AI, Brookings Cafeteria, Michael Webb discussed on The Brookings Cafeteria

Automatic TRANSCRIPT

Welcome to the Brookings Cafeteria the podcast about ideas in the experts who happen. I'm fred dues in February of this year. I had the chance to interview mark. Miro without his research on how automation in a I are redefining work. Now he's back on the show to talk about his new research. Showing how artificial intelligence could most affect better paid and better educated workers also in today's episode Ottawa. Sanders a new post doctoral fellow in foreign policy who focuses on artificial intelligence and nuclear weapons. Proliferation you can follow the Brookings podcast on twitter at policy podcasts. To get information formation about and links to all of our shows including dollar and cents the Brookings Trade podcast the current and our events. podcast if you like the show please go to apple podcasts. Ask leave us a review and now with the interview mark. Welcome back to the Brookings Cafeteria Fred. HP here is always your last on the program. I'm in February of this year to talk about your report on automation and artificial intelligence and how they are redefining work and now you here to talk about a new report also which AAC Britain and Robert Maximum which is a closer look at what kind of jobs or affected by a specifically so. It's a nice way to start the end of the year with this really interesting thing research that you're doing. Can you first define what artificial intelligence is as opposed to automation. A I is harder to the defined because in some ways it's more emergent it's newer it's changing faster and it's not discreet or tangible or physical like a robot and it's not as predictable and many are Software which is all about computer executing controls these technologies operate differently frequently can use large amounts of data to actually learn more broadly honestly they are simulating what humans can do and what would be called intelligence. That's a huge debate. What about when this becomes actual intelligence but they are mimicking the ability to perceive to predict to problem solve to to reason to learn and planning and then I want to say one thing or work on A? I is a subset of the broader realm home of automation that we looked at and we have defined a for all intents and purposes as machine learning which is perhaps the most discreet most established form of AI and that is the form of Ai that uses huge piles of data to discover patterns within that. So I just wanted to make clear bator. We're doing to make this very big in challenging analytic Olympic issue more manageable by looking at one piece of it so it sounds vital to understanding this research approach this framework that we not think about a I the same way we think about an assembly line robot that puts a rivet into the door there moves it down or some other kind of machine itself that perform some action repetitively. This is a lot more complex than that lot more nuanced and and that's a great point because all of these technologies allergies have been I think unfortunately thrown together into one thing and that's because many of of the analytic efforts around automation in general have actually pulled together all of these things they include robotics typically league include software and they say they include a I think they probably don't fully include a and at any rate the AIP of the story get swamped in much. More established and for now bigger realm of factory robotics products for instance other kinds of robotics and then the huge role of software. So we're trying to pull out. Let's now truly try to the isolate and identify dynamics affecting a alone and that's relatively new. There haven't been that many efforts to do this. There is no accepted single way to do it but we think it's really important that we understand the specific technology Ravin lumpy in with a broad God to Amorphous Bundle of automation. Technologies will mark. What would you say is your your co-authors topline finding about artificial intelligence and its relationship to work? One is that it's distinct building when I was just saying and its distinctive specific way robotics. Oh Botox and our work showed this above all has had huge implications and impacts in the factory world world. I think what most people their first impression or I thought about automation is about factory automation. which is robotics? And they're we've seen the massive impacts on blue collar workers the other huge bundle that we've seen is the story of software including an increasingly big office enterprise computing packages whether it's Microsoft or salesforce. These have had a largely impact on the low and medium skill workers within the office context were maintaining processes in the office or work shows that. Ai has a different footprint. It's not that blue collar footprint. Only though we see a lot of okay I in factory and a lot of it will affect workers in the manufacturing sector but at the same time the the broadest biggest and really newest recognition we make is that these technologies are going to especially affect those those in the upper and upper middle aspects of the white collar labor market as you go. Above the Fiftieth Fiftieth and seventieth percentile of income. You seem more not less automation exposure or involvement. So these sir technologies that are going to be used by white collar workers managers relatively well trained and well educated and well paid people. So there's this bundle there's AI in the factory and then there's this white collar cast a very different profile when reading your report. There's there's a long section on the new methodology that you and your co authors employed to do this analysis and I think that's really at the heart of why this analytical framework is so unique league and it's a different methodological Lens than looking at see factory automation. Can you walk us through. What that methodology is well? I I want to call out our incredible. Edible Tartan around this work Stanford PhD student. Michael Webb who is an acknowledged expert on patenting and develop our pattern based approach here. Now this all comes out of need it is unclear how to get out the occupation patients involved in one way has typically been essentially expert study expert assessment and ultimately subjective views of how particular technologies impact particular workers and that has been helpful to the field but we think the uncertainty certainty goes up. When you look at a Because it's such a young field and we think they're in general problems with subjective assessment and so we've relied on this method that very cleverly elegantly in Michael's development of this matches the in particular Noun Verb Noun Pairings of words in a Patents patents for a technology in batches them and and seeks the currents of the same words in actual federal occupational descriptions so it seeking overlap of words. Words it's the most literal kind of analysis and in that sense very objective about where there seems to be a match between what what a I can do as defined by patents that are looking forward and projecting possible commercial use and occupational L. Distributions which are descriptions of what workers do and we find all kinds of overlaps in those overlaps of the basis of this work. And there's a chart in your report. Page ten that compares some extracted verbs and characteristic nouns artificial intelligence patents. Can you give an example or two from this chart so that listeners can get a a concrete sense of your homing in on what I find endlessly fascinating here. Let's just do a few words. If someone say within recognize image recognize face predict performance detect abnormality determine similarity generate recommendation. So in way we can talk about what a is theoretically but another way to look at it. Which is Michael? Webb's contributions to just look at what it it does as defined by founders of companies and creators of Ip who predicted so these specific capability ability are really good definition of what a I does is. I've ever seen find this list of verbs and nouns fascinating but it also gives us a key to unlocking. What occupation will be affected so an AI? Application or technology. See that generates recommendations. You pair that with an occupation that also has part of its job description to generate recommendations. You could then and say that that job function is to some degree could be affected by AI. Because technology is there on the path so I used that example. Because it's essentially what we do. We follow Amazon Prime Prom to check out a movie that's made by. Maybe the same director Swan we just saw. Aw or net flexes recommendation of a cool related program so those matches point us at work that could be done by. Let's move on then to some more specificity around what that work is and who may be affected and how they may be affected by mark and you I I kind of generally talk about where you see in the economy right now yeah and are sort of maps graphics and hotspots suggests this so i. It's in the factory. That's where these kinds of algorithms are detecting defects in product cycles. All state are controlling processes are identifying problems. They're doing things like that and supplementing the work of people in the factory. But it's also in this white collar workplace often in consumer interaction call response. Automated Services Consumer recommendations on that flex and that whole zone of activity seems to be highly involved in a applications optimization is ation prediction wherever that's happening farther up the corporate occupational distribution. Those are big things that are happening. We see a lot of these matches of verbs and nouns in occupations like marketing sales. Computer programming is an interesting personal finance management medical applications. I think a lot of us have been hearing about New Developments In radiology for awhile. Where a I can can do as good or better job of reading scams so those are the two zones and then this very dramatic curve when you're in that white collar world g more? The worker is paid up to the ninetieth percentile the more they will be involved with with a AI. For better or worse that's important to add when you bring up occupational wages as one way to kind of slice the workforce in terms of how. Ai Might Affect it. What about about some other characteristics of the workforce that you see like educational attainment the kinds of roles that people have within jobs demographics age gender absolutely so educational attainment right off really jumped out and hear? The results of our analysis are almost opposite or early worked on automation In automation broadly writ with that focus on robotics and software that I mentioned you. You see a heavy till to the underrepresented and lower income lower education groups. Think of young people working in their first job at McDonald's you know which begins now heavily moving into using kiosk ordering and other systems or the factory factory were heavily hit. This is different. The highest most affected educational categories. Ba attainment so these is are better educated workers who will be contending with these technologies men much more than women now and that's opposite from earlier. I think this has to do with the role. Men play in. Corporate structures. Men are more senior. They're earning war. They're better educated somewhat and are going to be using these technologies more or more affected by them very different pattern one year. And you know we can get into geography as well so the map of the hot spots. It looked at first glance a lot. Like the manufacturing assuring geography because manufacturing is going to be heavily involved. But then you start noticing. A lot of big. Coastal cities are involved as well as big cities. I think this reflects two things. Software and high tech is going to be heavily affected by non honest product but as a tool for us and then big business centers so on the AI. Map Much looks the same with the mid last in the heartland glowing red but then places like Seattle the bay area look highly exposed. And I think that's because because those places have a lot of high tech manufacturing they have big tech industry and then they have big companies and and clearly the corporate world is going to be suffused with these technologies. You just use the term exposed which I think is a vital concept also for people. I'm trying to understand this analysis. I want to couch it in an old saw in release automation. And that's a robot's going to take my job the literal idea. Yeah that robot is going to replace your job and your task in this report. You're talking about occupational exposure to Ai. Technologies can you talk about what you mean by exposure. The word exposure is important in this whole field and we've continued to use it and it describes places where there is this match of capability. The technology can do that is currently being done by human so in that sense the word points to perhaps a threat the possibility of the work being replaced. I actually prefer the word and used the word more of involvement because I wanNA keep open at this point the possibilities that these these relationships with technology can be positive can be supportive of work. I should say our analysis does not make a normative call of where the relationship of a I with the work in edits identified is either negative or positive whether it's a substitution or a complementarity. So we're trying to be studiously. DOODY ously neutral on that our partner Michael Web is a bit darker and sees more of these relationships is likely negative looking looking at the precedent of factory robotics and enterprise software in the office but strictly speaking our analysis is one of involvement involvement now we think involvement can mean positive or negative impact. But it certainly is going to mean flux change and disruption and I think we can count.

Coming up next