Transcription
Good morning, everybody. My name is Paul Oyer. I'm a professor of economics here at the Stanford GSB. And I'm merely, merely a moderator. Uh, and I'm doing this because I do teach about AI. I don't know enough about it like these people do. But we're going to learn a lot. So, welcome to the session. Uh, certainly very timely, and I'm guessing a lot of you in the audience are heavily invested in this issue. Uh, many of you actually work at the companies that build the technology we're going to discuss, and some of you are probably building AI products that are going to affect the labor markets. Um, so I want to, we definitely want to, hopefully this will be great. So, we're in the interesting position of discussing a transformation while it's happening, and uh, with an audience that includes some of the people that are involved. So, I think that's a really exciting thing.
So, I want to introduce the panel real quick. Uh, at the end is Susan Athey, the economics of technology professor in the economics group here at Stanford. She brings economic rigor and a long track record on technology and labor markets. She's deeply embedded in the Stanford tech policy world, but also for the antitrust division of the Justice Department. Susan is a, a world-renowned economist. I won't list the awards she's won and the committees and the societies of which she's a member and officer. Let's just say it's a long list. And most importantly, when I asked, uh, Claude about this, uh, to describe our panel, Claude said that Susan is, quote, "the bridge between economic rigor and the actual boardrooms of big tech," which I thought was a pretty good description. I thought they did a pretty good job with that one.
Uh, I'm gonna go skip across to Nila. Nila is ADP's Chief Economist, and she and her team have provided a lot of valuable insights, uh, through her access to real payroll and employment data from tens of millions of workers. If you listen to NPR's Marketplace, like I often do, you'll often hear Nila commenting there. Um, she knows what's happening in labor markets right now, uh, more than almost anyone else. And, uh, her her role in the panel, per Claude, it might have been Gemini now that I think about it, but I don't remember, is, quote, "the person who sees the real-time heartbeat of the American workforce." Again, pretty good, don't you think? Pretty well on that.
>> I'm gonna get a t-shirt made.
>> And then, uh, Tami Bas Baseroglu, how'd I do? Good? Okay. Is the co-founder of Mechanize, which, and that's a company, a venture explicitly trying to automate knowledge work at scale. Tami's also done serious research on AI progress and economic impact. Uh, he provides an interesting perspective here because, uh, based on his outspoken belief that AI will displace many workers, and is backing that up through a direct commercial stake in displacement. Tami did a really interesting talk here on campus recently at the Digital Economy Lab, and I can tell you it was extremely provocative. Um, and according to whichever AI tool it was, his role in this panel is to be, quote, "the one tracking the exponential curve of the machines themselves." So, we're looking forward to that, Tami.
So, um, let's just start. I, I thought it would be most interesting because Tami is often, uh, seen as, as, you know, really provocative, to let him start by explicitly, uh, your your company's explicitly trying to automate knowledge work. And so, if you wouldn't mind starting by giving us your honest assessment of the timeline and how you think the labor market con, how you think about the labor market consequences of what you're building. That would be a great launching off point, and then we'll let the others jump in.
>> Sure. So, uh, for context, I co-founded this company called Mechanize. We build environments for training and evaluation where we teach models how to do kind of knowledge work. We, we focus on, on software engineering right now, which is the domain that, you know, the AI labs are competing most intensely over. And so, we work closely with the leading AI labs and help them advance the models' capabilities at doing, um, software engineering. And in, you know, in the future, we want to just support, um, you know, the capabilities needed to automate knowledge work. And, um, so, so my, my assessment of the timeline is that I think over the next, maybe, you know, one to five years, we're going to see this continuation of this kind of gradual automation of tasks that knowledge workers are doing, you know, in particular in, in software engineering and engineering. And that's been really exciting. But, but also more broadly in, in other domains of knowledge work, in, you know, consulting and, um, and, you know, uh, finance and banking and oil and gas and, and whatnot. Um, I think I don't expect like a very drastic amount of acceleration in, in automation over the next one to five years. I expect that in the next one to three decades, we'll probably have very widespread automation. So, I think, you know, over the next one to five years, I might expect maybe a couple single-digit percentage points of jobs being displaced or automated away. But over one to three decades, I, I expect that to be maybe the majority of, of work in the US economy ends, ends up being automated. And so, maybe one way of making, you know, a concrete prediction is economists pay attention to, you know, the labor share of the economy, which, um, you know, tells you the fraction of, of output that's paid out in wages to, to workers. And I expect within, you know, one to three decades that we'll see, um, more money spent on running, you know, AI workers in the economy than we will, um, spend on, on, on human workers. And I, I expect that that to be pretty likely within one to three decades.
>> Great. Thanks. So, Susan, what do you think of the timeline that Tami has laid out, and how does it square with what you're hearing? You know, you talked to a lot of policymakers, executives, and of course, I know you're in the middle of teaching MBA students as we speak.
>> Yeah. So, let me start from more of an, an academic perspective and then circle back to what I think I'm, how that squares by what I'm seeing. Um, I think a principle that I'm seeing across every, everything I'm doing is that it can be very helpful to separate technological capability from the, from the, the implementation and the change that actually occurs. So, we can have great technological improvement, but that improvement only translates into impact when it's combined with an existing system or an existing organization. And so, we, if we, and, and to reason about both the timing and the impact, I find that I need to really dig into the first principles of the scenario and then pop up from there. So, just take an example of drug discoveries. We can really widen the pipeline of drug discoveries, and somebody can show you how they can come up with new molecules, and it'll blow your mind. But then you have to think, okay, well, what about the human trials? How are we going to make the human trials go faster? How are we going to make the FDA in the US go faster? And of course, if we are piling up amazing drugs, eventually we will figure out how to invest in the capacity of the, the human trials, and we will change the FDA. But there are things like maybe that requires an act of Congress, and, you know, that requires political parties to be aligned. So, so there are things that can stop that, even when the, when the economic returns are very high. And then at a micro level, something probably all of us can relate to now is that we can see that it just got a lot easier to write emails. Great. I can write my emails really fast. But what did that do to your inbox? And what did that do to productivity actually? You know, that, that everybody is getting spammed, and, you know, I, my students preparing a slide deck for me, but now it's all done by Claude, and it's like not editable easily. And, you know, they, and it looks beautiful, and I can't actually tell until I read it that they didn't think about it, you know. So, so like, we, so we, it takes time. We're going to have to change how we communicate, and we can all, we could have a whole session about just how we're going to change communication and business and, and, and we will figure it out, but it will, it might be bumpy along the road. So, I think that's a pattern of every general purpose technology is that there's like a technology change, but then there are micro complementary innovations, and the innovation can be like an organizational innovation or a norm, and then there are sectoral factor reallocations, and it just one prods another, prods another, prods another, and it's iterative and it's slow. So, that's, and I think we're going to unpack this theme further, so I'll stop on that point now.
I want to come back to answer your question of what am I hearing? I think actually students, I have two kids in college right now. I mean, fear, big fear in MBA class, you ask who is afraid. What is your most popular topic of conversation right now? Passionate. It's like, "Oh my god, I have to choose my career, and everything I knew is wiped out." So, this is terrifying. And I also think that filters down to the high school kids and the parents. So, there's a lot of fear. Um, and, and that's something that also has huge political and geopolitical implications. In terms of the firms, I, I worked at DOJ where I helped, was on a leadership team of of hundreds of people doing white-collar work, and I also, we have a lot of professional services coming through Stanford, and I work very closely with some of them. Everybody's trying to figure out how to reorganize right now. If you ask any attorney, they're, "Oh my gosh, how are we going to train our new attorneys?" You know, how, what's the world look like? Um, the organizational change is pretty profound that's needed to fully take advantage of this. So, going from the pilot of like, "Oh, great. I can search my company documents and get a summary," to "I'm going to redo my workflow." I think people are just at the beginning of that. And I'll leave with one very interesting thing the MBAs told me last week was that the most spammed job opportunity they were getting on from LinkedIn, they were getting the most LinkedIn spam to be a forward-deployed product manager.
>> So, we, we can come back to that too.
>> Let me just add, let me add one anecdote to your point about like, you know, we're trying to figure all this out, and the junior labor market stuff is really complicated. Um, I love to use this. I'm going to use this in my class in a few weeks if any of you are in my class. In, uh, Bloomberg had an article about how OpenAI was trying to train people to be junior investment bank, train their models so that they wouldn't have to hire junior investment bankers. And then literally three days later, they had an article about how the investment bankers were using the junior investment bankers to teach them how to use AI. Which, you know, I mean, maybe they're talking about replacing them, but you couldn't help but think the demand for the junior people was even going to go up because of the AI. So, we'll see how that plays out.
Uh, Nila, what's your take on Tami's and Susan's assessments of where we are and where we're going? And, and you can either talk about that broadly, or if you're ready to certainly bring in some, uh, ADP data and what you're seeing on whether or not we have any real displacement, uh, through AI yet.
>> Well, thanks, first of all, for inviting me to, to be on the panel, and thanks to my co-panelists for giving me so much to respond to. I hope over the next 45 minutes, I can touch on three different themes: context, sentiment, and application. So, I'm already stacking the deck in terms of the questions. Um, but let's start with the context because I think it matters. AI is not happening in a vacuum. And so, I would love to level set on the context in which this transformation is taking place. And I can do so from the vantage point of ADP, which pays over 42 million workers around the world. It's tax day. ADP will issue over 80 million W-2s in the course of a quarter. And what it means is that I get to see about 1/5th of the US workforce in real time. We've started sharing that on a weekly basis with everyone. And what I can say, short answer is, we're not seeing a big effect on the labor market right now. But we can unpack that a little bit later. Here's what we are seeing. If you look at the last two years, three out of every four US jobs that were newly created came from one super sector: education and healthcare. And if you unpack that a little bit further, you'll find that the biggest occupation, the biggest sector that was seeing those job gains were home healthcare aides. Why? Not because of AI. Another A that is way more dominant in the labor market, which is aging. We are looking at mass retirements of the boomer population, and it is the wealthiest generation in history, and they are choosing to age in place. And there is yet to be an AI robot that can make their do their laundry or make them sandwiches. And that is what the customer ultimately wants, and we are far from that. So, where the jobs are coming and where the tech is going, totally different. Let's add on to that because the US is not unique. In fact, if you look at Europe, they're 10 years ahead in terms of the aging demographic. China, Japan. I just got back from Vietnam. The Southeast, Southeast Asia is at their demographic sweet spot in terms of the labor market, and yet they're only going to be in that sweet spot for about 10 years because of fertility policy. It's already ticking that time clock in which they can take advantage of having more workers than dependent individuals. It won't be forever. And so, where is the growth coming from? Africa. Three-quarters, and, and I know Susan can talk a lot about this, three-quarters of the working-age demographic is going to come from a continent that's not part of this conversation on tech. So, that's the context, and so that does color my perception of where the tech is going and this idea of diffusion or how it gets to application. I'm going to touch on the other two really, really quickly for sentiment and then application, and I hope we can jump back to those.
Sentiment. Yeah, I have two, one college-aged, one about to be a college student. Uh, they, they're on a predestined path. AI is not going to affect it much. Uh, but, uh, I will say there is a lot of insecurity out there. In fact, when we survey about 40,000 workers in 36 countries, only one in four say that their jobs are safe from elimination. They're probably in low-paying home healthcare aide occupations, a lot of them. In the United States, it's 28%. Um, what our research does show, though, is that if you invest in people by upskilling them, that safety number increases by five times. So, this is not like the weather. We don't just need to carry an umbrella or have a bigger jacket. We can actually change the course of the environment we're in just by investment. And that leads me to application. When you think about what the frontier research is saying, it's not macro. And,
>> I don't think anyone's a macroeconomist on this panel, so we can, we don't have to deal with that.
>> I was in charge of choosing who was on the panel.
>> Oh, you did.
>> Excellent. Excellent.
>> Side step.
>> Safe from that.
>> S&P has had a whole ton of. Okay, we don't have to talk about that. No, AI works at the task level. And so this idea of occupation, man, that's going to be an old concept after a while. We're not going to be talking about being an economist. We're going to be talking about the kinds of tasks, skills, and activities that define our work. And that's going to change because AI will change it. So, it's going to be a lot more fluid. Um, I joked with, um, Eric Bernolson, who's our partner at the, uh, uh, on the ADP employment report and also runs the Stanford Digital Economy Lab, that one day, um, we're not going to be talking about a jobs report. We're going to be talking about a task report, the number of tasks created and destroyed. And I said this in jest four years ago to an audience similar to this one, and Eric was in that audience, and he said, "We can do that." So, that's what we're going to do. Um, we are going to redefine how we think about jobs and occupations, and we're going to do it for the betterment of the worker because if we do that, then we get the promise of growth that is inclusive and shared.
So, let me ask one quick follow-up, and then we'll move on. That regarding Eric, who you, I knew you worked with, and when you said you're not seeing much in the data, that strikes me as a little bit in conflict with the paper of Eric's and the others that's gotten a lot of, uh, attention in the coal mine. How do you, how do you reconcile that? For those of you who don't know, Eric and a couple of co-authors have a paper suggesting that junior soft, young software engineer, the job market really has been negatively impacted by LLMs.
>> So, that is one of two papers using ADP data that use the word "canary." So, I, I think canaries and ADP, and the reason why, um, Eric was able to do this research using ADP data is because it's so granular. So, it says multiple things, not just one thing. Um, what they were able to do is use ADP's job title taxonomy, which we've been working on for a very long time, several years, and classified jobs in terms of AI exposure. Both AI-exposed careers like software developer or customer service agent, and also AI non-exposed careers like that home healthcare aide I told you about. And what they show is that during the rollout of ChatGPT and afterwards, in October 2022, you see a distinct drop in employment for early career people between 22 and 26 in AI-exposed careers, um, like software developers and customer service agents. It also shows that for more complex jobs related to more tenured workers, older workers, you see a ramp of of employment. So, I would say that the data does suggest that there is an automation of work, and the trick for the HR professionals is not, "Okay, those jobs just went away." It's upskilling those young workers to more complex tasks. The augmentation role of AI is what's really, what that paper talks about, not the automation work. If you look at augmentation, if you look at people older in their careers doing more complex jobs, what you see is employment growth. And that's not a foregone conclusion. I thought AI would just get rid of all the old, expensive people.
>> Right.
>> Right.
>> Especially because the junior ones know how to use AI.
>> Exactly. AI is not getting rid of me. It's going after my kid who's going to graduate into a role that no longer exists. And that is what we should be concerned about. Upskilling the youth to take advantage of the, of the, the tool which they are natively born to do, apparently. Instead of a silver spoon, they have claws. Um, but we need to upscale them into the new job, the new tasks that are coming, not the old ones that we have educated them for.
>> And I'm glad you qualified with upskilling the youth because the history of upskilling the older people who are threatened by tech change is not, you know, I like to say it's unblemished by success.
>> In if we look historically at upskilling older people who've lost their jobs. Let me move the conversation forward, and if from the context of your follow-ups, you want to come back to anything, feel free. So, is this wave of automation, do you feel like it's qualitatively or quantitatively different than the spreadsheet, the, you know, auto, electricity, the farm, you know, the plow, um, you know, compared to mechanization, computerization, offshoring? What pattern matching do you see that's appropriate, and what do you think's different from prior technical changes? And, uh, why don't you start us off, Tami?
>> Yeah, I, I guess one thing to say about junior software engineers, we at Mechanize are actually like aggressively hiring junior software, and we're looking like, "Oh, you know, junior software engineers are, you know, not going to get jobs." And then my, my own experience here is that I'm like, really aggressively trying to hire them. But, but the, the kind of the, the skills that we look for is very different than than it was a couple years ago, where we look for people's, we explicitly evaluate during the interview process, people's ability to use things like cloud code and, and, and make a lot of progress, produce a lot of code, and still like understand the code that the models produce, and being able to orchestrate models in parallel, and things like that. And we find people who are very good at that to just be enormously productive, and, um, and, and they just generate a lot of value. And so we want to kind of hire quite, quite a lot of them. And if there are any software engineers here who are interested, happy to chat.
>> You have raised hands here.
>> Uh, so is this qualitatively different? I mean, I think, you know, in software engineering, there has been, you know, many decades of automation. Um, that looks fairly similar to what I'm seeing with AI today, where, you know, um, you know, first you had compilers that automated the writing of machine code. Previously, engineers would kind of handwrite machine code that kind of got automated with compilers. You had higher-level programming languages like Python and so on, that kind of abstract away assembly code. Um, and so we've had this long run trend of higher and higher levels of abstraction in software engineering, where, um, you know, you have, um, you know, in, in the 2000s, you had web developers spend many hours writing boilerplate code that like an import statement in Python can do in, you know, in, in like a second. So, you know, we've had this long run trend of higher levels of abstraction, and it seems like AI is this continuation of this trend, where instead of writing a lot of the code, you know, software engineers are now just orchestrating agents to do a bunch of the engineering, and deciding which things to test, and how to instruct the models, and, and, and noticing the weaknesses with models, and maybe manually, um, kind of, uh, manually stepping in when there are things which models struggle with doing, for example, you know, testing UX and so on, which, um, you know, the vision abilities of coding models are not that great, and, and they have some limitations there. So, they have to figure out when to step in and do that. And so, I see this being this continuation of this long run trend towards higher and higher levels of abstraction, and it seems fairly similar in to to previous, um, kind of automation. Now, the thing that is pretty different is that, you know, compilers and high-level programming languages don't promise to be able to fully automate everything that an engineer can do. And AI, I think, you know, eventually will be able to automate everything that software engineers can do. I think that that'll take a couple decades, but I think we'll eventually get there. And so that is like a very big difference compared to prior technologies.
>> Does that mean you, you foresee a world where we just have excess workers relative to what we need to get the job done? Like, are you, you know, when you talk about these, this displacement in three to, in one to three decades, are we talking about unemployment? Or you're not, you haven't really...
>> Yeah, I'm talking about unemployment there.
>> Okay. Yeah.
>> Okay.
>> So, so it, it would just mostly be done by agents and...
>> But that's a big difference from the past, obviously.
>> Yes. That, that would be a big difference. So, so eventually we'll get to a place where there's a lot of unemployment.
>> Okay. Susan, uh, either picking up where Tami left off, or just going back to prior waves, I'd love to hear your input.
>> Sure. So, um, let me start with, uh, just a comment about like time scales. So, personally, um, I'm trying to focus more on the next 10 years because I feel like if we actually survive in our current form to that world that you described, like we're going to be very lucky. So, we can figure some of that stuff out when we get there. Um, but to me, it's seeing a path to get there is, is so, so not that it's not interesting to think about that that future, but it's, it's a lot of things have to happen first that we should be focused on, in my view. Um, on the, I want to first just respond to the, "What do young people do?" question. Something I'm thinking about a lot from teaching and also my, my, my kids and from, I'm in my work right now. I'm working on legal and professional services, and, and it's already pretty much a documented fact that with some effort, not off-the-shelf, but with some effort, AI is better than the most junior people at the most routine tasks. And so, that is just, that is a wave that's happening, and it's happening fast, and people are already thinking about how they're going to build and how they're going to restructure. Um, from the, "What do the young people do?" though, going back to this forward-deployed product manager. Um, my son's a STEM major, math, physics, never, not a lot of coding. First, a year ago, he wasn't really, he was in the worried camp. Then he got, um, got involved interning with a startup, and when he just got his offer letter, it was a forward-deployed engineer. I don't think he knew what a forward-deployed engineer was, maybe not even until he got his offer letter. Um, but that's what he was doing. And they wanted him not because of the skill, but because he was a problem solver. So, he was a, and, and nothing he learned in his math or physics classes, but just the breaking down of a problem into components and solving it. And there's a huge demand for right that right now because if you have an idea about how to conceptualize and solve a problem and break it down and measure and test, then you have these tools that can make that idea a reality. But in terms of like, what are the people going to do? If you think about all these people in India or other places that were like, they, they were doing some stuff before for big kind of IT firms. If the whole world is going to start digitizing and solving problems, like it's a, it's at least a 10-year journey to digitize and redo the IT stack of every organization in the world and create IT stacks for every kind of organization in the world. So, this feels like, just like your company, like this is a big expansion before it contracts.
Now, let me come back to, like, the, your, your first question, which was the, the, how is this wave different? So, let me contextualize one of my projects for the last year. I'm the faculty advisor. There's two of us for the World Development Report 2026 on Artificial Intelligence. So, we've got about 50 people working. I'm traveling all over the world to create a huge report that's going to inform World Bank policy for developing countries. So, I've been hearing, listening all different countries, Africa particularly, and we've been trying to customize this based on data. And there was a, there's just a few facts out of, we haven't written it yet, but there's a few facts that really focused my mind that you all might find helpful as well, that, you know, in the developing world, there's, depending on the country, you know, between, you know, 50 and 70% of people are working in one-person or one-family firms. So, they are retail and agriculture are two of the biggest ones. And so, when you think about the employer is the worker for the majority of workers, that kind of reframes everything because now it's not about, you're not firing yourself. Of course, the family farms might not be family farms. The retail might get taken over by a chain. So, they might, those firms might not be like that in the future, a little bit further out. But already, a lot of them are in small towns and stuff. They're in rural areas. So, it's not really going to all become chains immediately or get gobbled up by private equity in these countries. So, so then you're really thinking, if you think that there will be a lot of small firms going forward, then this, this automation is just making them more productive. It's allowing them to grow, and many of them are supplying things that actually are in, where there were capacity constraints before, and, you know, more local food, more local services, more goods. So, you're in that stage of development. This is just productivity enhancing.
The second piece is there's another category of workers that scales with the size of any country, and I think Nila set me up really well for this. Um, it's, you know, healthcare, education, elder care. The, the first two in developing countries are more important than the elder care, but they've got that too. Um, those scale with the size of the economy. And if you look rich to poor, they're way undersupplied in poor countries. You know, your nurses don't have a high school education. Your teachers are teaching science without any science education. So, the ability to upskill those service workers who then have positive externalities through your whole economy. And if you have agricultural workers, if you can all make them 5% healthier, that like, that like goes one for one into output, you know. So, I think those opportunities are huge. And so, when I, going back to like pattern recognition, a lot of the development economists are like, "Okay, here's another technology that's going to widen the gap between rich and poor." To pick up on Nila's earlier point, it could. It absolutely could, but it doesn't have to. And this is one of the few times in my lifetime where like a, a relatively small amount of dollars with relatively high ROI can actually improve the lives of like a billion people. So, this is like the opportunity of a lifetime. We may not take it, and we may, may fail to execute, but it's profound, and it's different, I think.
>> Can you be a little more specific about who we're, where for that opportunity to take? Who's using AI to do what to make that happen?
>> So, um, I think there's two for the small business. Just focusing on the small business, there's two vectors. One is that, um, app developers using low-cost software development tools of high quality will have AI-powered products delivered to small business owners through their mobile phones or through low-cost hardware, sensors, cameras, whatever. So, you've got cameras watching stuff go in and out of your little retail shop. So, now you have inventory, you have a bot in your WhatsApp, and now you suddenly have customer orders, and you have supply inventory and order management without doing a thing. So, I think like the advice five years ago is, if you want to digitize, you have to do A, and then B, and then C, and then D, and each of them is expensive, and by the time you do it, maybe it's obsolete, and like, "Just why bother?" Now, it's like the AI makes the adoption in principle so much easier. I think looking at WeChat is a great inspiration for what we're going to see, but even WeChat, a lot of that stuff is not as AI-powered as it could be. So, that's the first vector is like things that come and meet the small business owner where they are, in their context, in their language, with tailored, customized advice, bringing in local data, and overcoming all the frictions they face in their life, and integrated in their workflow. Huge upside. Farmers are way inefficient right now. There's not enough agricultural extension agents either. So, we can empower the farmer. We can empower the agricultural extension agent. The ex-agricultural extension agent becomes a technological extension agent and makes the farmers more productive. That's one vector. The second vector is, so that's a private sector vector, but it needs, it needs infrastructure. It needs mobile, and it needs an ecosystem. It needs financing and so on. It's got too many pilots and not enough implementation. The second part is government services provision. Government plays a role in education, health, and all that stuff, especially in developing countries. So, that's like philanthropy, NGOs, and governments delivering. The problem is actually getting it across the finish line. You talk to a government health official now, they tell you they've got a thousand vendors pitching them pilots. No shortage of pilots. We, because of the software coding, they can all, everybody, I can write software. Everybody can write software. So, now we've got a thousand pieces of software, but nobody's getting better healthcare. So, that's the problem there.
>> Okay. Thank you. That's helpful. Um, Nila, picking up on any of those themes, but especially, well, if you want to add ADP's perspective on this versus history, that would be great, but if you're, wherever you want to go with that.
>> Okay. Um, I hope the takeaway from this day and anytime someone speaks about AI is, it could, but it doesn't have to. I think that's perfectly put. Um, I also like Tami's comment about, yeah, we're still hiring software developers, just different ones than we're looking for, different skills than we used to. I think that's more realistic than not looking for them at all.
When I talk to, um, leaders who are like leading companies, big and small, 44% of the US workforce works in a company with under 50 employees, right? Three-quarters work in a company with under 250 employees. So, and then you scale that around the world. This is a small business economy, and that's not often appreciated. But, you know, they're struggling to hire people because they're looking for people who will work a 12-hour shift. Who wants to do that anymore? The, there are jobs available, but we live in a very fragmented labor market.
I'll give you a real, something that really kind of captured my attention. And I was looking at the median age of an HVAC worker, a plumber, and an engineer. The median age of an HVAC worker, critical to build a data center, um, has fallen by five years in six years. That's really fast from a demographic perspective. And it's not because there's a bunch of young people who want to do HVAC. It's because there's so many people retiring from that field. Wherever there is an infrastructure, not a digital, um, digital, uh, application, but an infrastructure application, whether it's in finance, there's not enough accountants. It's in caregiving, there's not enough babysitters or caregivers or educators or healthcare, or whether it's construction, there's not enough specialized trades. You have gaps, and you have supply gaps. And what AI does is it's funneling all of that. It's papering all over all of that and saying this is the only thing an economy can produce. And the biggest problem I have with it is, for what ends? Because if you destroy the worker, you destroy the consumer. And we're still a consumer-led economy. So, I have to question, um, the business model of destroying your customer base, because that's very different than the industrialized past that has given wind to AI. That's what makes me hopeful because if you take any of these scenarios to their ultimate, this is why I'm not a philosophy major right now. I realized when I was a philosophy major, if I picked any kind of canon, any kind of philosophy, the ultimate end was nihilism. And you can say the same thing about AI. I mean, if you take the worst, or even the best-case scenarios, you end up with no workers and no consumers and no economy because, and in my view, no society, because society is based on transactions and commerce and relationships. That's not, in my view, a very likely outcome. I think the outcome is we have to give up this notion that knowledge work is the only kind of work, and that AI is destroying knowledge work. Maybe knowledge work was on a path to extinction way before Chat and Claude came into the picture. Maybe we've set up our youth for an economy that was about to disappear and be extinct, you know, in 20 years anyway. Because that's what economies do. They constantly evolve. They constantly change. And AI is just one more ramification of that change. So, sorry to get philosophical for a moment. Um, I, I think really the task at hand is to define the task of AI, and I don't think we've done that really well from a humanist perspective. Um, my, my goal in our research and our data is to take jobs, break them up into occupations. I have a housing background, and we did something very simple in the olden days of statistics and econometrics. No one says econometrics anymore, but that used to be a thing. What, what we did is we looked at millions of houses, and we did something called hedonic regression. Everybody remember that? And that allows us, through millions of observations of housing listings, to actually provide the value of that fourth bedroom or the finished basement or the powder room in Palo Alto. It's got to be like $500,000 to have a powder room. You're able to do that because you saw that millions of times. And at ADP, we kind of do the same thing. We see millions of job listings, millions of occupations, and we see millions of wages. And our goal is to use that same practice, hedonic regression, to break down the elements of a job, the elements of a software developer, and price the value of the tasks so that we can tell you, an employer, like yourself, what are the, the higher value tasks, what are the lower value tasks, and how are those tasks changing over time so that we can go to a college campus, maybe not Stanford's, maybe my alma mater, Indiana University, and say, "These are the tasks and skills that you will need to be competitive in this industry over time." So that we can be very clear and specific with with hiring managers and with workers on where these tasks are heading. Um, I think there's a lot of empowerment in AI, but there's also a lot of blind spots on what the labor market really is right now. So, we're trying to elicit that with the data.
>> Great. Thanks. So, let me pick up on a theme that, that the last, that Susan and Nila just got to, which is inequality. And so, if we look historically over the last few waves of technology, that's always put pressure on the, you know, the, it's increased the level of inequality. And what worries me personally, I don't know about the rest of you, is that's led to a lot of political turmoil within the United States. So, let's focus on the United States for a few minutes. You started talking about other countries, Susan, but let's bring it back to the US. You know, as you look over the, the near term, because, you know, we won't go out to three decades yet, how do you feel about, um, you know, the distribution of, of skill and wealth, and how AI is going to affect that? Who's most at risk, who's least? And, and then, if you want to speculate on political issues, that's great, and if not, that's reasonable. So, why don't we start at that end and move our way back this way?
>> Yeah, so, I, I think we've touched on a lot of the foundations of your question already. That, um, from a, I do believe that the small business can be a great beneficiary, and that's very important for the political part of it as well. That is out of the narrative. And so, I, I think, um, we had a famous, um, politician visit campus a few weeks ago, Bernie Sanders, who gave a speech. He met with some of us privately as well, and that really focused me on, you know, he was, this, he was really talking about what he was hearing from his constituents, and the way that he phrased it, we met with that, you know, all the tech companies are making headlines by saying, "This many jobs are going to get replaced." And I can kind of understand why they want to do that because it's clickbait, and people want to read it, they're interested in it, and it also shows your technical leadership because your technology is so good, it'll get ahead. That makes sense. But it's at the same time, it's riling up fear, and that has political consequences. Um, and, and it's apparently not as clickbaity to talk about, you know, how it can help people. Um, if there was more public conversation about how it could help people, that might galvanize us to, here in the US and in other countries, to be putting pressure on our politicians like, "Why am I not, why is this technology so great, but I don't have good healthcare? I don't have good education?" The things that are important to me as a person, I'm not getting, and I'm powerless in this economy to get them. I have to afford a bigger house to move to a better school, and I can't do that. You know, this is what, you know, my grandma can't get a doctor's appointment. I can't get a doctor's appointment. You know, the, these are the things that are that are making people fearful and unhappy. And if we look back at previous shocks like trade shocks, um, immigration, um, COVID, you know, the, the fears that people have when people get scared of change, when things change and people are scared of change, they react, and they react strongly, and it doesn't always lead to long-term investments that solve the problem. Um, it can, and, and so I am very, very nervous that the fear gets ahead of the benefits, and then we incapacitate ourselves to solve global problems through electoral choices we make. So, that's, uh, that's really, um, where I'm coming from. And just to put like a little bit more of an economic lens on it, when I do this micro analysis, I, I, I really think hard for every, every impact on what are the bottlenecks and what gets in the way. One thing that gets in the way of benefits is that you need to test. You need to make sure it's safe. It's, it's good. You have good training data. You've evaluated your impact. If you are destructive, if you are a destructive force, whether you are a weapon like a military weapon, or you are a terrorist, or you are a cyber criminal, the trust and safety department doesn't slow you down. Okay? Legal doesn't have to review. You just do it. And a lot of those are very nimble, already digitized organizations that adopt new technologies very quickly. So, the fast harms are coming. If you've got leaders saying, "You're all going to lose your jobs," and tangible harms like, "Grandma just lost her life savings from a fraud and cyber hack or whatever," that's a recipe for not doing a good job investing in the benefits.
>> So, it's a great point. I mean, AI is incredibly unpopular. It's, it's almost as unpopular as most, uh, politicians right now. And then to your point, like, I think everybody in this room, I would guess, would agree with the following two statements: that is, we need to regulate AI, but if we do, we'll probably, there's a good chance we make it worse rather than better. In the pro, like, the ideal reg thing is to regulate it really well, but it's going to be very hard to do that.
>> Can I just, um, when I was at DOJ in the antitrust division, one of the things I needed to do was comment on AI policy in the US and the world. So, I spent two years redlining and putting comment bubbles on documents about AI regulation that had the, the, that said all the same thing: "This may impede new business formation. This may be bad for turnover. This is going to benefit incumbents and reduce competition." If you reduce competition, you won't get the benefits that this regulation is trying to create. And for the developing countries, it's even bigger because they feel helpless. So, they want to protect themselves, which, and have their values be respected, which they don't feel like happened in the past. But then they may not have anybody serve them, in which case then they are, they, they have no leverage, and they just, their laws aren't enforceable.
They they don't get the goals they wanted. So it's it's a you're kind of between a rock and a hard place in this problem.
Yeah.
Tami. Yeah. I I mean I definitely think that there'll be um a lot of really um large benefits acrewing to people winning and and and people also losing. I mean, one thing that I I see is that I I I'm I'm kind of struck by how large the variance is in how well people can use AI tools and how productive people can be.
Unbelievable. Yeah.
So, you know, one interesting thing is we have a a part of the interview that we do where we give people take-home and and we actually give them the ability to use whatever AI tools they want. And um and there's a thing you can do which is you can just give the take-home to an AI model and just let it do it and submit whatever that produces and the majority of candidates do worse than just like letting AI do everything and and submit that. And so and and that's like quite shocking like what's going on here? Um, I mean, I think people just fail to appreciate how good AI systems are and they kind of interfere and they kind of hobble or constrain what the the model is doing and like are very prescriptive about what it should do but but but like bake in bad design choices. And so um you know I I I just see that there's this huge variance in how well people can use AI tools and then that results in this huge amount of you know these these large um differences in how productive people are. Um and so I I I think you know there's there's going to be this um greater u kind of inequality seems pretty plausible to me. I I would say that like you know automation in general does broadly benefit people like even even though that there you know there are these kind of um transition costs associated with this transition and displacement and so on. You know if you just look at the history of automation and you know electricity and the combine harvester and things like this and a computer I mean those have just been broadly very useful to people and I expect AI to to also be broadly very useful.
Thanks. We're after Nilas I'm going to start calling on people. So if you want to come up to the mic please do and Nila take us away on this topic.
I I I think the concept of value ad is really important here. Um and what value the AI is adding not just okay it can do cool things but to what end is it doing those things? That's a really important question. And who's the ultimate customer? Um, is the customer other AI firms or is the customer the worker consumer? So, I'd like to go back to that primary economic unit being the worker the consumer. Um, in our data, we we we look at how people use AI and people who use AI daily, it's very interesting. They feel very much more engaged in their work. They're excited about it. It gives them passion. Um, not surprisingly, they tend to be in tech. Uh, but they also feel disconnected from their teams and they feel less productive, not more. Why? Because the things that make us feel productive, answering emails, going on sales calls, doing these wine things actually are being done by AI. So, what's left? The hard stuff. you know, figuring out the business plan, figuring out the new market, figuring out how to make money on all this investment that we put on AI when we don't have any customers because we killed all the workers. I mean, that's the hard stuff um that makes us feel less productive. And I am injust. I don't think we're going to actually kill workers. But um and so we have to change our measurement of productivity. We have to change our measurement of engagement. And the missing piece, the heartbeat in which you announced my my bio um of the worker is trust. And the way to a worker's heart is through investment and upskilling and reskilling. If we could do that smartly, we actually serve both ends of the economic spectrum, the worker and the ultimate consumer. That's value add and I think that should be top of mind. That helps with the inequality issue as well.
Thanks. Let's take an uh start over here.
Okay. So I wanted to dig deeper on Paul's question about the political consequences of this uh uh you know Susan you mentioned the trade shock that happened you know it's been started 35 years ago and um it overall made the United States wealthier but it created a lot of losers um and I think most people would agree that US failed in handling the transition costs for huge amounts of people and that's had extraordinary ary political consequences. It seems to me that this is going to be the disruption will be bigger, will be faster and may have more variance. Um, and you know, Nila, you mentioned, you know, nihilism. I think if this is generationally concentrated or concentrated in other ways, what sorts of social safety net and transition assistance should the United States provide? This seems to me to be a real leadership question for all of us. Um, because we failed it once before and we got to figure out policies that can, you know, address this challenge.
If I could take that one, the great news is we can all see it coming. It's not like it's a pandemic level surprise, right? And we have data and we have AI tools that actually could help with better outcomes. We can use AI to solve the issues of AI. But I would be remiss without mentioning tax policy that incentivizes capital over labor. So let's start there. Um that is something that we should have a conversation with about the incentive structure of AI investment relative to labor investment because that's tilting the deck in away from the worker. And if that's what society wants, okay, let's acknowledge it. But if if there is a an incentive that actually leads to scaling up the workforce, um let's talk about that as well. And I should say especially if a country is importing the automation technology then there's a big externality. They could hire workers. They could import an automation technology. The firm when the firm's indifferent, the country is not indifferent.
And so that's um you know keeping the money in the economy, keeping and and getting the positive benefits of your workers learning and everything else is is there. So I think you put that very well. Um I on top of what you said I think you know political movements come when there are concentrated groups of people getting impacted maybe a good or bad news we'll see which it is is that some of this is concentrated on richer people they're more politically powerful but also um have more of a built-in safety net. So, I think that's it's that's going to be interesting to see. And then if the robots come, like if what we're doing right now turns into faster physical automation, which I think is a few years out, then it starts looking like the China shock plus a lot more. So, I'm I'm fearful of that. I've been doing research to try to use digital tools to help people make transitions. I think we should be investing more in that enterprise because it's going to change from year to year what transitions people need to make. So we need to be good at the business or the practice of doing transitions not just one one-off programs.
Um, making predictions is particularly about the future. Uh so I I'm interested in whether you think a we should be doing more scenario planning uh and what types of scenarios you think we should be preparing for and what's an example of something that we could do so that if we lived in that future scenario um you know we we would have greater uh resilience.
Can you just clarify who you mean by we should be when you say we?
Well, why don't we start with the US? I I can I can I can answer this already.
Uh I mean that so um the scenario planning in general is is the tool that I'm using in a lot of my business consulting and in my teaching. And I think one one there's many ways you can interpret this but part of it is that I think people people's minds kind of explode when they think about all the changes and they can't really get them all at the same time. So then they fixate on one thing which may not be the right thing. So to me sen scenario planning is is partly is narrowing down what scenarios are actually plausible and in my view you can rule out a lot of scenarios like when I was in Washington everybody was thinking about the scenario where we could close down open weight models that was a massive waste of time I I mean I I told them that many people said it was a massive waste of time but other people said it was possible very senior famous technical people said it was possible to close down open models. And so we wasted a lot of time thinking about that when that was just a waste of time because of course China was going to make one. You know, it was it was it was 100% certain um that that somebody would. So rule out those scenarios and focus your attention on what is important. But the the biggest insight I bring to scenario planning is to think about the timing. And that's why I have been obsessing on thinking about bottlenecks and blockers and what goes fast and what goes slow because actually I do sit in scenario planning sessions that are focused on the wrong scenarios and they're also focused on decision irrelevant scenarios. Like if you want to sit around and say what are we going to all do on the beach to make ourselves feel meaning of life when the drones drop us daries? Like I I don't think that's a good use of time because if we get to the beach and drones are do dropping us daquiries, we'll have plenty of time to figure out how to make it meaningful. But like what I'm worried about is on the path to the drones dropping daquiries, some physical input is needed and somebody's got to choke up point on that physical input and we have a war or blow ourselves up trying to control it. So that's the thing I'm worried about, not what we do when we get to the beach. And so the scenario planning should be around like how do we how do we what is the timing of good and bad? What is the timing of decisions and what scenarios help us think about the decisions we're making today that we really need to figure out.
I I think also in addition to those types of scenarios, the the decor one would be my favorite of all time. Um we should run scenarios that are a little more introspective. What if our data is bad? What if we're told the wrong thing? Um what are what if our assumptions were not well played? What if our assumptions have changed? Um those kinds of introspective scenarios are missing I think in the discussion. I also think we should spend a lot more time on worker transitions and using that data, highquality data and looking at how workers can transition over time along with the tech, not just tech by itself.
So, we're out of I'm so sorry I'm being told we're out of time and I just want to give each of you a chance to wrap up with like what is su what if the optimists are right what does the world look like down the road? How how make us hopeful? What can the world look like? And you can pick your scenario whether it's two years, 10 years, whatever you want. We'll go that way, Nila. And it has to be quick because we're out of time.
The work the worker is empowered. The worker has several skills. The worker skills are adaptable, agile, and resilient. And that the growth and prosperity that ex is experienced by the world is shared.
Yeah. I I guess I uh am excited about just rapid technological progress and medical innovation and economic growth and um and for people to just broadly be a lot better off.
I think people's core needs getting met efficiently, cheaply, conveniently when and where they want it goes a huge way if you can keep your family safe, fed, clothed, educated, part of the community. Another thing I think has been I don't want to say like a waste of time but like disproportionate in all these panels and discussions I go to is everybody's talking about universal basic income which like most countries can't afford at a good enough level and and is not it doesn't actually make people happy. It's like goes along with the dairies on the beach. Instead, we need to be thinking about meaning and participation in society and the fact that there still will be work to do, but it's more like do we want to be like a king and surfs or do we want to have a society that um there there's still lots of work to do. Um, you know, it can it's just people don't have a very good life in one of those visions. Uh but it's not that people are sitting around on the beach. Um and so what what the question is can they participate meaningfully economically, socially and politically in this advancement.
Thank you very much.