Becker Brown Bag: Artificial Intelligence and the Modern Productivity Paradox

we're gonna go ahead and get started thanks to all of you for joining us for this edition of bfi's Becca a brown bag series featuring Chad's Iverson my name is Sam Morey and I'm the executive director of BFI for those of you who are joining us for the first time BFI is a collaborative platform for the diverse University of Chicago economics community with nearly 300 PhD economists on campus and even more scholars engaged in research related to the economy having an institute that brings our scholars together around issues of common interest and to coordinate allows us to leverage this work in a way that we think can have real impact now with that I should mention to learn more check us out at BFI UChicago edu and with that it's my pleasure to introduce today's speaker chad cybersyn chad is the Eli B and Harriet B Williams professor of economics at the University of Chicago in the Booth School of Business his research spans several topics with a particular focus on interactions of the firm market structure and productivity I'm also delighted to share that chad has just accepted that he'll co-chair bfi's new industrial organization research initiative so we're very excited for all the work that will be coming out of that Chad's work has been published in several top journals and he's earned multiple National Science Foundation awards he serves as an editor of the journal of political economy is a research associate of the National Bureau of economics research and has recently served on National Academies committees and as the chair of the Chicago census research data center board here at the University or actually I should say down at the Chicago Fed hopefully soon at the University so with that please join me in welcoming Chatta Lisa makes everyone thanks for coming today it's good to have you here to visit about some work I'm doing with Aaron Olson and Daniel Rock who are both at MIT and what we're looking at is how artificial intelligence technologies fit into what we've called the modern parrot productivity paradox and I'll explain what that is in in just a second okay so there's two elements to the paradox one is on one side you've got a lot of technological optimism out there here are some quotes from folks you've probably heard of or read in the business press talking about this great new world we have entered where we can do things that we couldn't do before and sort of the sky's the limit over the coming years in terms of of where we're gonna go with these new technologies and this isn't pure puffery there are real hard metrics of performance that this optimism is grounded in so for example machine learning algorithms have recently become have seen their error rates declined to below 5% which is about roughly the level of an untrained human so once you get an algorithm better than a person you can of course start swapping out the algorithm for the the person so I don't know if you can tell which of those pictures on the left is a blueberry muffin and which is the face of a chihuahua but it turns out a machine learning algorithm can do a better job than you can there's another example of how machine learning algorithms work you you train up the algorithm and then you give it a brand new picture and it figures out as best it can what it is there's actual examples from a commonly used algorithm and it you know just give you gas yes or no gives you probabilistic guesses and you can see how well it does there on some of those examples on the left that's image recognition another closely related technology is a voice recognition that also is starting to beat human abilities in a lot of applications so again the same sort of potential exists there and very recently it turns out there's been considerable progress in protein folding prediction which I have recently discovered is a thing so a little background on this every two years at the CAS P I don't know if they call it a Casper or what but that's a big meeting every two years where folks who tried to predict protein folding which is really important for lots of biochemical processes come together and sort of lay out best practice and how what kind of progress they've made these this is the progress over the last four meetings and you can see on the right there is the most recent meeting which was held just last year and the top performer you can see is the blue line pulling away from the second-place performer the dashed line is what you would predict the top performer would have done based on the progress of the top performers in the three prior meetings so you can see there's a considerable increment to the progress that's been made in predicting protein folding and the interesting story is the top team at the most recent meeting was from deep mind so you and there were about 10 people I'll be attend people with a lot of money behind them ten people none of whom worked in this field came in and did better than hundreds of people working in the field around the world did over the same period using machine learning algorithms so again this is just another example of sort of the unique and uniquely fast progress that's been made in terms of applying machine learning and AI techniques in the field okay so that's that's the one side of the paradox which is this technological optimism the other side is the fact that actual productivity growth as measured in the economy is terrible okay so to give you some numbers we are more than a decade into a productivity growth slowdown in the US but not just in the u.s. basically around the world every almost every OECD country that's basically the wealthiest most developed economies of the world plus with a few years leg also the larger emerging and most of the larger emerging economies of the world to have seen declines in labor productivity growth so in the US the magnitude of that slowdown is showing up here from 1995 to 2004 labor productivity growth was 2.9 percent per year that means a a given worker hour produced just under 3 percent more output than it did the year prior ok so you get 3 percent more stuff with the same amount of labor inputs that has slowed by more than half since the mid-2000s 2005 to 17 is 1.3 percent per year we were supposed to have the 2018 data that's delayed a month because of the government shutdown but it's not going to go up much if you look at the first three quarters we're problem maybe you could bump that up to one point four percent per year but probably not so we're still in a in a slowdown and again this is going on all all over the place here's some just highly smooth data from the conference board's total economy database showing labor productivity growth rates going back almost 50 years now you can see the world total peaked right about the mid-2000s right when the slowdown started in the US and it's been declining since then ok so how much does this does this mean well labor productivity growth is basically the speed limit on GDP per capita growth you can't get long-run economic growth without labor productivity growth it's just that simple hey and it's it's all about one for one so if long-run labor productivity growth falls by a percent in half per year you can sort of figure that if that's sustained GDP per capita growth is going to fall by about 1/2 percent per year now for one year a two year that's not the end of the world but you can pound that over multiple years over a decade or two decades it gets pretty big so for example had the labor productivity growth slowed down not happened in the US had we kept after 2004 kept growing at that 3% per year rate conservatively GDP this year would be four trillion dollars higher than it is four trillion dollars per year so that's $12,000 per capita per year that we are quote missing because of the labor productivity slowdown if the slowdown continues another decade we'll be quote missing almost half of half of the potential output that we would have expected in the beginning of the 2000s so this stuff adds up and it's important and the issue is how do you have simultaneously these fantastic new technologies that are doing things that no one's been able to do before yet we still see really lousy productivity growth in practice that's what we consider in these papers and we come up with four basic explanations okay here they are one we call false hopes which is just the optimism about the technology those specific examples aside is just misplaced and we shouldn't expect labor productivity growth to speed up again based on new machine learning and AI techniques the second story is Mis measurement which is the opposite which these technologies are here they're fantastic and they are producing gains the problem is our ability to measure the Gaines has waned since the early 2000s and so the slowdown is just a figment of our Mis measurement not an actual decline in the rate of technological progress the third story is what we call the distribution dissipation story there's a couple elements to that one element is that the gains are real of these technologies but they have fallen primarily on only a few players and you could probably name those companies if you wanted to tell this story okay moreover those companies because of the nature of the technologies have to expend a lot of resources protecting them from being taken and applied by their rivals and those rivals in turn spend a lot of resources trying to take those technologies away from the leaders in that process of protection and and attempted appropriation ends up burning up most of the gains that we see would see otherwise from the technology that's the distribution dissipation story and then the fourth story is the implementation and lag story which is the technology is here it's real its potential is real but it takes time effort and complementary products and capital to actually gain the benefits of those technologies and so there's going to be a gap period where the technology is around you can sense its potential but you don't actually yet see those gains in the data we consider each of these in turn and while we think little bits of each probably exist in the data we think the first three probably are not the primary explanation for what's going on so if we think about false hopes it's certainly true you can come up with examples of past technologies that people have been excited about that didn't pan out for example Fusion Energi has been 20 years away for 60 years okay and in fact the first generation of AI in the 1980s ended up being disappointed disappointing after initial excitement about its potential so we are certainly circumspect about that on the other hand it's not hard to come up with what we think are pretty reasonable scenarios where you can close that 1/2 percent per year productivity gap with only a modest number of applications of these new technologies and again in what we think are realistic scenarios and I'll walk you through a couple of those in a few minutes okay so we're not we're not quite so ready to dismiss the potential of these technologies the miss measurement story is something I've actually worked on a lot I have a whole separate study on it basically the while on its face plausible you know we get these new technologies and they're free we don't have to pay for them so they're not captured in GDP yet we spend so much time using them and get a lot of utility out of them etc etc etc that story while plausible implies a bunch of stuff that's measurable and none of that stuff is true okay you see none of the tell-tale signs of miss measurement going on moreover if you're in telling miss measurement story it's not just well we don't measure output very well it is we started miss measuring output worse in 2004 and in a particular direction and around the world and it's hard to imagine that all of that happened okay besides all the direct evidence that I talk about my other work so again while no one saying measurement is perfect we just don't think that's what's explaining the four trillion dollars of lost output in the u.s. the distribution and dissipation story is consistent with some things you've seen definitely there's sort of been an increasing sort of leaders running away from the pack dynamic in a lot of industries and you might say well those are the few beneficiaries from these technologies the problem is you have to explain how this four trillion dollars of missing output somehow is being gained and expended but we're just not capturing it in anything we measure so it's a little bit like the measurement story that way and if you told me there were 40 billion dollars of stuff we're missing in terms of rent dissipation I might believe that four trillion dollars is a whole lot of stuff to just not even know is is going on out there I don't I don't think that that's very quantitatively plausible so that leaves us with the implementation and restructuring leg story that the technology is real but we haven't seen its actual productivity effects yet because for various reasons that I'll talk about it takes time for those benefits to actually accrue and show up in the aggregate statistics okay if this story is right then the paradox of forward-looking technological optimism and present and past looking lousy productivity performance might be a paradox but it's not a contradiction moreover it's actually two sides of the same story that you would expect there to be a period of relatively slow productivity growth before an increase in an acceleration in productivity from the new technology because there's this period of retrenchment where you are figuring out how to take advantage of the new technology installing it in business processes and then in inventing the complementary processes organizational structures in capital that go along with the new technology and all of those processes take time and so you don't immediately see the benefits of the technology even though you can sense its potential ok so this is what we end up being thinking is the most likely story for what's going on the paradox I want to be clear it's not just well there are four stories and we don't think it's the first three Thor so it must be this one that's not solely what we're basing this on we also have an affirmative case for the implementation lag story and I'm going to give you that today okay so what are those what is that case there are three elements first I'm going to show you just as a statistical matter that past productivity growth doesn't predict future productivity growth so the mere fact that we've had slow productivity growth for a while tells us nothing in a statistical sense about what we should expect productivity growth going forward to be ok so the fact things are bad now does not apply we should expect them to stay bad just as a statistical matter the second element are those back-of-the-envelope examples of achievable productivity growth that I talked about I'll give you a couple simple examples where we think you can get considerable increases in labor productivity growth from the application of these technologies even in what are speaking in terms of the economy are relatively narrow applications and then third we're gonna make we make the case that artificial intelligence might be the next general purpose technology I'll just define what a general purpose technology is but as you can tell from the term they can have big effects and I'll I'll show you what that means and what that's meant in the past in terms of productivity growth and general purpose technology ok so first the statistical matter so what we did is we took data over a 60-year period and computed year by year we went backwards ten years and said I'll write in the prior ten years to this year what has been the average labor productivity growth rate and then we compared that to the labor productivity growth rate of the ten years that followed ok and then we just rolled that forward over 50 years okay so that's what's plotted in this figure the past ten years productivity growth is plotted on the horizontal-axis the subsequent 10 years productivity growth is plotted on the vertical axis and as you can see there's essentially no relationship between the ten years prior growth and the ten years future growth there's a very slight upward upward relationship but it's it's actually tiny if you take the entire span of productivity growth rates in the data you imply a differential of around less than two-tenths of a percent per year in protic predicted future productivity growth and it's statistically insignificant okay so basically again the lesson is if productivity growth has been slow recently it doesn't imply at all that it's going to be slow going forward it doesn't it's not negatively so sloped so it's not gonna imply it we must speed up in any sense it just tells you nothing right you can see this measure a number of different ways this is labor productivity growth if you want to look at total factor productivity growth at adjust for capital intensity also you find basically no relationship a very small upward slope but it predicts basically no significant change in productivity growth there's the regression numbers and for the sake of time I'm not going to spend too much time on that's just the regret the actual regression numbers from those plots I showed you already okay so that's the statistical element of the case the second element of the case are those examples of productivity growth we think could come from the applications of these technology okay so one you hear about a lot I'm sure is autonomous vehicles okay the Bureau of Labor Statistics reports three and a half million Americans work as some form of motor vehicle operator we think it's plausible that autonomous vehicles could reduce this number to one-half million in other words replace roughly 2 million professional drivers of various types ok private employment is about 122 million people so if you're getting the same output so that the algorithms will replace the drivers you're gonna move the same amount afraid around etc so you'll get the same output with two million fewer workers off of a base of 122 million that implies a labor productivity increment of 1.7% now this process of replacing drivers doesn't happen overnight so you don't get those that labor productivity growth immediately if it occurs over a period of say a decade which we think is not implausible that would imply an incremental labor productivity growth over that decade of 0.1 7 roughly percent per year okay so that's one from one technology affecting you know less than 2% of the labor force we would get an increment of 0.17% per year over the course of a decade another example of something Erika's worked a lot on in his work or call centers there are roughly 2 million 2.2 million people who work in a call center of one sort of the other based on conversations with people in the industry we think it's plausible that 60% of these folks could have what they do done by a machine learning AI algorithm instead again that Bay off of the base of 122 million workers that implies a 1% increase in labor productivity again that's not gonna happen overnight if it takes a decade for this switch out to occur that's going to increment labor productivity by 0.1 percent per year for a decade so just with these two technologies that are affecting in total you know three to four percent of the workforce you've already got an implied increment to labor productivity growth of over a quarter percent if you come up with five six seven more examples like this and we think you can you've basically explained the protocol activties slowdown away or reversed it is maybe a better way to say okay so we think that quantitatively these these new technologies have potential that it could realistically get us back on a 3% per year productivity growth rate at least for a decade if not longer these examples are labor productivity growth you can also see increases in capital productivity coming from machine learning and AI applications so Google used deep mind to start running in an experiment run their data center HVAC basically and so here's the plot of their energy use before during and after this experiment where they let the algorithm take over for the humans in terms of running the mechanics of the data centers and you can see that energy use declined during the length of the experiment and then when they shut the algorithm off it went back up to its initial level okay so this isn't just about swapping out humans it's also about making the capital that we already have more productive as well I'll note that the calculations I just went through are simple here's a job we're just going to have a machine do it instead of a person productivity growth rate differentials however as I'll get to in a second one of the greatest potentials for general-purpose technologies like perhaps AI is that they spur the creation of complementary types of capital and those technologies also increase labor productivity growth so they can have big spillover effects besides just the direct replacement effects I already talked about so you could imagine for example how different retail could look if we had little tiny autonomous vehicles taking everything you ordered to your house within an outer hour when you order it that's just gonna re configure large swaths of the economy none of that's in the calculation we talked about with their autonomous vehicles but you might think that those will have their own productivity gains tied to them as well alright so the last element of the case for the implementation leg story is that AI might be the next general purpose technology hey economist Tim Bresnahan and Manuel trottenberg wrote a book a couple decades ago about general-purpose technologies and they said there are three defining characteristics of a GPT it must be pervasive in other words used all over the place it should be able to be improved upon over time and it should create these complimentary innovations that I was just talking about and we think that AI does reasonably check the box in each of these cases so in terms of pervasiveness you know when it boils right down to it machine learning algorithms are a prediction machine and prediction seems to have application across all sorts of economic contexts and so the pervasiveness doesn't strike us as something difficult for AI and machine learning to achieve in terms of its ability to be improved upon over time well it's got machine learning in it by Nature it's supposed to be getting better as it's used and in fact if it works as we might hope this will be the first capital that actually makes itself better rather than having people make the capital better and then third in terms of ability to spawn complementary innovations I've mentioned a quick example already but there's some things like it's perception abilities of AI cognition abilities built into AI our building blocks that work together in a complimentary way in other words each one makes the other one more productive and so we think also that the ability and plus you add that to that it's pervasiveness you can imagine lots of complementary innovation being spun off as a result of applying these technologies to new to new scenarios okay so fine let's suppose you believe the case so far that AI is this great new thing that could be used everywhere will get better and spawn all these complementary innovations where are those productivity gains well that's from the implementation and restructuring legs bit and there's two elements to that first you simply need to build up enough stock of this new capital for it to actually move the aggregate dial and that can take a considerable amount of time I'll give you an example in a second and second the real gains are often harnessed from the new technologies once these complements are put into place with them but those complements very rarely exists already and are just pulled off the shelf and put into service with the new technologies typically they need to be invented and implemented themselves and that process also takes time and it can take a long bit of time as I'll show you in a second third there's also a measurement issue that I haven't mentioned yet that can happen when those compliments or intangibles when it's a kind of capital that we don't measure well things like business organizational structures brand equity know how your production know-how stuff that doesn't go on a balance sheet is capital but acts as capital in production then you can actually get a measurement phenomenon where even after the gains of the technology are accruing to the economy you're going to understate that initially I'll show you how that works in a couple minutes and then later you'll overstate the contribution of the technology we call this the J curve and I'll get to that in a second so for all of these reasons we might not expect these technologies as great as we might think they are even if we believe their potential is real we meant and expect them to show up now or even next year or in a couple years depending alright so let me give you some examples of how long these legs can be with general purpose technology now we titled this talk the modern productivity paradox there was a prior productivity paradox the famous solo productivity paradox based on a joke he made in 1988 where he said we see the computer age everywhere except in the productivity statistics he was basically talking about early IT the way we're talking about AI now this stuff seems to be put into place everyone's talking about how great it is yet productivity growth at the time was not very good it turns out when he said that in the late 80s computer stock had just that year finally hit its long-run level of application in the US economy so it hit 5 percent of the total equipment stock in the u.s. in 1988 and that's about what it is today all IT kapos about 5% of equipment capital it took 25 years 25 years after commercial availability of computers to get to that long-run level ok and only 10 years prior to his joke it was half of that so even after a technology's around and is commercially available it can take again a quarter of a century to be installed up to the point where it's actually moving aggregates in the way you would expect it to move those aggregates in the long run ok so the simple process of accumulating enough of his capital putting it into place and production takes time and that's the first element of this story another example go back to an earlier steel general purpose technology and that was the electric motor or I like to say portable power cuz it's really the electric motor and the internal combustion engine together which were commercialized around the same time they were commercialized in the 1890s okay but if you went and looked at US manufacturers in 1919 again about a quarter of a century later only half were actually electrified the other half were still running on coal or water power now we know electric motors are superior technology to power manufacturing relative to coal or water how do we know that because no one runs on coal or water anymore they all run on electricity okay but even after again a quarter century of the commercial availability of this clearly superior technology only half of manufacturers that actually installed it in part so that's going to slow down how long it takes for these new technologies to affect affect aggregate productivity yeah so we actually did a calculation like that and yes Pro average productivity if you account for sort of intangibles that were put into place intangible capital is put into place with that kind of those production processes yes measured productivity growth would be higher it doesn't explain the slowdown though because it would have been higher both before 2004 and after 2004 so there is some missing productivity growth there it's not responsible for this slowdown but yeah that the idea is the same well to be clear some of it is showing up we are getting 1.3 percent a year productivity growth and it's actually faster faster manufacturing because manufacturing has traditionally automated at a faster rate than other sectors it's just that it's slower than it used to be okay so these things again take a lot of time let me ask you to do one thought experiment with me that I found useful and thinking about these things and that is to compare those two general-purpose text technologies I just talked about the portable power of the electric motor and internal combustion engine and IT early IT okay so what I'm going to do is I'm going to draw an analogy between those two technologies and the years 1890 and 1970 I'm going to say that roughly speaking 1890 was the portable power what 1970 was the IT and that's kind of about right if you think about when those technologies were both invented initially and when they were commercialized that's roughly I think an arguably parallel situation okay so what I've done here is I've plotted labor productivity the level in the US from 1890 to 1933 okay and I've normalized it so it's level in 1915 is equal to a hundred so everything's relative to labor productivity in 1915 okay and you can see what happened is for this quarter century long period where portable power is commercially available but it's still being installed at a relatively modest rate in manufacturing labor productivity growth is kind of slow it's actually about one half percent per year it turns out from 1890 to 1915 and then in 1915 there's a bit of an inflection point productivity growth accelerates roughly doubles to about 3 percent per year and it grows at that rate until 1925 at which time there's another inflection point and it flattens out again for about the next decade okay so that's just history now I'm going to give you a little more recent history again I'm gonna ask you to think of 1970 as being the IT analog to 1890 for portable power I'm gonna plot labor productivity in the u.s. from 1970 until today and normalizing the level in 1995 which would be equivalent to 1915 in this chart to 1995 to a hundred so I'm just gonna overlay the two histories on top of one another 19 1890 1933 to 1972 today here we go all right I found that to be a striking parallel you have about a quarter of a century 1970 to 1995 of slow productivity growth towards the end of that period that's when solo said I look around I see computers everywhere down in the productivity statistics well a few years after he said that labor productivity did accelerate 1995 to 2004 it was 3% per year but as we've already talked about after the mid 2000s its flowed back down again now that brings us to today but we have the extra advantage that we can go back to the earlier period and see what happened after the mid 1930s to labor productivity so I'll show you that now it accelerated again to 3% per year through the decade that followed okay so what does this mean it doesn't mean given that productivity growth accelerated in the mid 1930s that tomorrow we start growing at a faster rate it doesn't work like that but what it does say is the productivity benefits of a technology did not have to arrive in one wave give what it's got and then go away and never be seen again it can come in multiple waves and if you think AI is sort of potentially the second wave of IT that would be a nice parallel to what we saw earlier with portable power where there were basically two accelerations of productivity growth with a quarter century of slow growth before and an intervening period of slow productivity growth as well one more example of these long legs this plot shows well just focus on the blue line this shows the share of gaff owe retail sales gaff oh and it's a basically it's stuff you'd buy in the department store kind of things like that doesn't include automobiles or gasoline some stuff that goes into the broad-based retail numbers it's more specific things you think of when you go to a store to buy basically okay this is the fraction of gafe Oh sales in the u.s. that are made on line the blue line okay and you can see by 1917 now it's about 30% we're at about a third now okay so we're at a third of those kind of sales being online sales in the mid 1990s of course they were that was small it was really small I'm old enough to remember the mid 1990s when Amazon started people completely recognized that it could really transform what retail look like okay but it didn't actually transform what retail look like arguably until maybe a couple years ago I actually have another paper on this with Alli her tatsu that showed the changes that happen in retail between 1995 and 2015 we're not really about e-commerce it was about the rise of the super center and the warehouse club okay maybe in the last couple years with some big bankruptcies and otherwise you can say okay now Amazon at all are starting to have an effect on the retail landscape but that's 20 years after people recognized they might in fact some people thought they very well will have an effect on the retail landscape it takes a long time for those changes to happen it's not just about all the production side things that have to happen the reconfiguration of supply chains etc customers had to be retrained to think about how they do retail - you had to be comfortable giving your credit card number to someone on the other end of a computer that you never saw you had to get used to the idea that things would come to your house and be delivered and you'd wait a couple days to get something you wanted and that you if you didn't want it you could put it back in the box and the company would take it back and give you your money back all that sort of stuff people had to get comfortable with overtime and so all these things that need to happen both on the demand side of the market and the supply side of the market the reconfiguration of organizations within companies etc etc all of that is the sort of thing that we think add to these implementation legs that create the gap between when you recognize the potential of a technology and when it actually starts showing up in our measures of market performance okay so the one last bit I'll talk about is this measurement element which is even after the technology is starting to be implemented there can be a further leg and measurement between what's going on and what you see in the data and the easiest way to think about it is to imagine that a lot of the capital the compliments that go along with AI are actually in tangible kinds of capital they're not things that we see as capital put into the national accounts and say oh this is capital and make stuff we're just going to be thinking it'll be measured as an expense in business instead and if that's true you might think and this is all I thought about in Ischl II okay if this stuff this intangible capital makes output but we're not counting it as an input we should overstate productivity we get some output out of this thing that we're not you know we get numerator without having to put anything in the denominator of productivity but that's not exactly the right way to think about it because that intangible capital is actually an output itself when you make capital that's an output as well just as when you make final consumption codes so you're not just miss measuring the denominator you're miss measuring the numerator too and it turns out under plausible conditions and there they are I'll give it a quiz in a couple of minutes on this I'll just tell you the story rather than all the equations it basically under a lot of plausible scenarios it all boils down to what grows faster the investment rate of intangibles or the actual stock of intangibles and usually you'd think initially the investment rate is growing faster than the stock and then later the two flip sides well it turns out those relative growth rates are what determine whether productivities under measured or over measured and for a lot of reasons we think again investment rates are going to be our investment rate is going to be growing faster than the actual stock will be growing early on so you're going to under measure productivity early and over measure it later basically what's going you're making all this capital now firms are restructuring themselves they're trying to figure out how to build these algorithms into the production process etc etc that's an actual intangible output but we're not measuring it so we're understating the denominator or the numerator later we'll be getting output from that stuff and we won't be counting all of that as capital and the denominator will overstate productivity that's that seems to be what what's going on here's a simulation now it turns out the aggregate numbers aren't quite big enough to move the dial much if you take you know a good I think our best estimate of AI investments in the last in 2018 is about 80 billion dollars plus or minus 80 billion if you you know you apply some multiplier to that to get an implied amount of intangible capital investment maybe we've got a misstatement and productivity of a few tenths of a percentage point so it's not by any means explaining the slowdown or something like that but going forward as AI investments continue to increase and they have been accelerating at a massive rate from a very low base we might expect this to be a bigger factor going forward all right so to wrap up we think the implementation and restructuring leg story is the best way to explain this paradox between forward-looking technological optimism but the present fact on the ground that productivity growth is is lousy we don't have a lot to say about the particular time when these legs will sort of resolve themselves and we should start to see productivity accelerate again it might be next year but it might be five years from now it could even be ten years from now okay but we've seen this before as I've just shown you in history with other general-purpose technologies if you believe that element of our case in the past those DVTs have taken a long time after their commercialization for them to be to show up in productivity but they did show up and that seems to be our best guess of what's going to be happening going forward in the future all right thank you very much everyone appreciate you coming coming today [Music] [Applause]

Channel: Becker Friedman Institute at UChicago - BFI

Views: 1,319

Rating: 5 out of 5

Keywords:

Id: 0YjFcqXBnKs

Length: 47min 20sec (2840 seconds)

Published: Tue Mar 19 2019