The Annual Hicks Lecture 2022: Redirecting Technological Change, Prof Acemoglu (MIT)

um welcome to the 2022 hicks lecture uh sponsored by opposite economic papers and organized here in the economics department here in oxford it is named after sir john hicks who's the famous economist who spent 38 years of his academic life here at oxford the lecture is what has been taking place since 1984 not every year but that is also 38 years ago that it started um we've had over the years fabulous economist talking uh about all kinds of uh topics and we've had here um and let me very quickly give the surname solo carl de montevideo becca lucas de roll send it in shelling and many more it actually means it's quite appropriate to have um this year's lecture delivered by diary nas moglu um darren is well doesn't need much introduction he's uh institute professor at the department of economics at mit where he's been since 1993 and he has a such a diverse portfolio of research and makes a lot of impact in all kinds of areas both in thinking in terms of policy side and today he'll be talking uh about technological change under the title redirecting technological change applications to energy automation and ai the format of the lecture is essentially um darin will talk for about an hour maybe just under an hour and then we'll have half an hour of questions we have uh an audience in the room we also have of course um um first 250 people online already that will be able to put in the q a uh q a questions um put it in the in the q a function um during the q a i think i will read out the questions that are posted and i will also of course give the room the chance to give some um uh give some questions here for further discussion so without further ado darren please the floor is yours thank you thank you stefan thanks for inviting me stefan and james and uh it's thank you to all of you for joining us it's a true pleasure to be delivering the hicks lecture albeit remotely so i'm going to talk about redirecting technological change applications to energy automation and ai and it's actually you know in my humble opinion it's a good topic for me to talk about because it does bring two of the areas that i have been passionate about over the last 25 or so years political economy and technology but hopefully with uh a number of new ideas for at least most of you in the audience it is also over it is also overlapping with a new book that i'm working on which will be out spring 2023 uh not on the same topic but with an overlapping set of issues so for some of you there might be more uh reading on this coming soon if you are indeed interested beyond the academic articles so what is the main theme the main theme here summarized in this first slide which i'm going to substantiate and uh build on in in the rest of this hour and then take questions is that technology is much more malleable than economists typically assume so there isn't a particularly preordained path of technology it's not only that we don't know what types of innovations are going to come that's already of course part of the thinking of most people but the direction of technology which types of techniques which type of organizations we develop there are many directions many paradigms that are possible and several of these have very distinct implications not just for that productivity but about the distribution of resources for example who benefits who loses whose political and social power is exacerbated and who is marginalized and of course those are particularly visible in areas like climate i will also argue productivity and work but also about the interaction between these technologies and political structures such as democracy moreover i'm going to argue that there is neither a simple market process that will automatically select whichever path is better even if we could define what a better path is nor is there any other market or social process that generates that different paths are going to have benefits that trickle down to everybody so the distribution of gains that are embedded in some of those paths have permanent implications that could be quite important there are conditions under which which will talk about uh institutions and norms and as well as market incentives might generate uh strong tendencies for better paths or for experimentation across parts but in a number of critical areas today energy general production technologies and ai i'm going to argue that the global economy has not done that type of experimentation so there is a need for redirecting technological change and hence the last part of the lecture will be about how to redirect technological change so here's the plan of the lecture i'll talk briefly about the need to redirect technology in energy briefly because that's obvious the bigger part which i think is more original and so i'll spend about 20 minutes or so is what's wrong with the current production technologies so in this process i'm going to provide a new way of thinking of production technologies which really highlights that different ways of advancing a technological frontier will indeed have very different consequences in terms of who benefits who loses what we do in terms of work or nature or work and wages and inequality and then i have a lot to say about ai but given that time is short i'll probably rush a little bit with on this part though i have a lot of slides on this so we'll see how much of this i cover but i'll explain why there are various different ways in which ai technologies can be used but again we have made systematically wrong choices on how we use ai today and and going more into it then i'll briefly talk about a general framework about how direction of technology is determined emphasizing the role of economic and social factors i'll provide some evidence from the energy sector and then uh discuss briefly how to redirect technology using some of these lessons from the energy sector and i'll conclude all right on energy i don't think i need to amplify this point any further there has been a huge buildup of greenhouse gases in the atmosphere uh the amount of greenhouse gases does fluctuate naturally and has done so over the last thousands or hundreds of thousands of years but today the amount is you know more than two standard deviation beyond the mean and has already led to an enormous amount of climate change the general consensus is of course quite clearly almost really beyond doubt that this is because of industrial production that has relied on fossil fuels coal oil and gas but it's not really a natural resource problem it's really a technology problem the reason why fossil fuels are consumed is because we invested in technologies that use fossil fuels and in fact these technologies have advanced a lot much more than their energy efficiency so as a result the amount of energy we produce has increased together with fossil fuel consumption and i'll come back at the end and talk about alternative ways of organizing the energy sector and those are relevant for thinking about why there was a particular direction of technology how it has changed somewhat over the last 30 years and what else can be done in terms of current production i want to do a soft introduction to this topic because i think this is one where there's a lot of new material here that i want to share with you of course in a summary form but many of you will be familiar with the color facts of balanced growth which you know were never really exactly true but were sort of a good approximation to the growth process of western nations over the over many decades in in the west and what we have seen in many industrialized nations not least in the united states is a very sharp decline in the labor share going against one of those cultural facts so here i'm showing you the aggregate labor share in the us in the private business sector it's gone down from around 68 to just over 60 percent and if you do industry adjustment in the red again you have almost a 10 percentage point decline which is really unheard of in the uk us or other advanced economy stefan do you do you have a question please no this is not just people giving already questions i'm sorry oh okay okay that's it i saw the end up so okay uh so there are many reasons why the capital share or the labor share in national income could be moving around markups monopsony capital deepening but there is growing evidence which i'm going to share with you that this is because of changes in what i'm going to call the task content driven by automation especially what i'm going to call excessive automation and that's going to be the main idea about general production technologies excessively going in the automation direction so there are many new ideas here packed into this one slide that's why i'm going to spend about 20 minutes or so sort of unpacking them but before i do that let me say that the labor share is not something that we should ignore and not care about precisely because many other labor market objects that are even more important such as inequality move hand in hand with the labor share so during the period in which labor share has been declining you have also witnessed very large increases in uh in inequality in the us so here i'm showing you the evolution of real wages for uh uh for men and women of five education groups going all the way down from high school dropouts in red to workers with postgraduate degrees in dark blue and what you see in the 60s and the same is true in the 50s if you go back with different data sources is that growth was broadly shared that the real earnings of all 10 of these demographic groups were growing more or less in tandem and more or less at the same rate as productivity growth about two percent a year in real terms but from around 1980s you see a huge increase in equality both among women and men and even more strikingly among men low and middle education groups workers with some college high school graduates and high school dropouts experienced significant declines in their real earnings even workers with just a college degree now postgraduate degree that light flu you see that they haven't really benefited all that much from the growth process of the last several decades these are in some ways idiosyncratic u.s facts but the increasing inequality is not a you just a u.s phenomenon by using an omnibus measure of labor market inequality such as the genie coefficient the peterson institute and the oecd estimate that inequality has increased pretty much in every oecd economy but even more interestingly there are very common changes in the occupational structure or the social structure of these economies as witnessed by this figure which plots changes in the occupation shares of different types of jobs in the major economies of the oecd in red what i'm showing here is the change in the employment share of middle-class occupations those in the middle tersile of the wage distribution and those are dominated by blue-collar occupations and clerical uh and and sales-related occupations in in offices and you see that in pretty much every country in the in the oecd those middle class jobs have been in retreat and what's relevant about those jobs is first of all they are the you know bad rock of the middle class where there's a lot of upward mobility middle class economic security and so on but also those are the jobs that are that have been at the crosshairs of automation so specialized software based automation in offices accounting software uh inventory control you know word processing and especially databases have been very important for automation of office work and robotics dedicated machinery numerically controlled machinery have been very important for the blue color occupations so those trends together with this type of automation related facts motivate what i'm going to present in the next few slides so i'm going to argue that there is of course nothing new about automation and in all current and previous episodes of automation going to the horse powered reapers harvesters stretching machines in starting in the 19th century machine tools industrial robotics and now software and even ai what you have is a process of technological change that ill fits into the usual way that the standard textbook economic model conceptualizes technology so the way that most economists start thinking about technology is to either specify an aggregate production function of this form or some sectoral production functions of this form where you use capital and labor and technology comes in the form of these ala k terms which are factor augmented they make either labor or capital more productive and especially if you want to be in line with the calder facts according to the uzawa theorem you would actually put a lot of emphasis on this a labor becomes more and more productive it turns out that this really is ill-fitting for many of these technologies that i have emphasized which are you know not the only types of technologies i will emphasize the malleability and the diversity of new techniques and new organizations but these have been an important part of technological change and really it's very hard to make sense of why you know if you think of what each one of these is doing is that it's replacing some tasks that were previously performed by labor and you cannot think of that either as making labor more productive and in fact i'm going to explain why you cannot even think of it as making capital more productive you have to think of it somewhat differently and this is not just a sort of uh a sort of epistemological issue of how we conceptualize technology it has actually critical implications about the implications of technology what it implies for wages for work and for inequality so let me unpack all of these by first providing an alternative framework for thinking about production and then contrasting this framework which is of course geared towards making sense of these types of technologies to the usual way of thinking about technology so to do that i'm going to specify production as consisting of tasks so think of tasks as the units of the production process so if you are going to produce a piece of garment you know you need to perform many of the sub units of production and each one of them is critical for that process so you need to design something you need to produce yarn by uh spinning and then you need to weave and put that together stitching and weaving you need to go through a variety of other production processes related to coloring chemical processes drying and so on and then finally you have a huge range of non-production tests accounting uh inventory control transport wholesale retail marketing and it's only when all of these tasks are completed that a piece of garment can actually reach its target so i'm going to capture that by saying output let's say in a sector but for now imagine a one sector economy i'll come back to multiple sectors in a second at least in verbal terms and you need to perform these tasks and y z is the amount of task performance or task services of task z and then i'm normalizing these tasks and show you in a second graphically as well to lie between n minus 1 and n that's just a normalization so they are uh they they they they integrate to one and then you combine them according to some aggregator that gives you the output and i've taken that aggregator to be a constant elasticity of substitution one that doesn't actually matter all that much but with would sigma is the elasticity of substitution the key is that these tasks need to be allocated to different factors of production so for instance take a weaving task are you going to perform that using a skill weaver an unskilled weaver or a machine like a weaving machine which is what you know the power looms that really drove people mad in the beginning of the british industrial revolution because artisans felt that they were being thrown away from their jobs so mathematically y z is automatable for some tasks those between the bottom and some threshold i and has not been technologically automated for the remaining tests if you're not technologically automated then you can only be produced by labor here i have removed the distinction between skilled and unskilled labor i'll come back to that in a second and and then that labor is going to have some productivity gamma lz and i'm going to allow a general productivity booster for labor to mimic this thing here the old way of thinking or the classic way of thinking about technology but if a task is technologically automated then you have a choice between doing it with labor or capital and capital itself has a productivity gamma kz and then also is also multiplied by ak now using this framework i'm going to uh i'm going to show you how graphically you can envisage this and in fact do some of the analysis and so to do that let's consider this graphic so on the horizontal axis i have the task at hand and the vertical ask access is the cost of production and i'm plotting two curves this one here is the cost of producing a task with labor so that's the cost of a unit task is the wage that you have to pay to labor divided by the productivity of labor which was a l over gamma l or you can produce it with capital but you can only do that for tasks that are technologically automated goes to the left of this threshold i and that cost is going to be the cost of capital r divided by the productivity of capital a k times gamma keys the firm's problem is nothing other than choosing the lower envelope in the assignment of tasks to factors so to do that in particular i'm going to just trace the lower envelope here well that means you assign all of these tasks to capital but then you have to stop because those tasks cannot be assigned to capital and then you have it jump up and then you follow this so that's a simple problem but i want to next highlight what technological change does in this framework so let's start with the classic examples or the standard examples of technological change so what does an increase in al do well you can see that in this framework an increase in al is a uniform rise in the productivity of labor in all tasks now you see two problems with this or you may see two problems one is that it's very unrealistic there are no examples of technologies certainly none of the ones that i gave you in the previous slide but going more broadly there are no examples of productivity the technologies that increase the productivity of labor in all tasks even something very general like electrification would increase the productivity of labor only in a subset of tasks the second thing is that you see that when you do something like al you generate a huge increase in productivity so the cost of everything declines as a result of that so you know uh if indeed technology did take this neighbor augmenting form technology would be a big gift to humanity as you know for instance freeman dyson has said technology is the gift of god well we'll see whether it is or not but this con this conceptualization suggested it now if you look at an increasing ak it's similar it reduces the cost of producing all of these tasks with count but none of this really is a good descriptive fit to what automation technologies do so what does numerically controlled machinery do or what does the spinning machines did what do what did the spinning machines do they took away some of the tasks such as spinning or such as stitching things together away from the workers that used to perform them so in this context that would provide that would correspond to a shift of this vertical line from i to i prime which i'm showing in this figure now you see two very different things that were absent and those are going to be both critical for my discussion one that now there is a direct displacement technology directly takes away some tasks from labor gives them to capital or enables capital to be substituted for them and as a result workers are thrown away from these tasks they may get employment and other tasks but we'll see what the implications of this are second the productivity effects are actually much more minor instead of a big rectangle here or rectangular object here now it's given like something like a triangle and in fact if i make the orange line close to the blue line i can make that triangle very small so while this way of thinking about technology is all about productivity in fact the way i have drawn it purposefully is that there is no reallocation of tasks here whereas here this is really all about reallocation of its small productivity implications and then the final thing i want to point out is that you can also add new tasks here so new tasks are the opposite of reallocation uh away from labor to capital because now they're creating things for you know i'm focusing on labor intensive tasks so they're creating new jobs for humans so they are reinstating workers into the production process okay so now let me draw some of the implications of this framework so to do that and this is i promised last figure that has last slide that has any math you can solve this model and aggregate it up and you will see that when you aggregate it up output can again be expressed as a function of labor and capital as in the standard model and in fact it takes this form here and if you look at it it will have some similarity to this ak times k a l times l production function but with a crucial difference there are also these orange objects which capture the task allocation and in particular if you want to look at the implications and i'll do that by focusing on the labor share you can write the labor share now as consisting of blue terms those are the classic ones that exist in the standard model w the wage relative to the productivity of labor the cost of capital relative to the productivity of capital but then there is this orange term which is that the task composition is shifting away from labor to capital and such things and i will argue in a second that it's really indeed the orange terms that are both empirically very relevant and also quantitatively very relevant and one way of seeing that is if the sigma here the elasticity of substitution is approximately one the blue terms are completely gone so they won't even impact the share of labor so everything's going to be driven by this orange term and the orange term here looks very complicated but it's approximately n minus i so it's really how much we are shifting tasks in favor of labor by introducing new labor-intensive tasks versus how much we are taking away from labor with automation all right so why did i put so much emphasis on the labor share first of all because it links to the calder fact and shows some major changes in the labor market but secondly if you want to think about labor demand overall labor share plays a crucial role because you can write the wage bill what employers pay to labor as output divided output times labor share and based on that you can look at what okay i had lied i had this one more equation but actually i'm not going to focus on the mathematical objects here i'm just going to focus on the labels but if you want to look at what happens to the wage bill when you create automate when you undertake new automation technologies it really boils down to what it does to output and what it does to laborship and it's very simple it always increases output as long as we are in competitive markets without any distortions why because you would only adopt the automation technology if it reduces costs and a cost reduction is the same as a productivity increased but it also creates this displacement effect the displacement effect is exactly this thing here workers are being displaced and the displacement effect always reduces the labor share that's where this orange term comes in this gamma is decreasing in i as the the simple expression here shows and so you're always going to have a displacement effect that goes against labor now two implications first much more robustly than in the classic standard theory of production now we see technology of one broad sort these automation technologies always reduce the labor share independent of what the value of the elasticity of substitution is or other issues and secondly the impact on labor is never unambiguous there is a belief in economics which we can face it's both theoretical and historical foundations but i don't have the time to do that if some people are interested we can do that in q a but you can think of it as like a productivity bandwagon increase productivity and ultimately it's going to trickle down to labor and automation technologies even in competitive markets and this is going to be amplified in non-competitive markets in a second as i'll explain but even in competitive markets implied that's not necessarily the case technological improvement do create a productivity effect that helps labor because it increases labor demand but it also creates a displacement effect and the displacement effect can easily be larger than the productivity effect and in fact for a class of technologies with pasqual restrepo and i in our work which develops this framework we call so-so technology productivity implications are not great so when you replace cashiers with automatic self-checkout kiosks you know there's some gain in productivity some reduction phosphorus is not major so in those cases the displacement effect is going to overwhelm the productivity modus productivity growth going hand in hand with uh labor market negative implications only new tasks do the opposite they create a positive productivity effect just like automation in fact this productivity effect would be much larger because you're creating new things you know how much productivity do you get by putting and check out kiosk versus by creating something that did not exist like new design jobs or new engineering tasks that could be much more major but also it helps labor by generating a restatement effect because you're putting workers back into the production process so to understand this i want to now show a couple of graphs and couple of pieces of evidence to build the case so this is how labor share has evolved in the u.s

from the 1940s to the 1980s early 1980s or mid-1980s and you see that the labor share is actually fairly constant within broad sectors there's some fluctuations but it's fairly constant but if you look at the last 30 years when the labor share declines start happening then you also see that a lot of it is actually within manufacturing there is some in mining actually there's a lot in mining but mining is a smaller sector but there's some in services but really the most important part in manufacturing and why is that well that's precisely because you see that's exactly where you uh experience the automation technologies and this is uh i want to show i'm not going to dwell on this in uh uh to to in detail but using a variety of different measures of automation technology so here i'm using the robots adoptions uh and uh and here other automation technologies in detailed manufacturing industries in all cases you see that adoption of automation technologies is indeed strongly related to the declines in the labor share in contrast if you look at other types of technologies especially those that are critical for labor for example introducing new occupations or or the emergence skills that have been identified like new job titles those tend to increase productivity but are also associated with labor share increases so this is already sort of building this notion that different technologies have different implications for labor different implications for the distribution of resources and i'll come back to that in the context of inequality in a second but this is all at the correlational level so let me take two minutes to delve deeper into one of these technologies robotics so on the basis of my work with pascual restrippo what i'm going to present now is a very brief i'm sorry a very brief overview of what happens to labor and in a second inequality when you adopt these new technologies such as industrial robotics that are very much targeted automation and as a background you know industrial robots are not everywhere they're only in some set of manufacturing but for that subset of manufacturing they have been transformation they have been associated with very large increases in productivity so if you believe in something like the productivity bandwagon that it's going to help labor then you should see that as firms introduce robotics they should also increase their labor demand they should create new jobs or or some such thing so to do that we'll look at different local areas in the united states 722 commuting zones their exposure to robots is a function of how much penetration of robotics there is in a particular industry at the frontier technologically and how important that industry is in a given local area and here are the results so what you see actually the lights have gone off here one second this means i'm not moving my hands as much as i should because the sensors is not things i'm not here so if you look at more exposure to robots where robots have been adopted more often you see not higher but lower employment this is showcased by the industrial heartland of the united states like detroit uh lansing defiance city etc but it's not confined to it so if you cut those highest robotics areas you end up with the dashed line instead of the solid line even more importantly perhaps it's not just employment it's not that some you know welders are losing their jobs wages in that area is very severely depressed for the next 20 years so this is the change in log hourly wages after controls are included but without controls it's similar and if you look at what's driving this is exactly what you would expect it's people who used to work in tasks such as operators assemblers inspectors welders they are losing their jobs they are losing their jobs in routine manual occupations and these specific occupations so this very much already suggests that you know a we cannot ignore the distributional effects of some of these technological advances and contrary to the most optimistic read that new technology somehow will always trickle down in terms of benefits to most workers you see general depression of the labor market but this is really the the tip of the iceberg robotics is only one part of the automation technologies and the bigger effects of automation are distribution to do this in a subsequent paper pascal arrester point i we construct a measure of task displacement which is essentially well let's not look at the mathematical thing that we just explained in words is whether a demographic group where demographic group is defined by education gender age and ethnicity whether a demographic group specializes in automatable tasks such as routine manual tasks that i showed you in a second in industries that are undergoing automation so if you are a worker with a male worker with a post-graduate degree in your 30s you are working you typically specialize in tasks that have not been automated and the same is actually true if you are a high school dropout immigrant you tend to work in very manual low wage [Music] skill sort of service skill workers and it's unskilled service worker work which is also not being automated so but if you look at other demographic groups for example older male high school graduates they tend to work in many of these welding type jobs that have been automated very fast so here is a summary of uh what i want to show there each circle here is a demographic group and this is the amount of task displacement they have experienced from 1980 to 2016 according to this measure that i constructed and here is the change in their hourly wages and what you see and the color coding helps you recognize which group is what that there is a very strong association between task displacement and the changes in the u.s waste structure so about 70 50 to 70 60 to 70 percent of the changes in the us wage structure is explained by the fact that some groups predominantly used to specialize in tasks that have since been automated and here i'm showing you that these groups were not on differential trends in fact as i showed you already 50s 60s early 70s were a period of shared prosperity they were growing more or less in tandem with the more with the less less automation exposed groups and that's why the relationship between hourly wage changes between 1950 and 80 and task displacement after 1980 is completely flat before 1980. so okay this was a whirlwind tour of saying well there is a different way of thinking about production and it is richer in terms of what it implies about the consequences of new technologies and it brings out these distributional consequences of new technology but the question then is well if automation is so inequality generating and so important for you know making gains from production productivity not shared between capital and labor how come we had several decades of much more shared prosperity so to do that in a different paper with pascua restrippo we try to say to try to decompose what the different types of technologies were ongoing in the u.s so we don't find much of a royal

for this capital augmenting and labor augmenting technologies but two types of technologies are critical those that create displacement automation and those that create new tasks the reinstatement and if you look at the period between 1947 and 1987 what you see is that there's a lot of displacement generated by automation but it is almost perfectly counter balanced by restatement and that's why the labor share is constant that's why growth is broadly shared the same is true by the way in manufacturing as well very rapid automation because there are a lot of machine tools and numerically controlled machinery being introduced already in the 50s and the 60s but it is counter balanced with a lot of real estate but then fast forward to the more recent three and a half decades or three decades or so and you see that automation accelerates and restatement slows down in manufacturing for example you have about twice as fast displacement and automation as you did in the decades of the 60s and the 50s and almost no reason so put differently therefore the decline in the labor share and the inequality implications may be at least in theory we related to the fact that the technology portfolio of the us and many other countries around the world when you look at it from an international perspective has become much less balanced and much more biased towards automation but you might say well if that's the technology that gods of techniques give us perhaps there's nothing to do but i'm going to argue against that because of course redirecting technological change is predicated on the endogenicity of technology and in particular there is no need for the direction of technology to be efficient or even technology adoption to be efficient and so here i want to plant the idea of excessive automation in particular automation is excessive when it goes beyond what's efficient and it also creates significant externalities that are not internalized the significant externalities that are not internalized i'll come back to them especially in the context of climate but they are present in other contexts as well uh but let me try to explain why you would have technological change going in the automation direction beyond what's efficient so here is one easy scenario if this curve here is the opportunity cost of labor but because of bargaining or rents firms perceive the labor cost or the wage that they face is actually above that then automation is going to be comparing this curve here to the orange one a social planner that's utilitarian would like to stop here at this point where the blue solid blue curve intersects the orange curve but firms would want to go all the way to the maximum in fact they would go all the way to here now this is an example of excessive automation in two ways first of all you're deviating from what the social planner would like to do but second as a result of that you're creating this red area and red area is negative productivity so if you have here the productivity effect that's given by the green triangle but then the red triangle subtracts from that because you are actually automating tasks and using them using tasks with capital that's socially more expensive than the social cost of flavor so you're really not going to get much productivity benefits from automation but you're going to have all of the negatives in terms of labor inequality all right so now let me do a small course correction and talk a little bit about ai because i've taken more on the production technology side as i feared i might but we have created we have we have covered a lot of ground so i want to argue and i'm going to do this more more quickly uh so without going into the sort of the framework the mathematical framework and the evidence as much but i want to argue that despite claims to the contrary ai which is promised to bring all sorts of goods is actually an unusually centralized technology dominated by a few firms and a very monolithic vision of how you're going to monetize ai and as a result it is actually not generating as much productivity gains as one might have hoped and it's empowering large corporations with a variety of negative economic social and political implications to do that one has to look at the specific ways in which ai is used and there are many and since time is short i want to just focus on three of those actually four of those but very briefly and then i have a couple more in my slides but i'm going to skip those one is about ai and data collection data is the lifeblood of ai ai is completely useless the current vision on the current technologies of ai are not you know there are exceptions like alphago or alpha zero that are not so much based on data but most of ai is a combination of machine learning and big data and the question is where that data comes from who's compensated for that data how you acquire that data this intersects with privacy concerns but lots of other things and the proponents of the current direction of ai say privacy concerns are not so important and users are empowered to choose the privacy settings that they want so i want to argue against that and i'm going to base this on a paper by myself and uh and several co-workers ali makdoumi as a rashmalikian and also so imagine that you have a setting in which users have a privacy value either intrinsically or in order to protect their consumer surplus from price discrimination but also enjoy payments or prefer free services on platforms normalize the value of data to the platform to one and normalize and denote the value of uh privacy for users by vi for user on the key here is the data is social when i share data i'm not sharing necessarily just data about myself i'm also sharing data about my demographic group about my neighbors about my friends and so on so to do that you know you need to co assume that what data signifies the information data signifies is correlated across users so let's assume i'm not going to go into the mathematical details but if you want to visualize it assume a normal distribution with a correlation coefficient between the information of users given by rule and let's take a very simple example which i'm going to use to illustrate what you might call the privacy boondog two users one who is not very privacy conscious and is willing to share data for whatever reason so v1 is less than one but the other one is very privacy conscious feature so since v1 is less than one if the second user did not exist social planner would say sure of course user one should share her data with the platform and in fact in any market uh where you can construct with explicit prices or with free services the platform is going to capture that data even if the user is russian but when there is the social nature of data when user one shares her data she's also going to reveal information about user so to illustrate this let me take the extreme form where rho is approximately one so you reveal a lot of information about user so what that means is that there is a first order negative externality you are violating the privacy of user2 but it doesn't stop there it's actually much worse than that because what happens is that once user one shares her own data she's revealing so much information about user 2 that user 2's information becomes not so valuable to protect so even though user 2 would in an ideal world prefer not to share her data after user 1 shares her data she said well all of my privacy has been violated already so she would also be willing to share her data but once she starts sharing her data that now enables the platform to acquire user one's data for cheaper and cheaper in fact in this example as role goes to one the platform can capture both users data for zero price and that would be despite the fact that v2 can be very large or there may be lots of other negative social externalities from acquire acquisition of data in fact you can easily show that even if both users don't want to share their data there's a coordination problem between the two and they move they might both end up sharing their data again creating a data monopoly in the hands of platforms now this is not just bad for privacy reasons this same sort of forces also imply that firms that acquire a lot more data will become more and more dominant in the product markets that they supply so economists generally think that competition is the most effective way of controlling misbehavior by firms but as you create more and more data monopolies competition starts malfunctioning why because if one firm has a lot of data and has acquired a lot of data that makes other firms much harder for it to compete against and as a result they are neither a good uh sort of protector against the monopoly of the data rich firms but even worse in general in such environments this is another project uh you can show that pricing decisions are strategic complements which means that actually when one firm has a lot of data and can use that for price discrimination and charge higher prices that will enable other firms to charge higher prices as well so what you are seeing there for is yet another dark side of the data and i'm sorry i'm gonna have to get up and turn on these lights again these sensors are misbehaving sorry okay well we're running out of time and i want to sort of uh talk about other things as well but let me just say another aspect of uh this sort of uh ai's effects on industrial organization is what you might call behavioral manipulation as firms know and more and more they might actually know more about the consumers than they themselves know and that opens the way to a lot of different new forms of manipulations but i want to link all of this to the effects of ai in the job market as well so ai is a broad technological platform can be used for many things it can be used for creating new tasks actually we have some very great examples of that it can be used for reorganizing work in a human friendly manner or it can be used for automation and ai there's been a lot of talk of ai but if you try to measure ai you see that there's very little of it before 2015 in the u.s but there's a huge increase in ai for example as witnessed by vacancies for ai specific skills in the us from 2015 onwards so there's about a four-fold increase in those postings in the course of a few years but which are the firms or establishments that actually drive this well in recent work with joe hazel david otter and pascal restrippo what i find is that this comes almost entirely from establishments that have a lot of tasks that can be automated by ai and those are the fourth quartile and the third quartile depending on different measures that we use felton at all's measured suitability for machine learning or based on michael webb's work uh exploiting patent text in all of these texts in all of these cases about half of the u.s establishments that don't have many tasks that can be automated don't post many ai things but those that the fourth and the third quartiles do and when you look at the data and we do this much more detail much finer variation but here as a summary when you look at their hiring then you see that exactly those that are introducing ai technology slow down their hiring the other establishments continue to higher and those ones essentially are flat so overall the evidence is that ai even though it could be used in many different ways it could be used for reinstatement or displacement using the same terminology as i have used you see that it's going much more in the automation direction and this is not relevant just for developed economies right now we're seeing it in developed economies but i've also argued that ai is the quintessentially inappropriate technology meaning that it is a technological path uh you know it's it's it's it's a particular direction of technology that is developed in the west and as it spreads and shapes the global division of labor and is adopted in the developing world it's very unsuited to the conditions of the developed world because the real strength or competitive advantage of emerging market economies is they're uh semi-skilled and low-skilled workers and ai is about the substitution of algorithms and very skilled engineers work for those things so it will be another force creating inequality there are many other aspects of how ai is being used today monitoring and also the ad based model which i have worked on but let me not sort of talk about those since i want to conclude in five minutes and i have i could just about do that and leaving enough time for uh discussion at the end so these were all sort of to say that in the area of ai in the area of general production technologies and in the area of energy we have significantly different consequences from different directions of technology some of them are good some of them are bad and the question is which ones we end up choosing so for that we need a theory of endogenous direction of technology which is what a lot of my work over the last 20 years has been focused on so let me now summarize that in one slide or two slides imagine that we have two abstract objects a g and a b good technology and bad technology clean energy versus fossil fuel or excessive automation versus more balanced technology and the key is what is feasible and what are the incentives so what is feasible is summarized by the innovation possibilities frontier imagine that we can put some more resources are in these rich workers or research spending sg and that will give us more faster growth of ag or we can put more of them in sp and that will give us more of ap what will determine how the market allocates these things well three things one is expectations about what is feasible so that i'm going to capture by expected with imagine this h.i.g here is not

known so as the expectation of atg and eight expectation of aw government taxes and subsidies tao g and tau b and of course what firms perceive as the pre-tax profits from the two technologies paigee and pipe so under some weak assumptions and some market structure for research you can generate an equilibrium condition that says the equilibrium is going to be such that expected profits from putting more money in sg has to be balanced with the expected profits of putting more money in sp so now you can see three sets of things that are going to influence this technology how profitable is pi g versus pi b what are the policy distortions and what are the belief distortions there is no guarantee that the equilibrium is efficient even absent any belief or policy distortion why because pi g and pi b are not necessarily proportional to social benefits so for example if pi b is fossil fuels more fossil fuel profits are associated with more climate damage or if pi b is using ai for ad manipulation that's going to be very profitable for facebook or google but it's not necessarily going to be in society's objectives but in addition we can also have subsidies making things better or worse so for instance fossil fuel industry today receives about five billion dollars a year in terms of subsidies globally and of course there could be distortions in terms of beliefs what researchers find is the more attractive area for pushing research forward well now i want to quickly show you what redirection of technology could look like starting from a baseline and to do that i want to focus on energy and point out that despite the fact that you know our history with energy technologies has been one of failures especially failure to introduce carbon tax or other regulations there have been actually reasonably striking advances in some renewable energy solar and wind especially so here is irena's estimates of the cost of solar and other renewables and wind was already not so much more expensive than fossil fuels in the 1980s but solar energy was prohibitively expensive as late as 2010. but there has been a tremendous advance in both of these and actually some a few other renewable areas how did this come about well it was due to the redirection of technological change so if you look at what's going on is that during this period you have a run-up in the number of renewable and green patterns relative to either all patents or patents that use fossil fuel and this is in the u.s up to about 2010 and that's what you're seeing after the 2010 the decline as a result of these innovation stock increasing the same is true in canada france and germany even more so than the us as you would expect since they were the ones that provided even more funding and some carbon tax but from 2010 there is a reversal why is that well actually because from around 2010 natural gas becomes very cheap and starts killing the incentives for renewables so here is one figure to sort of make that case i'm showing you in blue now the same figure here green fossil green versus fossil fuel patterns in the in the car industry and sorry in the electric electric uh uh generation and and uh and the dashed line here is a natural gas price index so when alternative energy was attractive renewable energy was attractive people invested more and then when natural gas started becoming cheaper then you had complete reverse so in terms of the framework this really emphasizes this pi b versus pi g part of course there is some tau b and tau g but they're not huge right now but they have you know they have played some role but not as much as tai chi what about in the area of automation and ai well i want to argue and then i will conclude after this that there are several factors that have made the general production technologies and ai both of them go in the direction of excessive automation global competition for western firms have forced them to cut labor costs and many firms thought that the way to do that is actually automate so in some sense that has created a boost for pi b even more important and we can go into the details of this in q a the business models and the growing size of big tech has been a huge factor in both through the expectations vision how to conceptualize new technologies what they are about and also profit incentives most of the big tech have a business model that works on the basis of algorithmic automation and as they become bigger that's like a boost for pi b we have also had changes in labor market institutions as i said labor market need not be competitive and when it's not competitive then another thing that determines the direction of technology is worker power and worker power of course is not very good for automation and as worker power gets diluted that really acts as another boost for occupation and then finally most western countries especially the us subsidizes capital so if you look at the us the marginal tax effective marginal tax on capital was always lower than the labor which is taxed more heavily but it's become variety for a variety of reasons again we can go back into the q in the q a has essentially plummeted since 2000. today software and equipment capital used for automation are taxed at less than five percent if you hire workers instead that's taxed at 25 plus so therefore what i'm arguing is that the general direction of technology in all of these areas fossil fuel general production technologies towards automation and ai have all been biased in directions that are not necessarily socially efficient for ai these visions what ai researchers are really excited about is really important as well but let's leave that to the q a but if you look at what ai researchers are motivated by it turns out to be the wrong problem in some sense it's the it's the it's it's it's a it's it's a side effect of the turing test type of reasoning that cool ai research should be reaching human parity and it really distorts what type of ai research gets done both in universities and in in the context of the silicon valley and cambridge and other uh sort of ai hubs all of this also gives us a lot of lessons about how direction redirection of technology could happen in the area of renewables you see what it took there was some subsidies to energy which of course required a new measurement framework to be developed there was changes in social norms awareness among consumers then there was price incentives coming from the natural fluctuations of gas and oil prices but all of this was in a very ad hoc basis there wasn't in the context of solid regulation or carbon taxes globally and that's the weakness of the institutional context and you know what really is missing in all of these is this sort of broader institutional context for redirection of technological change so that has some political economy elements that's why i said at the beginning this is all about political economy as much as technology but again i'm going to leave that to the q a i'm sure some of you will have questions on that so let me now conclude you know i have argued you know at a high level that technology is much more malleable its direction much more flexible than generally assumed and this matters especially because different directions create different externalities different productivity benefits and different distributional consequences and the market process doesn't generally get that right many too many externalities and and then i try to illustrate these in the context of energy production technologies and ai i spent quite a bit of time in the area of production technologies precisely because that's where you can see how the type of economic model one uses biases the way that one thinks and i try to make that point by developing these task-based models and also talked a little bit about how to redirect technological change let me now stop sharing the screen and i look forward to your questions and comments thank you well well thank you very much dan um there is um we still have 25 minutes for questions i'm just going to turn first to the room here uh whether anyone would like to raise a question we have plenty of questions on the chat and q a as well i'm just seeing where the hand goes up because if not i think i will [Music] see whether i think jim jim malkinson has a question and so i'm not sure whether we can get jim to um yeah that would be great to hear jim's voice yes indeed um can you hear can you can you hear me yes absolutely it was really more a comment darren that um this argument about factor prices impacting the direction of technological change is something that bob allen argued about in the industrial revolution there's a big effect on what the technology was developed and where it was developed particularly in the uk it was just really that comment that so it relates to the same point you're making i think that's right thanks jim thanks for raising that yes uh well actually you know the the that idea sort of appears in marx then it's in hicks the theory of wages uh i should have i was going to make that comment since this is the hicks lecture but then as i was running out of time i forgot that point so uh it's it's actually a comment that i wanted to make and it's it's relevant because this is the hicks lecture but it's also hicks is the first person to make this comment that high wages would create a particular type of bias in technology and the and that has two historical relevances one is much less important but it's my own history of thinking because you know i started working on directed technological change in the 1990s and then after i started working on it i discovered pics and i first tried to make sense of hicks using the type of models that i was using at the time and

2022-03-10 00:11

Other news