Dr. Tom Mitchell, Carnegie Mellon University: The Impact of AI on the Future Workforce
Van Ton-Quinlivan: Welcome to WorkforceRx with Futuro Health, where future-focused leaders in education, workforce development, and healthcare, explore new innovations and approaches. I’m your host Van Ton-Quinlivan, CEO of Futuro Health. We’re going to tackle a big question on today’s episode that has major implications for workers and employers alike: how will AI change the future of work? There are experts who warn of the so-called “robot zombie apocalypse,” and others who see workers being better off for having AI if it is harnessed the right way.
We could not have a better guest today than Professor Tom Mitchell to explore these and other questions related to the impact of AI and machine learning. He co-chaired a U.S. National Academy study on AI and the workforce five years ago, and is currently updating that work at the request of Congress with the results due out next year. In his nearly forty-year career at Carnegie Mellon University, he established himself as a giant in the field, and in fact, founded the world’s first academic department of machine learning. Among many other prestigious positions and affiliations, Professor Mitchell is a member of the U.S. National Academy of Engineering and the American Academy of Arts and Sciences. Thanks so much for joining us here today, Tom.
Dr. Tom Mitchell: It’s pleasure to be here with you, Van.
Van Ton-Quinlivan: We’re so delighted. Before we get into the workforce issues, I wonder, Tom, if you can generally describe your work in machine learning and cognitive neuroscience and the applications of it in which you are most interested?
Dr. Tom Mitchell: Sure. I’ve worked in the area of machine learning, developing algorithms for machine learning, for many years and a lot of that research — even though it didn’t stretch from theory to practice — a lot of that research really is driven by practical problems and interesting scientific problems. You can think of a computer program doing machine learning as a program that can improve its performance at some task through some kind of experience. For example, we might train a computer program to diagnose medical conditions by showing it historical data about the symptoms that patients had, and the correct diagnosis.
Some of the applications I’ve worked on over the years that I’m most keen about…we’ve done a lot of research on brain imaging and trying to understand the human brain using machine learning on data that we get from the brain. In one study, we showed people who are in an fMRI scanner different words — common words, like celery, airplane, palace — and we’d record the activity in their heads with an fMRI scanner. We were able to train our computer program to then, given a new brain image, tell us which word that person was thinking about.
So, you could basically think of the machine learning algorithm as learning the subtle patterns of the spatial distribution of neural activity in the brain that indicate whether you’re thinking about the word computer or the word tomato. The significance of this is that it allows us to study interesting scientific questions. For example, is the neural code that your brain uses the same as mine, and we study that question by turning it into a machine learning question. If I train my machine learning algorithm on data from your brain, can it then successfully decode which words I’m thinking about? The answer to that question turns out to be yes, and we’ve had a number of demonstrations where we bring people into the lab who we have never seen before, and showing them different stimuli, our algorithm could successfully decode which word they’re thinking.
Another application that I’m currently very excited about is online education. Think of it this way: for the first time this decade, we finally have online computer systems that have taught literally millions of children subjects like mathematics, simple arithmetic, and algebra. Millions of students! That’s more teaching experience that these computer programs are collecting than a human teacher could collect in a 100-year career of teaching. The challenge now is to use machine learning to go through the step-by-step trajectory that each of these students took as they took the course, and to discover which practice problems are the best ones to show to which students at which points in the curriculum. In other words, we want to use that vast, collected historical data to learn to teach better.
Van Ton-Quinlivan: Both of your examples give me goosebumps because they tie so well to the current challenge we have, especially for adults, with technology gaining so much rapid adoption. It’s been hard for the workforce to keep up with new skills, and so by being able to use these types of technologies, could you foresee in the future that workers have a better way to learn, an accelerated way to learn instead of what they have available to them now?
Dr. Tom Mitchell: That’s a really good point, and it really is true that as technology makes the rate of change in work and the workforce faster and faster, the need for continuing education for retraining, for upskilling, for understanding and learning about the newest technology gets more and more intense. So as a society, I think we will have to depend on new approaches to education, just-in-time-education, to allow the workforce to keep up with these changes and I think the role that online education will play in that is bound to increase.
The interesting thing is that I don’t think robots will replace teachers. But what they will do is they’ll act as teaching assistants. For example, I teach a course on machine learning, and it will be difficult during our lifetime for a computer to replace all of the things that I do. However, I hope that during my lifetime, computers will help me by, for a simple example, giving the homework online and grading it automatically. But once you step down the road of say, a computer program helping in that simple way to deliver and grade the homework, you’ll quickly notice that you have a new opportunity: the computer could also adapt the homework to my needs as a student. Instead of handing out ‘here’s your homework assignment, it’s these ten questions,’ we could create new, more intelligent adaptive homework assignments that might give you the same first question that it gives me, but if I get question one wrong and you get it correct, then I might get a different question two then you do, along with some remedial information to help me understand the concept that question one was testing. In that way, once computers start even doing simple things like automatic homework, we open up the possibility for personally customizing the education to each individual student.
Van Ton-Quinlivan: This future that you’ve laid out for the learner of having personally customized education, it doesn’t sound scary at all. As a matter of fact, that sounds exactly like the education that we would want for our children or our nieces or nephews. And yet, when we talk at a high level about AI, we worry so much about the technology taking over our jobs. Tell us a little bit more on how you would coach us to think about these things.
Dr. Tom Mitchell: Maybe the way to begin is simply to summarize the main results of the National Academy study that we conducted a few years ago, and then maybe we can talk about what’s changed since then. At that time, we looked at the question of how is information technology generally impacting the workforce? As you point out, many people are concerned wondering whether computers will take over our jobs. What we found in that study was interesting. We found that there are actually many distinct forces that information technology is exerting on the workforce. For example, it is automating some jobs. If you’re a tollbooth operator, you should start planning for a different career. That job is going to go away because it’s very easy for technology to perform it.
However, we found that most jobs are very different from tollbooth operators. Most jobs are actually a bundle of different tasks. Tollbooth operators are kind of an exceptional job because there’s just one task involved and if we can automate that, the job goes away. But if you think about the job of, let’s say, a doctor…a doctor has to diagnose patients, has to come up with candidate therapies, has to have a heart-to-heart discussion with a patient about which of those therapies the patient elects to adopt, has to bill the patient, etc. So, you can think of a doctor’s job or a teacher’s job as really a bundle of tasks. In some of the research that we’ve done — before and after the National Academy study — what we found was that there are very few jobs that are going to be completely automated. Instead, what’s very likely to happen is that the majority of jobs will be changed by automation because computers will be able to have help with at least some of the tasks in most jobs. But we found almost no jobs outside of tollbooth operator where all of the tasks are going to be automated.
To come back to the example of a doctor, I think in the future doctors will receive more and more assistance from computers in checking the diagnosis, maybe suggesting diagnoses, checking the proposed therapies, maybe pointing out drug interactions, looking over the shoulder and suggesting therapies. But the doctor will probably receive much less help from the computer at the task of having the heart-to-heart chat with the patient. Similarly for teachers. Teachers, hopefully, will get more help from computers, but not in all the tasks that the teacher has to do and teachers will still be the primary motivator for the students.
Van Ton-Quinlivan: Your same 2017 report predicted that the advances in AI and technology would mean workers would increasingly need to be more creative and adaptable with stronger interpersonal skills as well, as opposed to — as you were alluding to the tollbooth worker – being good at doing repeatable, predictable tasks. How do you think that is coming to pass?
Dr. Tom Mitchell: Well, it’s interesting. First of all, the reason that those skills are going to be increasingly important is that, as we just discussed, computers will be able to assist with the more mechanical routine tasks involved in the job, but it’ll be much more difficult for computers to help with the aspects of the of the job like the doctor’s heart-to-heart chat with a patient, where it’s really a human-to-human communication task. For that reason, we see in the future which skills are going to be more important and which are going to be more supported by computers.
The ones that are going to be more important for people to excel at are the ones that the computers won’t, which is the human-to-human communication. Now, what’s interesting is that — if you look at even university curricula these days — there’s more and more of an emphasis on teams and how to get students to experience working together in teams. For example, here at Carnegie Mellon students who are freshman in our engineering school take a course, which is really a team design and building course, where they design artifacts that require multiple stakeholders to weigh in on what the design should look like. So already, you see universities really making a beginning of a shift toward more team-based projects and that’s all, I think, going to help with the education of people and how to interact with people in the soft skills.
Van Ton-Quinlivan: So, you mentioned the soft skills, you mentioned the teaming and collaboration. I also believe even in your examples of teaching the doctor, being able to work with the AI-assisted tool also means that you have to have the core digital literacy, the familiarity to technology. Those occupations didn’t need you to be technology literate, or digital literate in the past, and yet the new type of doctor, the new type of teacher that you’re describing, will have to be grounded in these skill sets.
Dr. Tom Mitchell: It’s true, and there’s a kind of aspect of that scenario that we touched on in our National Academy report which is the question of how will the computer work with the person? Will the computer simply automate that sub-task or will the computer assist the person in the sub-task? So, to ground this, again, we go back and think about doctors in the future. Will computers take over the task of diagnosis or will they advise the doctor who has ultimate responsibility? I think in that case, most of us would prefer that the computer advise the doctor — two heads are always better than one — and the doctor can always take the advice of the computer if it’s good. But depending on the job, depending on the tasks…for example, if it’s in assembly line work, there are going to be aspects of that that are simply automated instead, so the computer will play the role of automator rather than assister.
But the big determinant of how the future of work is going to play out is how we develop these technologies, and how we choose to adopt them. Do we, in particular, adopt computers as assistants that allow people to do their job better? Or do we use them to automate the task, maybe at the same level of competence more economically, and the future is really ours to define? It’s not a given. It’s up to us as a society to figure out which of those things we want to do and probably you could imagine either scenario. We could have the same kind of quality, but automated and so less expensive, or we could have better quality work being done, and people retaining the jobs. My guess is that it’s going to be the latter, but what we want to do is not replace people, but help them do their jobs better. But again, it’s a social choice. It’s a political choice. It’s an economic choice by corporations and by unions and even by the technologists who are designing the systems. In the very design of the system, you’re already making some kind of commitment about whether this is going to be a replacement or an assistant to the person.
Van Ton-Quinlivan: Tom, I’m curious, given that you have a holistic view of this field and the technology that is developing, if you were to counsel any of the listeners on what area of study to go into in order to be active in this field, I wonder if you could provide some advice and counsel?
Dr. Tom Mitchell: My advice is to go diverse. Do get a liberal arts education, as well as a technical education. Do not leave out the technology, because you will need technical literacy and the technology in the future, which we can’t really predict, is likely to require even higher levels of technical literacy than today’s computers do. So, I would say get a diverse education and include technology in it. And do what you love. That’s the most important thing. You’ll learn twice as much if you do what you love.
Van Ton-Quinlivan: One of the implications of technology on the workforce is that there could be a new way of getting educated that is perhaps different from our traditional two-year degree, four-year degree and graduate degrees. What do you foresee happening in the world of education and learning?
Dr. Tom Mitchell: You already can see this beginning to happen. In my own parents’ generation, it was common that people would get a degree in their teens or early twenties, and then they would work the rest of their life and that would be all the education they need. These days, we’re already seeing much more of a reliance on continuing education, on upskilling, on learning new skills to take on new kinds of jobs and it’s hard to imagine that trend won’t simply continue.
What that means is that we need new kinds of conceptions of education and how to make it accessible to people, and what are the organizations who are going to provide this education? Are four-year universities or community colleges also going to take on the task of providing three-month courses that will allow a welder to become an electrician because they have many of the same skills already but they need training in specific areas? What are the organizations that are going to put together and offer those kinds of continuing education?
There’s also a question that I’ve heard you bring up in the past, which is, if new training short-term training is required, then what are the standards for educational certificates? Who is going to develop those standards that will then allow employers to look at the training that somebody has and evaluate whether they have the appropriate skills for a particular position? So, it does introduce an opportunity for a much more diverse kind of education with different time constants than the usual four and two years that we’ve become used to. But it also introduces this kind of need for figuring out how to organize this in a way so that people can, for example, get a certificate in how to use spreadsheets in a way that is recognizable by employers, even though they might get that training from several different alternative sources.
Van Ton-Quinlivan: I would imagine that the pace at which AI is developing is even more rapid than other technologies and if you look at how IT and IT skills have played out, our higher education institutions have not been able to keep up. That caused the proliferation of coding boot camps, for example. I’ve wonder if in this future world, to keep up with the advancements in AI and what that means, there will be another infrastructure that gets created in order for people to skill and upskill like you’ve laid out?
Dr. Tom Mitchell: Yes, I think that’s a very good point. I agree with you on that. It’s interesting to think about the speed with which AI advances can propagate. If you think, for example, about the invention of the automobile, which was a great invention, but it took literally decades for that to propagate around the world because you have to produce a lot of automobiles and you have to set up an infrastructure for delivering fuel, you have to develop repair shops and highway systems. And so, automobiles had a tremendous impact on society and on work, but it unfolded over a matter of decades because of the physical infrastructures – plural — that had to be developed.
Now think about software. Today, if somebody has a brilliant idea and they write a piece of software then that software doesn’t take decades to propagate to all of us. We could have it all tomorrow morning in the app store. There’s no physical infrastructure, there’s no delay. We’ve already got the delivery mechanisms, and manufacturing of software is simply a copy command. I tell this story because it helps us think about the speed with which these kinds of software advances, including AI advances, can be sent out into the entire world once they happen. Somebody develops a translation system that can, in real-time translate French to English, and we could all have it tomorrow. In fact, that software kind of exists already.
Van Ton-Quinlivan: Well, that brings up the topic of the refresh that you’re doing now to update the 2017 study. What are you seeing, Tom, as some of your early findings, or at least some sub-topics that you’re focused on?
Dr. Tom Mitchell: If you think about what are the main things that have changed over the past five years, the glaringly obvious thing is remote work. We all discovered over the last couple of years how surprisingly effective we can be communicating with each other by video conference, and doing our work remotely and using computers. I think I was certainly surprised and I think most of us were surprised at how effective that could be. There’s still the question of to what degree can it replace real physical, face-to-face interactions, but there is no doubt that remote work is now a real alternative in many cases. If you combine that with thinking about the future of work that’s a very interesting impact. Remote work, I think, makes it easier for part-time work, and just-in-time self-scheduled work.
One of the implications of remote work is an acceleration of the trend that we saw even five years ago with gig kind of work. For example, Uber drivers. That was a very interesting development because it allowed people to self-schedule when they were going to work to schedule working eighteen hours a week if that’s what they felt they could do instead of forty-nine, and so I think remote work is kind of an acceleration of this trend toward jobs being increasingly available that can be self-scheduled and done at the discretion of the worker.
Now, there are a lot of questions people have raised about gig work and whether the working conditions are equivalent to full-time positions, whether benefits are being covered correctly, whether retirement accounts are being set up, and so forth. Again, those are social and political questions, but there is no doubt that a lot of people are taking advantage of this kind of work opportunity, this option. So, that’s one of the changes.
And other changes…the rate of progress on AI is even faster than it was five years ago. In particular, one area where we’ve seen really dramatic improvements in artificial intelligence is in natural language processing. So, we now have not only systems that can transcribe spoken words to the equivalent text, which we’re all used to — now we use them on our phones every day — but increasingly, we see systems that can answer questions, that can be trained on very fast quantities of text and that can then answer a pretty diverse set of questions. We see systems that can author text on their own, and even though that text might not be perfect, we’re already seeing an impact on jobs where these AI systems that have only been developed within the past five years are being used in white-collar jobs that involve writing advertisements, or manuals, or various kinds of text. They’re being used usually not to fully replace the human author, but to provide a first draft that can then be edited. So I think that is the second trend that we’ve seen is, AI has really developed more quickly than we expected in some areas.
However, in other areas, like self-driving cars…five years ago quite a few people publicly predicted that by now we would have fleets of self-driving cars, but we’re not quite there. To be fair, if you look at that area in technical detail, there’s also a tremendous progress there and we might expect that even though we didn’t make the 2022 deadline for self-driving cars, we might well make a 2030 deadline. I think it is coming.
Another trend is AI and information technology playing a larger and larger role in providing education and training to people. One of the unfortunate things we noticed about that is that these educational opportunities that are provided by, for example, online software that could teach you high school equivalency kind of courses, those educational software systems are used much more by people who are already good learners, disciplined learners, and they’re not as useful to people who are not already successful learners. So, even though it’s an improvement in education in general, it seems that it might be further skewing the distance between the haves and the have-nots.
Van Ton-Quinlivan: Maybe there’s a role for public-private partnerships there where the government can shepherd the application of the technology so that it’s helpful to a wider group of learners, perhaps?
Dr. Tom Mitchell: Yes, that’s a great point. I think there’s another big opportunity for public-private partnership if you think about what can governments do that can make a difference here. But a second thing is simply to get a better handle on — and make it possible for citizens to understand — what job skills are employers looking for, where there is an oversupply or an undersupply and if I want to get one of those skills, where can I go to get the training that I need? Well, one of the conclusions of our 2017 study was that we’re flying blind into the future of work. We really don’t have much visibility into which technologies are influencing jobs most prominently in the past six months. We don’t have very good visibility into what skills are in highest demand and short supply, and so that’s another thing that governments could do and should do.
Now, if you think about how can you get the data, it turns out the Department of Labor does collect a certain amount of data about the workforce. But the vast majority of data is in private hands. It’s in companies like Indeed.com, which has a huge number of online job opportunities. It’s in companies like LinkedIn, which has a huge number of online resumes. The government needs to, in my opinion, work more closely with these private organizations that actually have the data to provide and then reflect back to all of us the statistical summaries of where are the opportunities, where’s the need for new skills, and how can we get them?
Van Ton-Quinlivan: Tom, that’s a good call to action for the private sector and government to work together and look ahead at where skills are heading. I was wondering if you could wrap up today by giving us some insights on the competition that we face with China. As much as we’re debating the ethics of AI and the applications of AI and whether AI will affect our jobs, there’s competition for the United States on this technology, and I wonder what you see when you look at the two countries?
Dr. Tom Mitchell: There certainly is competition and it’s not just with China. The whole world has picked up on the idea that AI is going to have a huge influence on the future of their economies, too. Regarding China in particular, China is a rising powerhouse in AI, and the U.S. is also a powerhouse in AI. Historically, the U.S. has had — largely because of its university system — has had the upper hand in the rate of development of AI. When I say because of its university system, I mean because of the openness and the popularity of the U.S. universities, we happen to have the smartest people from all over the world come here to get educated, and many of them stay. And even those who go back contribute a lot to the research enterprise while they’re here.
We can choose policies that make it less attractive or more difficult for people from other countries to come to our universities, but I think we’re shooting ourselves in the foot if we do that. We need to have policies that make it more attractive, that continue to give the U.S. the leg up that it has had historically of making it attractive for people that come here and get a good education and contribute while they’re getting that education. China doesn’t have the same kind of reputation, doesn’t have the same kind of draw in its university system. But it has alternative systems that are clearly producing AI companies with very talented people that are going to provide a very active challenge to the U.S. going forward.
Van Ton-Quinlivan: Well, Professor Mitchell, we are so delighted that you could share time with us today and thank you so much for your leadership in this space in our country, but also globally. Any final words that you’d like to leave us with before we close?
Dr. Tom Mitchell: I really enjoyed the conversation. Thank you for what you do with this podcast. I’ve enjoyed some of the other podcasts that I was watching. I’ll just leave us all with this thought: currently, we have a very large number of jobs in this country that are going unfilled. People blame the “great resignation” or the lack of skilled candidates to fill those jobs. At the same time, we have 3% of the workforce unemployed. So, we have a very large number of jobs going unfilled, and a very large number of people going without work. Can’t we make a priority of the task of matching these up in a way that makes our citizens more productive and more economically supported? I think we can, but it’s going to involve some public policy and some intention to really making that matchup work. So, I just want to leave with that observation.
Van Ton-Quinlivan: It’s a great provocation. Thank you, Professor Mitchell. I’m Van Ton-Quinlivan with Futuro health. Thanks for checking out this episode of WorkforceRx. I hope you will join us again as we continue to explore how to create a future-focused workforce in America.