If you thought the battle between machines and jobs – the dislocation of labor and society resulting from digitization or automation – has been one-sided so far, just wait. The next wave of attack is well underway, and it’s called AI.
Artificial Intelligence, most simply, refers to computers that perform tasks that normally require human intelligence – things like visual perception, speech recognition, even decision-making.
Earlier this year the management consulting firm McKinsey famously wrote that “25 percent of the global workforce will either need to find new professional activities by 2020 or significantly broaden their technological skills. The World Economic Forum’s “Future of Jobs Report: 2018” states, “By 2022, the skills required to perform most jobs will have shifted significantly… [and] no less than 54% of all employees will require significant re- and upskilling.”
The greatest concerns are not just that AI destroys jobs, but that it increases inequality – that low-wage employees get displaced, while high-wage employees maintain or even extend their value that’s more difficult to replace with a machine.
Of course, new technologies have disrupted existing processes for centuries. Steam engines. Electricity. Microprocessors. Will the experience with AI be different than with previous technologies? Most importantly, what can governments, corporations, small businesses and individual workers do to not just avoid massive disruption, but rather position themselves to take outsized advantage of the opportunities?
To find out, I recently hosted an excellent roundtable discussion at Clayton, Dubilier & Rice with NYU Professor Robert Seamans.
Prof. Seamans’ studies how technology and governance structures affect strategic interactions between firms, affect incentives to innovate, and ultimately shape market outcomes. Previously he served under President Obama as a Senior Economist at the White House Council of Economic Advisers. He also co-authored an important review titled, simply, “AI and the Economy,” which explored the potential impact of AI on productivity and labor, and considered the various roles for regulation around antitrust, wages, data portability, and even immigration.
Transcript: Prof. Robert Seamans, AI & the Economy
Chris Riback: I think we should start with what I’m sure you will feel and agree is the pertinent question. If we thought that globalization and version one data of the digital age were bad for jobs, machines replacing humans, just waiting until the full impact of AI hits. We’ll have machines that can think and continuously improve, machines that can perform cognitive tasks and that’s when many jobs will really take a hit.
You might’ve read just this week that Microsoft invested $1 billion in Elon Musk’s Open Ai to pursue artificial intelligence that’s smarter than we are. So, I feel compelled to ask as a fellow who makes his living off of his brain, are you in trouble?
Rob Seamans: Thanks Chris. No, I’m not in trouble and none of us at the table right now are in trouble, at least for a very, very long time. I’m excited about the potential for AI to spur economic growth, but I think that the effects on labor, they get talked about in the popular press, have been misleading. There will be some substitution, but I think it’s going to far more be the case that there’ll be a lot more complementarities between AI and jobs. I think that there’ll be, as a result, a lot of employment growth. Those stories don’t get told enough, but I’m pretty optimistic about the net effect of AI on jobs.
Chris Riback: And we will talk about that, we’ll talk about the net effect on jobs. We’ll also talk about productivity and the differentiation between productivity and labor, but also some of the public policy aspects because while you may be optimistic, it sounds like about the net effect of AI in labor, I did pick up on your phrase net effect.
Rob Seamans: Yes.
Chris Riback: Because at the micro level-
Rob Seamans: You’re using your cognitive skills there.
Chris Riback: I don’t have a plan B, if AI replaces… I got nothing. That’s all I got, man.
Can we just start, quickly, with some history? New technologies have disrupted existing processes for centuries, steam engines, electricity, microprocessors. We are in the digital age, and have been for some time now. We’ve seen disruptions, but we’ve also seen productivity gains, obviously.
What makes artificial intelligence different from the run of the mill digitization and from the impacts we’ve seen to date?
Rob Seamans: Yes. So, let’s start first of all with history, so going back, right? So, you mentioned steam, electricity, maybe the first wave of IT, all of these have led to productivity growth. The thing that doesn’t maybe doesn’t get talked about a lot is that it takes a while for the productivity growth to take effect, right? So, just to give you an example of this.
Research that’s been done on factories switching from steam to electricity in the US in the early 20th century. It would take five to seven years, or so, before that factory would see any productivity gains. And the reason for that is that you can’t just take steam out and plug electricity in. You actually have to rearrange the way you do the production process. That takes time, that takes co-investment, investment in other assets. You don’t know what those other assets are. You don’t know what that rearrangement is going to be ahead of time, and so there’s a lot of experimentation, there’s dead ends and it takes time before you see this productivity growth.
So, I mention that for a couple of reasons. First, is to tie to current productivity growth that we are not seeing. This is one thing that people bring up. The recent increase in what AI can do for five, maybe 10 years, depending on how you date it. And yet it doesn’t seem like it’s showing up in the productivity statistics yet. And the reason for that is that it does take time for firms to learn how to use this new technology, how to deploy it. That’s point one.
Point two is, now this is my hypothesis, but firms that engage their workers, their workforce, in terms of figuring out the best ways to use this new technology are the ones that are actually going to see the biggest productivity boost from it. And so that’s probably the reason why I think that the story is ultimately, and on net, a good story for employment, and particularly for firms that really engage with their workforce.
Chris Riback: Will the integration of AI…? Do you see that gap, or the delta from the lack of productivity gains, at the beginning, to the speed of, let’s call it perhaps, exponential productivity gain, and you can characterize what you think the productivity gain will be, how quickly do you think we can see that that gap brought down? I would assume quicker than we saw from the shift off of steam into electricity, for example.
Rob Seamans: Yes. One would hope so, I guess, given that example, given that that was in the olden days, quote unquote. whereas presumably now we’re more efficient. so we can move more quickly, and things like that. I hate saying this, as a professor, right?
Chris Riback: Mm-hmm (affirmative).
Rob Seamans: But as an academic, I’m fine saying this, I don’t know. Okay? And moreover, anybody who says that they do know, is lying to you. They don’t know, they’re just making it up. But let’s start with a few conjectures here. So, I think there’s three big questions and we’ve been dancing around that a little bit.
First, is how much productivity growth will we get? Second, is when? And then third is what’s going to be the effect on labor? And we’ve started to talk about that a little bit.
Chris Riback: Yes.
Rob Seamans: In terms of how much the productivity growth will be, I think one good benchmark, and I would view this as the floor, as opposed to the ceiling. So, in other words, a reasonable guess would be a 10% increase in economic growth from AI. And I’m basing that off of research that’s been done looking at the effect that robots have had on economic growth over the prior two decades.
In terms of when, so I gave the five to seven year number a moment ago. Just extrapolating from that, it’s been five to 10 years since we’ve had AI, or these rapid gains in the lab, in terms of what AI can do, and we’re just starting to see commercialization of AI. And so, I think, we’ll start to see this increase in productivity growth, modest increase in productivity growth within the next three years, let’s say that. That would be my guess. Now, of course, if there’s a recession at some point that’s going to push stuff off a little bit, but that’d be my guess about the minimum amount that we would see, and when we would expect to see it.
Now, you said exponential growth. That’d be great if we could get exponential economic growth out of AI. I’m less convinced that we will. I think we will get growth, but I don’t think it’ll be as dramatic as some folks think. And part of that is we can look around the room and I think every single person here that I can see at least has their cell phone with them, or maybe tucked away as Richard has tucked away. There’s AI in here, we’re already using AI. I used AI to navigate my way from my home to this building. So, is that increasing? Is that somehow driving productivity growth?
It’s driving a little bit of productivity. It’s making me a little bit more efficient, I didn’t get lost on Broadway. But in terms of leading to 10% economic growth, no, we actually need more. We need something more. And so people start to think about what the killer apps, if you will, might be. So, some of that might be energy efficiency, some of that might be autonomous vehicles, or something like that. Maybe dramatic changes in terms of how we do logistics and transportation. You start to think about where you might get those killer apps that that would then drive a lot of economic growth.
Chris Riback: So, let’s turn to the other part of the equation, labor, which you said previously you’re bullish on the net effect, but at the micro level, I can imagine how things might be a little bit different.
You outlined in the paper three effects, three approaches. The theoretical perspective and empirical, or a historical perspective, as well as attempts to make granular predictions about nascent technologies and the effects that they might have on labor. So, if you would, give us your overall thesis about the impacts on labor overall?
Rob Seamans: Yes, so I think it’s good to think about three buckets, as we know, three different buckets then we talked about in the paper. and these are buckets of occupations.
One bucket of occupations are those, “Will there be some substitution and maybe those occupations will disappear.” That’s one bucket. The second bucket might be new occupations that are created. And then, the third bucket is all the existing occupations and the changes that will happen in those buckets, or in those occupations, as AI becomes more and more a part of everybody’s life.
I think those first two buckets, those are the two buckets that people talk about, jobs that disappear, and the new jobs, new occupations that are created. I think those are the smallest buckets. There will be some jobs that disappear, there will be some jobs that we can’t imagine right now that get created. I don’t think there’s going to be a lot actually happening though in those two buckets. I think it’s going to be more that third bucket, all the existing jobs and how these jobs are changing.
In terms of the effect on an individual and how they’re going to be affected? The topic of what you guys are talking about today, skills and retraining, to the extent that an individual can quickly access the training that they’re going to need to gain some new skills, that they’re given the proper incentives to get those skills, those individuals will be able to learn the skills quickly that they might need to transition what they’re doing in their occupation from perhaps the way it used to be done, to the way that it’s done in the future.
And just to give you an example of this, I’ll take my own profession. So, as a research professor, I spend half my time teaching, but half my time doing research. And 30 years ago a professor in my shoes would write out everything by hand, and hand it to a secretary who would type everything up. It’s now the case that you have to have the skill of typing as a professor, you have to have it. Indeed, in every occupation you have to have it, right? But it’s a skill you need to have it. It’s a skill that I learned over time, back in college. And so, occupations, we’ll see a lot of changes within occupations. It’s going to be incumbent on people to learn the skills that they’re going to need to do well in those occupations.
Chris Riback: So, the incumbency to learn the skills, and the ability to learn those skills, and the opportunity to learn those skills brings me to the next two areas that I want to talk to you about, inequality and public policy regulation in there.
Inequality, it is my own personal belief that inequality gaps are at the core of many of the challenges that not only our society, many societies globally are facing today. There are all sorts of other impacts as well, but the various forms of inequality and access to opportunity, et cetera.
You also describe~ in the paper how labor rates, say between skilled versus unskilled jobs, might be maintained. You just described it there, but at a high cost, or increased inequality. Will you describe that paradigm, how worried are you about AI driving an increase in inequality and how much rent does that take up in your mind?
Rob Seamans: Yes. So, for starters, in the paper we only touched on this topic.
Chris Riback: Yes.
Rob Seamans: There’s space constraints and other focuses.
Chris Riback: That’s why I wanted to talk to you more about it now.
Rob Seamans: Okay. You’re putting me on the spot. Yes, so I do worry about the inequality piece. I’ll say two things on that. First, some of the current research that I’ve been doing is looking at how advances in AI have been affecting labor, both employment and wages, over the prior five to 10 years. And on net, I see roughly no effect on employment, but I do see a positive effect on wages.
However, the positive effect on wages is concentrated in those occupations that had high wages to begin with. Okay? So, that that’s suggestive of the fact that AI might increase inequality, at least across occupations. You might imagine it would happen within occupations as well. So yes, I do worry about that.
From a policy point of view, one might imagine that one could create training programs that would try to help people transition from one occupation to another, that might help people get skills that they need to do well in their current occupation. And one of the promises of AI, if you will, that people talk about, is that there might be a way to digitally deliver some of this skills training.
Now, I think that’s true, that’s probably right. However, this is an area where I feel like a certain type of inequality, that I’ll talk about in one sec, is very worrying, and that is the inequality in access to broadband.
Chris Riback: Mm-hmm (affirmative).
Rob Seamans: So, if we think that we might be able to sort of deliver training to people remotely, or something like that, so that they can learn the new skills that they will need, then we need to make sure that people can actually access this training.
In fact, however, that’s not the case. So the statistic that I’m going to give you is from 2016, but as of 2016, half of the households in the bottom income quintile in the US had no broadband at home. So, everybody here in this room, this is not an issue that we deal with, we can really easily access whatever we want from the Internet, we can train ourselves easily via the internet, or via our phones.
Chris actually turned the tables on me this morning and asked me what I thought of the New York Times article that was published this morning, but I had to admit I hadn’t yet read. So, while he was getting his coffee, I very quickly pulled it up, and I’m waiting for you to quiz me on that. But you haven’t done that yet but so I can do that because I have access to the Internet. I can retrain myself quickly. I can learn skills that are remotely delivered to me.
However, half of the households in the United… Sorry, half of the households in the bottom income quintile of the United States, precisely the households that we might worry about, are going to be adversely effected by some of this, don’t even have access to the technology, that we might think they would need access to, in order to undertake some of this retraining.
Chris Riback: Lack of access to broadband is one area. Are you more worried about inequality gaps expanding because of the lack of ability to access re-skilling and retraining? Whether that’s because I don’t have broadband, so I can’t retrain myself, which by the way, even if I did have broadband, seems like a pretty high bar if there isn’t maybe a public policy program around it and there aren’t enterprise, or pro business incentives around retraining.
Rob Seamans: Mm-hmm (affirmative).
Chris Riback: We’ve just gone through the beginning stages of globalization where we’ve seen what’s happened with manufacturing jobs and the lack of investment in retraining and re-skilling that occurred while there were all sorts of economic challenges in the world. So, are you more worried about that aspect? Or are you more worried about the part of AI displacement where the rich get richer? Meaning the higher skilled get more access and more higher skill, there becomes a greater wage gap on that level?
Rob Seamans: Yes. So, you’re asking a good question. I think both are worrying. I don’t how I weight the two. However, my focus on broadband is because the fix to this is a relatively cheap fix. Relatively cheap meaning several billion dollars, right?
Chris Riback: Yes.
Rob Seamans: In terms of trying to spur, let’s say, rural broadband or something like that.
Chris Riback: And next week is infrastructure week again, isn’t it?
Rob Seamans: Again or was it a recess next week? I don’t know.
Chris Riback: Oh, maybe it’s a recess.
Rob Seamans: But just to put a few numbers out there. So, the earned income tax credit, which is a fantastic program that we have in the US, that helps incentivize people to learn the skills that they need to switch maybe from one job to another job. We spend 75 maybe $80 billion on that a year, eight billion for broadband is a 10th of that. I don’t want to shift money from one to the other, but just in terms of orders of magnitude, it strikes me that increasing access to this new technology via investment in broadband, or other types of connectivity, is a relatively cheap way to try to start to address some of this.
And by the way, also a bipartisan way to try to address some of this. This an area where the Obama administration was very focused and where, also, the current administration has been focused.
Chris Riback: I’d like to ask the folks around the table start thinking of your question, or questions, because we’ll have time for a one or one or two of those.
Let’s shift then to regulation, public policy and the pop quiz on today’s New York Times story, which was all about the value of data and got to data portability, which is one aspect.
And there was a figure that was quoted, a report. We all know every click mouse stroke thing that we do on a computer, or a phone, there’s really valuable data behind that. And the trade is, generally to overstate it, we get free stuff. And Google and Facebook, and then others, get our data and that data has a ton of value and this one study put that value at $76 billion.
And so, the questions around that are, what should we do about that? What should we do about that? Perhaps whether that’s an inefficiency or is that a fair transfer? Are there private sector solutions that should think about redistributing that value? Are there public policy questions? So, around data portability, there are also immigration issues that need to be looked at. So one, this is the easy one, does AI require some level of regulation? Two, if so, what level of regulation? And maybe even three, should there be an AI specific agency?
Rob Seamans: So Chris you, you threw a lot out there, you even toss in immigration, right?
Chris Riback: Yes.
Rob Seamans: Immigration is the easy one. We need a different immigration policy than we currently have, or than this administration has been pursuing. Anecdotally, I haven’t yet seen systematic research on this, but anecdotally, a lot of really good AI researchers are leaving the US for Canada, or not coming to the US in the first place. And so, if we as a country want to be a leader in AI, we need a different immigration policy than we currently have. That’s perhaps a separate…
Chris Riback: It’s a separate. And can I throw one more in there?
Rob Seamans: Oh sure.
Chris Riback: Yes, why not? Antitrust.
Rob Seamans: Antitrust, yes. I was waiting for you to say that. Great, okay. On the antitrust piece, I mentioned bipartisan a moment ago, one thing that strikes me as very different now, relative to say five years ago, on the antitrust front, is how dramatically things have shifted for the large technology companies. They used to be the darlings of what we’re doing here in the US. And now you have folks on the conservative right that are worried about bias, let’s say, in terms of bias search results. You have folks on the progressive left that are very worried about some of the antitrust issues perhaps.
And so, I think it’s a tough environment for a tech firm right now. Probably a great environment to be a lobbyist, right? That was meant to be a joke. Okay. But I can say more about the antitrust in a minute and maybe tie it to the data portability, but let me just touch on the AI regulation.
There are a bunch of aspects to this. So first of all, no I don’t think, and I think we described this in the paper, no, I don’t think that we want a specific agency whose entire job is the regulation of AI. Right now the system that we have in Washington DC is, we already have agencies that are very specifically focused on their different areas. So, for example, we have National Highway Traffic Safety Agency and that’s focused on roads. We have SCC that’s focused on trading and the stock market, And the like. And there’s just a lot of deep expertise there that we want to remain there. And ideally what you’d have are some additional people in those agencies that are really well-educated about AI and what it is that AI can do. We don’t want to somehow try to have all of the issues related to AI as it pertains to public markets, or as it pertains to roads, somehow be siloed in their own agency. That doesn’t sound like it makes a lot of sense.
But you could imagine maybe a sub agency, or a subgroup within each of those existing agencies, that’s focused on AI. As a side note in Congress, we certainly want more tech expertise, and one way to do that is via something called the Office of Technology Assessment. This was something that we had back 70s, 80s into the mid 90s that we then got rid of. I think something like that makes a lot of sense and it would be relatively cheap to bring back in, and it would be a way to keep members of Congress educated about technology.
So, I don’t think we want an AI-specific regulator, or an AI-specific agency, but we do want more expertise in the existing agencies. And you could also imagine maybe, at a minimum, a working group. It might start at the White House that’s focused on some of the issues around bias in AI. I think something like that would be really important to look at.
Chris Riback: Questions from around the table? Yes?
Audience Question: I would just love to hear your thoughts on how organizations are positioning themselves as AI leaders, or experts, or leveraging that technology within their world. It doesn’t have to be in the healthcare space, but it just seems like there’s a lot of hype out there.
Rob Seamans: Yes, so I completely agree with there’s a lot of hype out there. There’s no question about that, there’s a lot of hype out there. Let me just take a step back for a minute.
So, when you mention healthcare, earlier Chris, when you were asking about productivity gains and where we might see them, one exercise one could do is go through different sectors of the economy and try to think about where we might see a lot of gains from AI.
I mentioned autonomous vehicles as being one. Healthcare you could imagine being another, using AI to try to identify at risk individuals earlier than we currently are. That might lead to people being diagnosed quicker than they currently are. That might lead to lower costs for that individual in the long run, perhaps a slightly longer life, a more productive life. You can play that out. And I think healthcare is one of the areas where I would expect that we’d see a lot of productivity gains. Education would probably be another.
Where you probably would not see the productivity gains is in areas like marketing, right? Where it’s like a zero sum game where if you can market to an individual a little bit better than some other firm, that just increases the incentives on the part of the other firm to invest in a little bit more AI to try to market it a bit better, but at the end of the day, you’re still just going to sell one product. You’re just fighting each other to see who can do it more efficiently. So, that was sort of the response on healthcare.
In terms of the hype piece, I’ll give you an example again from the autonomous vehicle sector. So Elon Musk loves to talk about… You brought up Elon Musk earlier.
Chris Riback: Yes.
Rob Seamans: Loves to talk about how there’ll be fleets of AI-enabled autonomous taxis within a year. So, that that’s one end of the spectrum. That’s the hype end of the spectrum.
At the other end of the spectrum is someone who knows a ton about autonomous vehicles. It’s an individual named Chris Crimson, so this is a former Carnegie Mellon University professor. He won some of the early DARPA self-driving car challenges in the deserts, I believe, of Nevada and was probably the one of the first employees at Google X focused on self driving cars. And he ran what became Waymo, he ran that unit for a long time before leaving to found his own startup. So, he knows a lot about this space. He’s also very bullish about it and he thinks that there will be a lot to happen here, which is why he’s running his own startup. And his best guess as to when half of the vehicles out on the road will be autonomous is 30 to 50 years from now. So, this is somebody who knows this stuff really well and it’s the opposite end of the hype.
So yes, there’s a lot of hype out there and my suggestion there would be to think about people that really know what they’re talking about and try to get their sense on what AI is doing, and what it holds for the future.
Chris Riback: Which leads perfectly to the close, which is yet another Elon Musk reference. In your paper, Musk said that you quote him as saying, “AI is a fundamental risk to the existence of human civilization.” He might be prone to a little bit of hype. So, is the takeaway, we don’t have to worry about a human civilization from AI just yet?
Rob Seamans: Chris, we don’t have to worry about human civilization being negatively impacted by AI. We’re going to be around for many, many, many, many centuries.
Chris Riback: Okay. Well, our next conversation will be on climate change.
Rob Seamans: Yes, there we go, thanks.
Chris Riback: Rob, thank you for your time.
Rob Seamans: Thanks.