But when it comes to AI, Autor says the exact opposite could happen.

Today, On Point: How artificial intelligence can help American workers.

David Autor, professor of economics at MIT. Codirector of the Labor Studies Program at the National Bureau of Economic Research (NBER). Author of a recent essay in Noema Magazine titled “AI Could Actually Help Rebuild The Middle Class.”

Elise Azara, director of marketing for Zero Emissions Northwest in Spokane, Washington. She says AI helps her at work.

Part I

MEGHNA CHAKRABARTI: This is On Point. I'm Meghna Chakrabarti.

(MONTAGE)

CNN NEWS ANCHOR: Tonight we're taking a closer look at a new technology that's making waves in the world of AI.

FOX NEWS ANCHOR: ChatGPT.

BBC NEWS ANCHOR: ChatGPT.

WALL STREET JOURNAL NEWS ANCHOR: ChatGPT. A state of the art conversational AI model developed by OpenAI.

CHAKRABARTI: Well, since the public release of its demo in November 2022, OpenAI and ChatGPT completely changed the public perception of artificial intelligence in everyday life. And it put one question in the sharpest relief yet: Is generative AI a job killer?

(MONTAGE)

UNIDENTIFED PERSON ON TIKTOK: Here’s how ChatGPT is going to kill jobs.

AL JAZEERA ANCHOR: Let’s talk about AI and whether it’s going to take your job.

UNIDENTIFIED PERSON ON TIKTOK: I think it’s just the start of tech layoffs. Let’s break it down.

CBS NEWS ANCHOR: Artificial intelligence could replace millions of jobs.

UNIDENTIFIED PEOPLE ON TIKTOK: Will AI take your job? / Yeah. / What do you do for work? / I work at an AI company. / Oh nice.

CHAKRABARTI: Well, last March, economists at Goldman Sachs estimated that AI could automate as many as 300 million full time jobs globally by 2030. So at first glance, one might expect a labor economist like David Autor to share those concerns.

He's a professor of economics at MIT and co-director of the Labor Studies Program at the National Bureau of Economic Research. And for years, his research has shown that major steps forward in technology and automation have historically hurt middle and low-income American workers. And as generative AI climbs up the income scale in terms of the work it can do, those job loss and income reducing dangers seem to be firmly moving into white collar jobs, too.

But David Autor's latest research has revealed something completely unexpected. He says AI could offer benefits and even help restore "the middle-skill, middle class heart of the U.S. labor market that has been hollowed out by automation and globalization." He's written that in an essay on Noema magazine called AI Could Actually Help Rebuild the Middle Class.

Professor Autor joins us now. Welcome back to On Point.

DAVID AUTOR: Thank you, Meghna. It's a pleasure to be here.

CHAKRABARTI: So I want to actually start by going back to several years of your prior research about the workplace — or American work and technology. How would you generalize what you've found as each wave of technology and automation, you know, settles into workplaces?

AUTOR: Sure. Thanks for the question. So, it's often, the question is often asked, is this time different? And the answer is yes. Actually every time is different. And the era of computerization, which many would say sort of really began in earnest in the early 1980s when it really reached offices and so on had a kind of a polarizing effect.

Computers were really ideally suited for carrying out tasks, jobs that could be described by a set of formal rules. So the rules that you would use for doing accounting or proofreading and typesetting, but also in repetitive production and operative environments where people are doing skilled work, but following a very well-defined set of procedures.

Traditional computers, pre-AI, are really good at following rules and operating tools. So things that can be codified as a series of steps that a machine that doesn't solve problems, doesn't have novel ideas, doesn't have innovative insights, but that can do that work rapidly and inexpensively, that was very displacing for people in offices and for people in many assembly line jobs.

Now, the effect of that was polarizing because on the one hand, if you were a college-educated worker doing professional or technical or managerial work, a lot of your work is decision making. It's deciding how do I treat a patient, or how do I redesign a building, or how do I architect a piece of software, or how do I rewire a house, or, you know, plumb a new building. That's decision making work. And computerization is really helpful for that, because it provides the information and, and calculation that, that helps support good decisions. It doesn't make the decisions for you, but it provides input. So it made people much more effective in that type of work. So, computerization was really great for college educated workers and for many high paid workers.

However, if you were in the middle group — in the clerical, office, administrative support, production, operative work, that work started disappearing. And if you couldn't move up into these professions, many people found themselves instead in services — food service, cleaning, security, entertainment, recreation, some low-paid home health care. Now, that's socially valuable work. I don't mean to denigrate it at all. But it's poorly paid. And the reason it's poorly paid is because it doesn't require much training or certification or expertise. It kind of uses a relatively generic skill set so that many people of sound mind and body can do that work without a lot of preparation.

CHAKRABARTI: Yeah.

AUTOR: And that means that wages tend to be low. And this is true across the all industrialized economies. Some pay those jobs better, some worse, but they're always at the bottom end of the earnings distribution.

CHAKRABARTI: So let me summarize at least part of your work this way. It wasn't all that long ago I think that you coauthored a paper in the Quarterly Journal of Economics called New Frontiers, right? The Origins and Content of New Work from 1940 to 2018. That was a very interesting paper, Professor Autor.

AUTOR: Thank you.

CHAKRABARTI: Just to sort of quote a couple of things here. One is that the study found that since 1980, that technology replaced more jobs than it generated, right? And that 60% of jobs in 2018 actually never even existed in 1940, which makes a lot of sense due to the changes that you talked about. So there was an overall replacement there that technology did of jobs that outpaced the augmentation, which is the other part of the picture that you talk about that. um, That that augmentation or the ability, I guess, to do more with technology did add some jobs to the economy, but it wasn't as many as that were lost. Is that a fair way to just sort of roughly summarize that paper?

AUTOR: That's correct. But let me let me offer a qualification.

CHAKRABARTI: Okay.

AUTOR: We're not running out of jobs. And so the discussion of the number of jobs, I think, is a little bit — is very alarming to people. But in fact, we're in a  period of sustained labor market scarcity. Our population is growing very slowly. Immigration has been highly restricted. Fertility is really low. And the labor force, the number of adults of working age is basically going to be relatively flat. It's growing at the slowest time in American history. And in fact, in most of the industrialized world, working age populations are declining.

Now you might say, well, why is that a problem? We have fewer people. We need fewer jobs. However, there's a growing share of the population that is past their working age and has earned a well deserved retirement. And those young people are going to have to support them. And that requires work, and that requires productivity. And with shrinking numbers of young people, able-bodied workers, and so on, that actually creates a period of great scarcity. In fact, we're seeing this now in the United States with the extremely tight labor market after the pandemic.

So the question is not the number of jobs. That's not what we should be worried about. It's the quality of those jobs. Do those jobs require expertise like those professional workers? Or do those jobs tend to use generic skill sets, so like those food service, cleaning, security workers? Again, intrinsically valuable jobs, but poorly paid. And that's what we should be focused on, expertise.

And let me give you a concrete example I like to give. Think of the job of air traffic controller and crossing guard. These are basically the same job, right? The job is to prevent things from crashing into one another: planes into planes, cars into children, et cetera. And yet air traffic controllers are paid more than four times what crossing guards are. And again, the reason isn't social value, right? We don't want our children to be run over on the way to school. And if we had to pay crossing guards a lot of money to prevent that from happening, we surely would do so. The difference, again, is expertise. In the U.S., almost any adult of sound mind and body can become a crossing guard with with no training or certification, but to become an air traffic controller requires several years of air traffic control college and then hundreds to thousands of hours of apprenticeship.

So, a world in which everyone is a crossing guard, a world in which we don't have much expert work, is less good for labor than a world in which everyone is an air traffic controller or doing some of their professional activity. So that's what we should be concerned about. Whether jobs will use expert skills that reward knowledge and competency and specialization. Or jobs that require, you know, human labor to do menial stuff.  That's not as good as a scenario.

CHAKRABARTI: Okay. But I'm just going to go down this automation versus augmentation rabbit hole with you a little bit more, because I have to say I'm emerging right now a little bit confused. And then we'll dig into the AI part of your research here. Because I thought, and maybe I was mistaken, but I was under the impression that your research over many years had shown that regarding automation specifically, that overall the effects on employment of automation were overall negative?

AUTOR: The effects on wages.

CHAKRABARTI: On wages, okay.

AUTOR: Exactly. And, I should say, it's not just me. There's a large set of really distinguished researchers who've worked on this problem. I think it's widely agreed that this polarization has occurred as automation has kind of hollowed out the middle and that has put downward pressure on people without college degrees. Not by reducing the number of workers per se, but by pushing workers into non-expert work that's paid poorly. And that's the primary concern. And so it is — that automation, if automation just takes the work that you're good at and allows the machine to do it better, cheaper, and faster, that is good for productivity potentially, but it's not good for you. (LAUGHS) Right?

CHAKRABARTI: Yeah.

AUTOR: And that's the challenge. To give another useful concrete example, I think. Since 1865, London taxi drivers have been required to acquire what's called "the knowledge."

CHAKRABARTI: "The knowledge," yeah.

AUTOR: That's right. And it takes almost three years. And it's a complete memorization of the streets and highways and byways and overpasses and underpasses of London. It's an incredible feat of memorization. Literally brain scans show the changes in people's physiology of their brains coming from this act of memorization. Once smartphone-based routing was introduced, that knowledge was completely economically irrelevant.

CHAKRABARTI: Uh-huh.

AUTOR: Not only does, you know, Waze know all those highways and byways and streets and underpasses, but it also knows how much traffic is on them at any given point, which is something you could never memorize. So, a friend of a friend of mine is a London taxi driver, and now considers himself to be an entertainer and tour guide, not a driver at all.

CHAKRABARTI: So he's trying to find new ways to add value.

AUTOR: Exactly.

Part II

CHAKRABARTI: Some people say AI helps them work smarter, not harder.

ELISE AZARA: We have this kind of inside joke in the office where the boss is like, "Hey, have you tried running this through AI?" You know, like no matter what the task is. And it has worked and it has proven as a helpful tool in some aspects.

CHAKRABARTI: So this is Elise Azara in Spokane, Washington. She's director of marketing for a company called Zero Emissions Northwest, which helps farmers get grant funding for using renewable energy sources.

As you heard Elise say, her boss is pretty big on AI, or at least on testing it out. He hopes it'll make their team more efficient. And Elise says so far she has used it to help write things like press releases and blog articles. There's an AI tool built into their project management software.

AZARA: If I upload a start or like a headline, it'll start to try to draft it based on that. Or for instance, we used a lot of AI prompts to write the website, you know, just to try to give us, "Hey, what would be a better way to paraphrase this page?" It's kind of like having like a friend there that's like, "Hey, have you thought about doing this?" I can bounce ideas off of it. I mean, not to like personify AI. (LAUGHS) But I don't necessarily think my first idea is the best. I always want to kind of see what else could be out there. So in that sense, yes, it's very helpful.

CHAKRABARTI: Also in that sense, she's doing what she would have usually done with other human beings, but doing it with AI saves her a lot of time. That's time she can now use to work on other things or spend with her two young children. But the AI's writing is not perfect. Elise still has to edit it. And sometimes trying to incorporate AI has actually created more work for her. She says one time her boss wanted to pull data from a bunch of different utility bills, and he spent hours uploading the documents.

AZARA: He's copying and pasting, you know, the utility bills, or he's trying to convert it from a PDF to a Word or whatever, so that way he can get the data in there. And then he's asking the AI, he's prompting it, saying, "Hey, I need total kilowatt hour usage," and this and that. And like, it would pull from the wrong columns. Or it would pull — it didn't know exactly what it was looking at. We had to be really careful because we're like, we don't want to rely on that. It was just easier for us to manually go in and kind of like pull that data ourselves.

CHAKRABARTI: But in general, Elise says she has been pleasantly surprised by AI. She was initially a skeptic. And she does still get a little nervous about it.

AZARA: It kind of was this moment of, "Oh my gosh, I just wrote a press release in five seconds. That's insane." It's awesome because it means I can devote more time to other projects, but yeah, I'm like, oh gosh, this is just the beginning. Like, where is it going to go from here? You know, am I writing myself out of a job? I don't know. I hope not! Because I would hope that you would still have — everything still needs an editor. That's kind of where I see my job going is, okay, I can use it as a tool, but it's not going to like overtake what I do.

CHAKRABARTI: That's Elise Azara, a marketing director at Zero Emissions Northwest in Spokane, Washington. Well, David Autor is with us. He's a professor of economics at MIT and a very well known labor economist in the United States. And he's recently written an article that appeared in Noema magazine called AI Could Actually Help Rebuild the Middle Class.

So, Professor Autor, Elise is a really interesting example there. But use her to make your case that that you state in your essay that unlike automation or technological advances in the past, which we talked earlier generally has had a net negative effect on wages, that perhaps AI could rebuild or help rebuild the middle class. I'll be honest, professor, I'm very skeptical, but give it your best shot. (LAUGHS)

AUTOR: (LAUGHS) Great. So, first of all, there was so much in what Alicia (sic) just said. I thought that was terrific. And let me try to draw it out a bit. But I want to just step back a bit. I don't think computerization has lowered wages on average. It's that it's raised wage inequality a lot. So it has raised some people's wages a lot and other people's wages have fallen. And that's not, and I think that's, you know, not a deal I want to take.

CHAKRABARTI: Oh yeah, I would agree. And I think actually a lot of the ways that we talk about economic and labor changes due to technology is quite misleading as we keep pointing to those averages. But it's the rise in the extremes that's really problematic. But go ahead.

AUTOR: Right. We would be less productive and less wealthy in general, if we didn't have computerization to make many more things more productive, but some people would be certainly be better off.

But let's go back to Elisa (sic). What's really, you know, terrific about her example — and many things are terrific about her example — is she's talking about a use of AI for which there was no pre-AI substitute, right? Your word processor previously could check your spelling and look for grammatical errors, but it couldn't complete your thought and write a whole, you know, you take a headline and it produces a new paragraph and suggests different ways to write marketing copy, right?

And that's what illustrates something that's really distinct about AI. Previous computer technology followed rules. It just carried out the well-defined steps. But writing isn't like that. It requires actually going beyond the material, beyond the rules and putting in additional ideas and thoughts. And we don't know the formal procedures for doing that. We don't have a script for writing, you know, compelling marketing prose and AI doesn't either. What it's done is it's gone through vast amounts of unstructured text and ferreted out kind of patterns and ways of doing things that it can then replicate back for us.

And so it learns from unstructured information. And a lot of the work that we do is ultimately  in that way. It's informal. We do many things that we don't have formal rules for doing. And so, for example, you know, writing or telling a joke or coming up with a great idea or hypothesis or a new product, right? We don't have a program. We haven't historically had a program for that. So AI goes beyond what was previously feasible. You can think of traditional computing is like, you know, studio musicians reading from the sheets of music and AI is more like a jazz musician that can improvise and riff on the material that you give it.

But notice that what Elisa (sic) did not say is, "This does my job for me." What she said instead is, "It helps. It speeds things up, but I need to be there supervising it because it comes up with crazy stuff." (LAUGHS) Right? And so it actually requires her expertise to do this work. She did not say, "Anybody could take this machine and do my job." She said, "I need to be there to supervise." And this is true of a lot of AI applications. AI, because it's like a coworker, as she said. When you collaborate with someone else, you have to filter their ideas and their suggestions and determine, you just don't take someone else's opinion as ground truth.

And we see this with AI all the time. So, for example, some colleagues of mine, Nikhil Agarwal and Tobias Salz and their co-authors, ran an experiment with an an X-ray — an AI radiology tool. It's called CheXpert, Chest X-ray Expert. And what it does is it reads scans, and then makes diagnoses, you know, this might be pneumonia, edema, et cetera.

And it turns out this tool is really quite good. Working just with scans, without any ancillary information, it's about as accurate as 60% of radiologists are reading scans. Which is really amazing. And you can see this a beautiful application for AI because there are no clear, bright red lines for reading X-rays. It's a matter of judgment. You've seen many of them, you see patterns and you begin to recognize what it looks like. So AI is perfect for that because you don't have to tell it the rules, you just give it examples and eventually it infers the kind of underlying decision process.

So you might think that given that this tool is as good as 60% of radiologists, radiologists using the tool would be even more effective. But it turns out they do worse using this tool than they do on their own reading scans. And there's a reason, and the reason is what my colleagues call "correlated uncertainty." So when the doctor is uncertain what a scan shows, the software is also often frequently uncertain, and it will report its level of uncertainty. And when the doctor is certain, the software is also usually very quite confident, and it reports its confidence.

What tends to happen is when they're both uncertain, the doctor will defer to the AI. And when they're both confident and yet they disagree, the doctor will tend to override the AI. And in general, neither of those appears to be the right decision. (LAUGHS)

CHAKRABARTI: (LAUGHS) Huh!

AUTOR: When they're both uncertain, the doctor probably ought to go with his or her gut because they have a lot more ancillary information. When they're both certain and yet they disagree, the doctor should at least ask, look, this tool has seen millions of scans, more than I will see in my career, and we have a different diagnosis. The doctor should at least ask why they differ. That doesn't mean that machine can't be wrong. It just means it requires consideration. So here's the point I'm making, this comes back to Alicia's (sic) example — I'm sorry, did you say Elise or Elisa?

CHAKRABARTI: It was Elise.

AUTOR: Yeah. Elise. I'm sorry. Come back to Elise's example. It takes judgment to use the tool. The radiologist needs to supply his or her own judgment because this machine is opaque. It's error prone, and the fact that it appears confident or unconfident doesn't mean it's necessarily correct. And so when it comes to writing the marketing material, the software will come up with some good ideas, but they may be wrong. And you need someone who has an understanding of the job to filter that. And so she can --

CHAKRABARTI: Oh --

AUTOR: Yep, go ahead.

CHAKRABARTI: Oh, no, no. So let me, so just to cut to the chase here, what you're saying is that — I mean and this was, in fact, I think the meatiest part of your paper — which is that you're saying that AI is transforming what expertise actually is in the labor market and its value.

AUTOR: Precisely.

CHAKRABARTI: And it sounds like you're saying, well, expertise now is going to be replaced — well, the expertise in terms of the manipulation and synthesis of information, AI will be doing, but the new value add will be the judgment from human beings.

AUTOR: That absolutely right. So the most valuable work in the economy is decision-making work, right? When you're a doctor, you say, well, how do I treat this cancer patient? Or you're an architect designing a building or developing piece of software, or even figuring out how to land a plane or how to, you know, rewire a house. And in all those cases, these are high stakes decisions. They're one-off decisions, right? You're not making a decision for thousands. You're making a decision about an individual house or individual, individual patient and so on.

And that's where human judgment combines with information, knowledge, and formal training, but often, but usually it's a judgment call. And AI is helpful for supporting those decisions. It doesn't make them for you in most cases, but it's helpful to kind of give you guidance and guardrails. Guidance like had you considered this? Guardrails, as in don't do that, right? Don't prescribe these two drugs together, they're negative, they'll negatively interact.

And the good scenario, the way to use AI well, is to enable more people to do valuable decision-making work with this tool. In other words, to open up that high value work to people who are not quite at the elite frontier of the profession in which they are engaged.

CHAKRABARTI: Mm.

AUTOR: So an example I give frequently — and I want to be clear, this is not about AI specifically — is the occupation of nurse practitioners. Nurse practitioners are registered nurses who have an additional master's degree in nursing as NPs, and then a lot of training. And they do tasks that used to be limited exclusively to MDs. They do diagnosis, they do prescribing, they do treatment. And this is a great use of talent, of expertise. It lowers costs. It makes the care more accessible. It creates great jobs. Nurse practitioners are well paid.

And it didn't come about because of technology. It came about because of actually a social movement where nurses recognized they were underused and fought against the kicking and screaming of the American Medical Association to create this new credential. But at this point, they're very heavily supported by technology, right? They have electronic medical records, they have diagnostic software, they have prescription software that looks for drug interactions, and this enables them to do a greater variety of care tasks. And it's easy to imagine a future where nurse practitioners have a larger scope of practice, where they can do more valuable work.

And AI is a tool that can enable more people with the right foundational training and judgment to do more valuable work. So, for example, most software development is done by people with a fair amount of training in computer science. In the future, software development will be open to more people because you won't have to have as much formal knowledge of the tools. You'll just have to know what you're aiming for. Or if you're a contractor remodeling a house, you'll have better tools to visualize and show the customer and consider the engineering side of this.

And even in writing, just as in Elise's example, you will have tools that help produce text faster. But very likely, you will need your own judgment and expertise to decide what's suitable, what's valuable, and what should be discarded.

CHAKRABARTI: Okay, so Professor Autor, let me say that I have respected your work for years and years and years, but you're not quite shaking me out of my skepticism here. Because, because I wonder if we — I do think it's a potent argument that you're making. And let me just read a little section of your essay that summarizes what you just said.

You said that, "AI can enable a larger set of workers equipped with the necessary foundational training to perform those higher-stakes decision-making tasks currently arrogated to elite experts." You mention doctors, lawyers, software engineers, et cetera.

AUTOR: Professors. (LAUGHS)

CHAKRABARTI: (LAUGHS) "That AI used well can assist with restoring," this is the key part, "the middle skill, middle class heart of the U.S. labor market that has been hollowed out by automation and globalization." Okay. I am a little concerned though, or my skepticism rises because are we not falling into the trap of privileging the present, right? That very persistent human bias of not being able to accurately project that far out into the future. Because people probably made similar arguments in the past for previous technological disruptions. And to the point that you're making, it wasn't the number of jobs necessarily that went down, but the value of the new jobs being created.

And so I'm not sure I'm convinced that we will have a lot of high value new jobs being created in this AI future that you're describing. I mean, the other thing that Elise said is that she still has that nagging worry that she's training the AI out of her own job, right? Because what would stop her boss in the future from saying, "Well, I actually now don't even need an Elise. The AI has gotten good enough that I can just say, 'Write the the press release,' and I'm pretty confident that it'll write exactly what I need?"

AUTOR: Good. These are great questions So, let me see if I can respond effectively to a subset of them. So the the other thing I actually wanted to come back to on Elise is you might be concerned there's too much of a good thing, right? So if everybody can do it then it's not very valuable, right? Expertise by definition means something that some people can do and not everyone can do. And so one concern is that eventually it'll just displace expertise in some areas completely. And that will certainly happen. I gave you the example of Waze, right? In certain types of software development, the software develop will develop itself. And, uh, and this has happened many times in the past, right?

So, you know, we've had incredible productivity growth in agriculture, and as a result, we have very few people working in farming anymore relative to what it used to be, and yet we produce so much food. So we can certainly automate ourselves out of a job. But what has been true for individual sectors of the economy has never been true for the economy as a whole, right? It's not that productivity growth has ever made us poorer or less able to — or let me put it differently. Whenever we see a lot of productivity growth, we see booming employment and booming consumption, not the opposite. But that doesn't mean it isn't really negative for a subset of people, right?

CHAKRABARTI: Mm-hmm.

AUTOR: And we've seen this in many technological transitions. So in the movement from the kind of artisanal era to the industrial era, you know, skilled artisans were wiped out, the so called Luddites, right? I their replace, we eventually got a huge industrialized workforce in factories and offices, but it took decades for that transition to occur.

Part III

CHAKRABARTI: So, Professor Autor, let me get back to you because you were saying that in previous technological transformations, right, the impact really played out over the course of decades and you were headed in a particular direction. Go ahead.

AUTOR: That's correct. So the Industrial Revolution, it took about 60 years before it benefited rank and file workers. We had a good century of it from about 1880 to 1980 and it was a very productive way to organize work. It was complementary to people with high school educations and it built the middle class of a lot of the industrialized world.

However, the transition was ugly. It displaced a lot of artisans and it created hardship, and also capitalists used it to basically indenture children and unmarried women into dirty and dangerous work. And so it was, it was a rough start. And the era that we're in now also will have real pain.

So when I hear people talk about AI, I tend to think they're simultaneously too pessimistic and too optimistic. So let me start with too optimistic. People who are excited about AI will tend to say, "Oh, you know, no one will be displaced. Everyone will just be more productive and more efficient and they'll have more time for their children, et cetera." That's not true. There are certainly types of work that are — like those London taxi drivers I talked about — where the expertise that people have developed will be stranded. And there are still people driving London taxis, but their skills are not as valuable. Because most people can drive, or at least most people think they can drive.

And we will see that in some software development and some types of advertising writing. I'm very worried about content creators who do visual work and music because I think their intellectual property is being liberated without compensation by AI at present. And I think that's a regulatory issue. So we should be prepared for that. And not to, and I do not mean to dismiss that at all.

Why do I think they're too — why people are not sufficiently optimistic? Because I think there's great potential that gets overlooked in the discussion of which jobs will be automated. The most fundamentally important uses of new technologies are not for automation. They're instead things that enable new human capabilities that were not previously feasible, right?

So think about, you know, airplane and mechanical powered flight. Airplanes didn't automate the way we used to fly. We simply didn't fly before we had them. Or if you were to go back to ancient Greece and automate everything that the ancient Greeks were doing by hand at that time, you wouldn't have modern America, you'd have ancient Greece without horses. It wouldn't have flight. It wouldn't have electricity, wouldn't have penicillin. Wouldn't have telecommunications, right? Automation — or sorry, new technologies — have been fundamentally important not because they allow us to do the same things faster and cheaper, but because they allow us to do fundamentally new things.

And that's why new work is so important in the paper you mentioned earlier. A lot of new work, which is the type of work many of us do, requires expertise that didn't exist prior to the technologies that supported it, right? Those air traffic controllers, if they didn't have a GPS and a radar and a two-way radio, they'd just be a person standing in a field staring at the sky. They wouldn't be able to do their work.

Similarly, we wouldn't have pediatric oncologists without all the technology and tools and expertise that goes along with that. The same for all of these computer technologies we use, but even for many of the high-end services that we experience — in tourism, in travel and food. This is this is specialized work. It requires human expertise. That expertise is made valuable and it couldn't exist without the technology and tools that support it, as well as the higher incomes that come from it.

CHAKRABARTI: Okay. So, okay. You've half sold me, Professor Autor. (LAUGHS) 

AUTOR: Oh! Okay.

CHAKRABARTI: You're making progress here! (LAUGHS) But so let me refine it, because when you're talking about that new jobs created the nature of expertise and its value will shift, that makes perfect sense because it follows previous patterns as you talked about with each revolution in technology.

But when we're talking about rebuilding the middle class, class is a very specific word because we're talking about incomes, really, and people's sense of security in their lives. And you had just said previously, that both you and many, many other economists out there have studied extensively on how even as technology has changed the workforce and required higher levels of education for that elite expertise income, that it has for working class Americans, as we know, for the past 50 years or so, wages have been flatlined. Or, you know, perhaps more accurately, have just gone down relative to the percent of the national wealth that capital gets versus labor. I don't think you're disagreeing with that, right?

AUTOR: No, not at all.

CHAKRABARTI: Okay, so in the case of AI, what I wonder is, is that — why wouldn't that problem get anything but worse, right? Because we're just, we're relying on the vicissitudes of capitalism, right? To make the social and political policy to help keep the middle class robust and strong? I mean, I don't think that that's going to happen. And my concern is that the AI transformation of the nature of work is going to be so rapid. We've already failed to protect the middle class over the past half century. There's no guarantee that in the next five years even that we'll have policy makers who will do what's needed to be done to help, you know, as you're saying, rebuild the middle class with these technological changes.

AUTOR: Great. So first of all, there's no guarantee. Absolutely. And in fact, my article, it describes a good scenario, a way we could use it. It does not say it will happen that way. But let me say why I think it's useful for rebuilding middle skill jobs, a phrase I also prefer to "the middle class." And that is because we have displaced so much middle skill expertise from offices from factories and so on, pushed people into low paid services. And the opportunity is to use this tool to enable them to do more valuable work, right? To move into the health care services, to move it into software, to move into skilled repair and skilled construction and so on.

Now, that doesn't mean anyone can just do it. It means giving the right training and education. So it doesn't mean anyone could just simply do that work. If it were true, it wouldn't be expert work. And why this tool is useful for that — it's really fundamentally different from traditional computing, right? This is the world's leading technology that, by the way, can't do math and can't keep facts straight. It's really opposite. And so it would actually be somewhat surprising if it had exactly the same effects as the technology we used before. Now, I also, I want to underscore another point that you made, which is it is not the technology itself that makes these decisions, right?

CHAKRABARTI: Correct.

AUTOR: It is the choices of people, the choices and the incentives of the market and the structure of policy. And we have a variety of choices to make, and we could certainly use it badly.

For example, China uses AI very effectively to have the world's most comprehensive surveillance state, the world's best real time censorship system, and that's an amazing technological achievement, and it requires AI to do it. But that doesn't mean that's because that's what AI does. That's choice. You can use AI to make healthcare more affordable and accessible, to make education more immersive, more engaging, more like simulation. We can use it to help skilled repair people do a broader variety of tasks. We can use it to enable more people to do software development. So there's many, many ways to use it.

And I agree that I'm concerned about the incentives as well. I don't think corporations have necessarily the right incentives. It's not that they're malevolent, but why should they be concerned about aggregate employment? It's not their job. So policy is necessary. And we could talk at greater length about that.

I think we'll know we're being successful if we see more people without four year college degrees — which only a third of Americans have — doing high value work in education, in healthcare, in design, in repair, et cetera. And if AI is a threat to work, as we were discussing earlier, the people to whom it is actually most threatening are professionals.

CHAKRABARTI: Yeah.

AUTOR: Doctors and lawyers, software developers and engineers, right? It could make some of their work less expensive, more of it done automatically. Now, that's a mixed bag, but it's not altogether bad. First of all, professionals have seen a terrific last five decades, right? A lot of the growth of inequality is the rising wages of decision makers who are scarce, and yet made more valuable by computing.

And if we have a period of declining inequality, which we are in right now, that's not altogether bad, especially if it means that those high-paid professionals see more competition from skilled people who aren't quite as elite. If doctors have more competition from nurse practitioners, I'm all in favor of that. If lawyers have more competition from other legal experts who don't have as much training, but still do a great job, I'm in favor of that. If PhDs in computer science face more competition from people who have an associate's degree and the right tools, that's a good thing.

Not only is it good for the people who do those less highly paid jobs, it's also good for the rest of us who buy those services. We all pay for education. We all pay for health care. We all pay for software. I'll stop. (LAUGHS)

CHAKRABARTI: No, no. So let me just jump in here because the scenario that you're laying out right now, I can begin to see what you're talking about in terms of the increasing competition for work that was previously sort of narrowly defined as high value for a few, relatively few number of people.

But okay, we have another little thought experiment here about people who don't feel anywhere near as optimistic, okay? So some of those folks include British voice actor Mike Cooper.

MIKE COOPER [Tape]: I received a marketing email from a company that produced voiceovers. So I clicked through to their website, went to listen to their voice samples that they had listed, selected a couple of parameters like male, English, UK, because obviously I have a British accent, to see who they had on their list, and pressed play, and heard a version of myself coming back to me.

CHAKRABARTI: Okay, so again, this is Mike Cooper, who is a British voice actor, and last February he talked to Scott Tong on NPR's Here and Now.

COOPER: It was obviously me, it's the sound and tone of my voice, but it doesn't quite sound like me in the same way because we have nuances in the way we speak which aren't fully replicated in the AI versions.

CHAKRABARTI: Okay, so here's what he's talking about. We have a recording here that Cooper provided to the Washington Post last year, and that recording will then be followed by an AI replication of Cooper's voice produced by a voice generator called Resemble AI.

COOPER: My work has been described as exceptional. My performance is spot on in the first take, and my delivery as having an effortless authority.

AI REPLICATION OF COOPER'S VOICE: My work has been described as exceptional. My performance is spot on in the first take, and my delivery as having an effortless authority.

CHAKRABARTI: Okay, so not a perfect replication, by any means. Cooper's real voice is crisper, warmer, audio quality far better, and it captures really the entire range of human nuance that can be in our voices. But, that was the AI voice clone a year ago, okay? So, in just 12 months, how far has it come?

UNIDENTIFIED MALE VOICE: Okay, this is probably going to be too easy, but let's try it anyways. Tell me what you think.

CHAKRABARTI: Okay, so Professor Autor, what we just heard now is the voice of one of our producers, On Point producer Jonathan Chang. Or, actually wait, is it his voice? I haven't heard these yet, Professor Autor, so we're gonna listen to a second clip, and then you and I are gonna try and guess which one is Jon's real voice. So here's the other clip.

UNIDENTIFIED MALE VOICE: Okay, this is probably going to be too easy, but let's try it anyways. Tell me what you think.

CHAKRABARTI: Okay, Jon's in the studio. It's the second one. Is that you, Jon?! Oh, he shakes his head yes! I got it! So Professor Autor, it's not like completely accurate yet, but it was pretty close. I mean, could you tell the difference?

AUTOR: Not in the second, not in the second instance.

CHAKRABARTI: Yeah, right. The only reason why I could, I think, is because I know Jon really well and I listen to him talk every day. So just one more thought here from, um, from Mike Cooper. Because that's just the advances that AI in terms of replicating the human voice has made in a year. Like, what's it going to be like in six months or another year? And here's what Mike Cooper says.

COOPER: The idea now that you could take a minute of somebody from an audio book that they'd recorded or a minute of somebody giving a speech at a conference, create an AI model and then get Emma Watson to read Mein Kampf, in some ways it's like the the horse is out of the stable already. But as we've seen with things like ChatGPT and Dall-E, it's beginning to impact all kinds of creative people at this point, so we need to be careful.

CHAKRABARTI: Okay, Professor Autor, I'm afraid we only have a minute left in this conversation today. There's so much more to talk about. But, you know, again, I just wanted to surface that as the concerns that people continue to have. So what's the last thought you'd leave folks with?

AUTOR: Thanks. It's been a great conversation. So first I want to say Mike has a beautiful voice, Mike Cooper, it's fantastic to listen to. Second, as I think I said earlier, I think this is intellectual theft. I don't think — this is a misuse of the technology.  And it's a failure of our legal system that we don't have a framework for dealing with this. Think of music streaming in the days of Napster. Right? People were just recording their CDs and sharing them with everybody and no one was getting royalties for that. And we solved that problem with Spotify and with Apple Music. We need to solve that problem for AI.

What is happening right now is not appropriate and it will wipe people out. And in my opinion, it's simply stealing. It can be fixed. There's — it's not a technologically hard problem. This is a question of assigning rights and setting up a system for compensation and for setting up a system of intellectual property ownership, just like we do with publishing novels or patents or movies. So it's very problematic, but that's fixable. It's not intrinsic.

CHAKRABARTI: Right. Okay. So that's a great thought to leave us with, because again, it's fixable through human political decision making. (LAUGHS) And therein lies my persistent skepticism, not with what you're saying, Professor Autor, but with the ability for political decision making to be good.

AUTOR: But think of what the actors and television writers just did to negotiate this problem for their industry.

CHAKRABARTI: True, true. So maybe good positive change is possible. With that, I'm going to have to wrap up.