AI-tocracy: The Political Economy of AI, with David Yang

good afternoon welcome to the director seminar series at the fairbanks center i am winihib and interim director of the fairbank center one of the functions of the fairbank center is to provide a place and a platform for scholars to come together to share their work and their research that are of common interest for all of those who are interested in china and one of the goal that we want to achieve in the coming year is also to encourage more young scholars to be part of our intellectual community therefore in the spring series we have invited a few young scholars and i call them rising stars in the area of china's studies broadly defined to come share their work with us today we are very thrilled to have david yang assistant professor at harvard economics department to present to us his uh i think very innovative and interesting work on ai procrasti the political economy of a.i um david's research focuses on political economy behavioral and experimental economics economic history and cultural economics in particular his study he studies the forces of stability and forces of changes in authoritarian regimes during lessons from historical and contemporary china david has his undergraduate degree ba degree in statistics and business administration from uc berkeley and also a ph.d in economics from stanford so without further ado i'd like to welcome david to present his work and for the participants um please type in your questions in the q a chat box we will uh reserve the q a session to after david's presentation however if you have questions for clarification that are important for david to address during the talk feel free to type it in and i will do my best to give the question to david so david welcome thank you thank you so much for for having me i'm very uh excited to uh present this this is a very much work in progress uh so this is hopefully going to be you know part of a research agenda that we'll have and very much would like to hear your thoughts and feedbacks uh as we'll progress a lot okay so the the title of this project is aitocracy and as you can see from the title uh we'll try to understand the relationship between ai as a sector and autocracy as a political regime uh um and this is joining work with martin barraja from mit and noam nookman from lse so i want to start sort of a bit sort of uh sort of intellectual background that there is this at least within economics and political economy in particular this conventional wisdom that there exists a misalignment between autocracy and technological innovation and this can come in various forms uh such as sort of exploration risks or lack of protection of popular rights overall speaking in autocracy uh induces sort of less investment from the innovator because what's the point of innovating if it's going to be exploited afterwards it can also render into sort of you know the autocracies themselves might hold up innovation because the innovation generates disruption to the society that might induce loss of political power and political ranks from memories and so on okay there has been a lot of speculations in the last couple of years in the public domain suspecting that such misalignment between autocracy and innovation may no longer be relevant in the age of ai especially as it becomes sort of increasingly important sector and there are sort of good reasons to suspect so the people have argued that ai has become part of autocrat's new toolkit that autocast has always been wishing to control the population and they have been doing so increasingly by predicting people's behaviors and ai has been sort of an incredibly important tool to to achieve that our own prior work has has sort of documented that you know states has been a big player in collecting huge amount of data uh from various functions and that data has been used by ai firms uh around the world and in china in particular and that those data that state has collected allow the firm to innovate not only for the government sector but the same data can allow the firm to be innovative for the private sector with generating sort of what economists call the economies of scope which can be important uh that benefits the the ai sector's innovation trajectories okay so what we want to do in this particular project is to ask the question and from from an empirical perspective do we actually observe elements that support a sustained equilibrium between entrenched autocracy and in a in an innovative ai sector and in particular i want to break this down into into two sub questions because for a stable equilibrium or alignment between autocracy and asexual to exist we should have benefits run in both directions okay so the first question we want to ask is do autocrats actually benefit from ai benefit from ai secretary and innovation and does the ai sector in particular the commercial ai sector not just the government sector but the commercial air sector benefit from working with autographs okay now answer to these two questions actually are not very accidentally obvious uh the first question do autocad benefit from ai you know despite sort of a lot of the speculations you know from news that you read these days we don't have very strong direct empirical evidence that suggests autocross demand for ai is directly politically motivated and we don't have to the best of our knowledge direct and causal evidence that ai actually helps autocross to stay in power and generate instability also part of the goal of this project is to provide such evidence and try to document this uh hopefully in a convincingly causal causal way that other guys potentially do benefit from from ai and demand ai in return for the second question does the ai sector as a whole in particular the commercial part of air sector benefit from the autocrats now you might imagine that autocrats take advantage of ais in the same way that they take advantage of many technology that help them sustain power but this need not be the case that uh or induce sort of the ai actually sort of having direct benefits uh from from working with autographs in generating more innovation okay even if the government's a sector do benefit and actually do sort of grow after the the government work with it with the ai firms the political you can imagine several reasons that political motivated investments of the alpha a sector may actually distort or crowd out commercial sector innovation because the pivots of the firms are focusing on to serving the government and that's actually going to leave a bit of a vacuum in the commercial sector which which which uh you know it's going to be the the lion's share of the market that that that if that's the case that wouldn't sustain sort of this this this equilibrium between autographs and ai sector as a whole okay so we're going to take a take a step back and and try to try to answer these two questions uh empirically and see if they are indeed the case now before i i sort of show you know what we're doing purely i want to sort of you know give you a very sort of overly simplified sort of conceptual framework or view of how you know if such equilibrium exists how they will exist and how how does it work okay so in a nutshell you know ai sector or innovation in the air sector is data intensive that you know generating more output from the uh uh from the ai firm's perspective requires access to large amount of data autocracy demands certain kinds of as software for political purposes presumably because it allows them to predict people's behavior or catch the criminals catching political prisoners and so on and in in the process of of using ai to process data and so on the autocross is going to collect a large amount of data the ai firms when they work with those autocrats to process the data as using artificial intelligence it's going to be implicitly receiving access to a lot of government data uh because the data is going to pass through from server for example when they have to process it uh and so this there exists this ecosystem between the autocad demanding certain potentially existing equal system between autoclass demanding certain ai software collecting large amount of data and that data can be beneficial to the ai firms now an important feature here is that when the data is now accessed by the government data is accessed by the ai firms and the firms presumably under certain contracts are required to produce certain services and goods for the government the firm could use the exact same data to train a different algorithm or or use the data to develop a slightly different software that's meant for the commercial market that's the government don't directly care about and this is what again we call the economies of scope that means that you know there could be a limitation to the to the to the amount of distortion that's going to be generated in the sector when the government data flows in into the commercial sector because the commerce sector commercial fish recognition i firms take an example and use the same data to do facial recognition supermarket checkout booths that the government don't directly care about but relies on the same data okay so so so in that to the extent that such economies of scope of government data exist and is large enough access to government data could outwit various disincentives for those uh commercial ai firms so to invest and innovate and that can generate again also part of the part of the the synergy and equilibrium okay so let me be more specific about about what we do and give you a preview of what we find just in case i don't have time to go through all the results i'm going to focus on facial recognition ai sector in china and using the sector as a proof of concept and as one of the very important sort of frontier ai sector in the world i'm going to document that autocross such as china do indeed benefit from ai that is in locations in china that have occurrence of episodes of local unrest that leads to subsequent greater investment of the local government in public security ai products and those increased investments in public security ai products subsequently suppressed future future unrest and generate greater political stability okay so the the ai software does seem to be benefiting from the local from the local governments in china in terms of for political purposes and local governments uh demand more of such uh as as a result and second ai sector does seems to benefit from working with the autographs here again specifically what i mean is is local governments the the ai firms who received contracts to work with the governments in china subsequently innovate not just for the government sector producing more ai software for the government purposes but they're becoming more innovative and productive in a broader commercial market that government don't directly care about that's going to suggest that this existence of economic scope that the sort of the data and other resources that these firms ai received by working with the government allows them to take that resources and take that data as an input and be productive in in a much wider market okay these two elements is going to suggest sort of a a an alignment and then a stable equilibrium where ai sector helps enhance the autocratic stability in terms of political stability and autocross demand for ai can help sustain the growth of aai sector and presumably growth of a much wider economy if ai become increasingly important share of the the economy and it can potentially bias the technical change towards the data intensive innovation uh which will discuss more in more detail uh uh towards the end of the talk okay so if i have sort of no remaining time in in and i'll also show you some some uh sneak peek of some of the the more recent sort of our our next project uh in the in this agenda to to look at sort of if this sort of this alignment between autocracy and ai indeed exist this is implied that countries such as china have a competitive advantage in ai sector and what are the consequences of this in terms of global ai trades do we observe china being a net exporter of of ai trade in the last couple of years and how do we think about sort of potential sort of policy responses and so on okay so so without further ado let me let me quickly sort of give you a a overview of this the various uh data sets that we use uh to to to to help us to implement empirical exercise and then we'll show you sort of the uh introduce the empirical strategy that we will use for for each one of the question uh that that uh that i post and then show you the results okay so to to to to perform sort of the the to answer the question i just post we have to sort of piece together uh you know uh broadly sort of three different kinds of data sets uh that help us to sort of paint uh different dimensions of of this relationship between the the government and then the sector the first data sets that we need to build together and we would would build this for from scratch for this project is is the sector is the data that that describe public security ai investments and procurement by the by the local governments of china this actually involves sort of piecing together two different datasets we're gonna we're gonna identify to the best of knowledge sort of the the close to the universe of all the fish recognition ai firms who have ever operated in china over the last decade or so that's that amounts to about 8 000 of such firms and this comes from sort of you know both a chinese-based uh firm database and then a us-based uh data center okay we're then gonna put together a sort of i think this is very close to the universe of all the public uh procurement contracts that the local governments of china has ever issued in the last two years there are about three million such contracts coming from the official source of the government procurement database of maintained administrative finance okay we're going to focus on the subset of such contracts that's issued by the local public security arms government so these include local police departments local sort of you know uh uh bureau that help maintain political stability and so on there are about a little over ten thousand such contracts uh that that sort of you know issue by local security public security arm of the government and being awarded to an ai from in our database okay so so that's gonna suggest so a particular local public security arm of the government is procuring goods that's been supplied by ai firm and very likely it's going to be a a facial recognition yeah product given that we we're looking at sort of all the facial recognition affirms uh here okay so that's that's that's the first thing it says i should note here that i want i want goes through uh this in in the talk today we could we could sort of as alternative look not just sort of the the fifth recognition yeah software that the local government is the public security arm of the government is purchasing we can also look at sort of ai capable of surveillance camera this kind of hardware that that's that's accompanied this uh this ai investments uh through these contracts and and so on and then most of the results i'm gonna show you is actually very similar if we look also look at the these sort of hardware purchases on top of the the software equipment okay so that's data space number one database number two that we need is some measurements of local political pressure in which the local government might be responding to when they're when they're investing in in ai and also subsequently the greater ai investment might help the government to suppress okay so we do this in two different forms the first one is a somewhat standard the the protest and and the occurrence of other episodes of political unrest this is coming from a global data center g delt and we use the china version of of this it give us about sort of ten thousand of such events that involves the political unrest protest various form of demand for rights and so on that occurred uh between 2014 and 2020 throughout china which is the period where we're mo the bulk of our ai contracts is coinciding okay uh our alternative uh is secondary sort of measurement of local political pressure and this uh which will interpret very much as potential preemptive sort of demands or a pressure for for public security the local government might be fueling which is the share of weaker minority population in a particular profession this becomes quite salient from a local government's perspective uh since 2009 uh uh uh and we would take sort of the minority share just from the former statistical yearbook again these are the the local uyghur minority share conceptually is quite different from the the occurrence of of episodes of public unrest because to the extent the local government might be responding to the local rigor manager is not necessarily responding to actual occurrence of the of the unrest but very much potentially due to uh preemptive and and sort of just just a high salience of of of of of eagerness to to to to stop any potential political unrest in in the future okay finally we we need to have a data set when we look at sort of the economic outcomes how ai firms are benefiting from the local uh from from working with with local public security arm of the government we need to sort of have a way to measure uh these firms innovative activities so to to for this we're going to look at all the software releases that this ai firm has registered with the ministry of industry and information technology in china which is required for them to release major software to the to the to the public we're going to categorize these software into into several ways uh we're going to categorize them by intended customer for example whether the software is used for by government such as traffic control and public civilian software or whether it's used for the commercial market for example some traffic control for a supermarket uh uh supermarket sort of you know checkout booths and so on which which personally argument government don't have a direct interest in we're gonna also categorize them by intended use for example whether the software involves components that relate to civilians or monitoring uh to look at whether you know that potentially pivot the the the ai from it to innovate in a slightly different direction and to do this categorization i i won't go into detail today but we would use a machine learning algorithm where we sort of manually trained uh manually labeled about 13 000 such software into different categories and then we would we would use the recurrent neural network to apply this labeling to the rest of the software uh where we think our algorithm is fairly accurate in a prediction okay so that's the three data sets that we have it data set that that links local public security arm of the government to their ai software purchases uh with the ai with these ais of uh ai software firms in fish recognition sector in particular uh various measurements of potential local political unrest or pressure of local investor local government may be responding to and finally the the the the the the ai firms sort of software output and the most important output of these ai firms okay so so now i'm going to get back to our questions the first question do autocad actually benefit from ai okay first i'm going to look at demand whether the local government is actually buying more ai software and more efficient recognition as of in particular uh suits public security arm of units of the government when they have sort of latent political pressure okay when we look at sort of the variation coming from the local weaker population uh the the unpurpose factor is going to be fairly simple and just look at dude localities with higher weaker population share experience greater uh public security investment in face recognition ai in us in in those localities we're going to focus on regions outside of xinjiang uh just because the wheelchair uh in young region is is a very different magnitude than the rest of the of the country okay so this is you can think about this as sort of any sort of factors that may be maybe triggering sort of local public security responses and whether that's inducing a higher uh civilians uh as software purchases okay to look at the actual occurrence of the local unrest we're going to ask the question do episodes of local unrest in one quarter in a particular place in a particular prefecture let's be precise these two greater public security air procurement in that prefecture in a subsequent quarter okay this is going to be fairly sort of high frequency uh responding uh responses by the by the local government we can do this in in various ways we can we can do a simple oh as progression where we we would do essentially a panel fixed effect specification controlling for location and time fixed of x but of course uh one need to interpret the os with various caution in terms of in terms of causality because of conforming factor that maybe on the background so we could also implement an instrument and variable strategy work on instrument the occurrence of these local protests with local weather condition the intuition is that holding fixed a lot of the potential local political demands on the margin if a large protest or large episode of unrest breaks out depends on whether the weather condition is favorable if it happens to be a huge sort of super rainy or super cold the protests may not occur just because the just because it's it's relatively marginally more constantly okay so we can we can use uh weather conditions uh in uh throughout china to to to to predict sort of protests in in on on a particular day and we can be hands off in exactly what the configuration of the weather condition in predicting the protest occurrence uh uh and using a lasso regression and and our first stage is it's very strong in the sense of you know very much aligned with with your prior extreme weather condition such as thunder hail or very high temperature for example leads to fewer protests in in a given place just because it's marginally making it more costly for people to take it to the streets okay we then take this first stage to average it up to the quarterly level and we can see whether you know whether this protest uh whether these protests that's coming from slightly more exogenous sources because of the weather conditions uh leads the local government to invest more in the in the fish recognition ai in the subsequent sector okay so let me get back to the to the to the to the results on on how local government with high share of weaker population in in a given prefecture uh uh their their behavior in terms of investing in public security ai investments what you see here in column one and two is that we find that if using the in in you know os regression places with higher share of weaker population relative to places with relatively lower share the local governments are spending significantly more in terms of investing in public security ai this relationship is actually if anything stronger for the for the male uh weaker population against suggestive of this potentially dependent uh sort of pressure of of preemptive public security of local governments this actual emphasis is not a generic minority share a sort of presence of minority uh sort of a phenomenon if anything so in columns two and four where we would put the non-weaker population shared side-by-side by a weaker population share we see a negative relationship between sort of you know higher minority share uh leading to less ai investment that's very much sort of it's mostly driven by the the places with high managers tend to be also poor and then local government just not spending as much uh in terms of public security and high technology private security but this friendship flips once you look at sort of the the weaker share population uh where we're again given the the the the you know the the public uh domain knowledge uh where local governments do worry about uh the public security and and in terms of you know let's start invest quite heavily in in in the official recognition ai okay to to look at sort of the actual occurrence of the of the of the unrest uh and and and and and political protests and how does that lead to to governments investing in in ai um uh what i'm showing you here uh in the plot is predicting the subsequent ai investment using the previous uh uh using using the the the protests that occurred uh in in the in each of the quarter either leading up to the protest investments or after the investment zero is sort of the immediate quarter of quarter after so so the way to read this table uh this this figure is that in the quarter where a large amount of local unrest took place in the subsequent quarter we observed a a a sudden increase in local government's investment in facial recognition ai by the public security arm this did not occur previous to the protest occurrence and it didn't sort of you know lasted longer than the more than a quarter so it's a quite sort of fast-paced contemporaneous responses where the governments are responding to the last quarter quarters protest occurrence and and starts to invest more uh in ai uh you know that was the that was the ols progression if i were to do this using instrumental variables just looking at the part of the the protest occurrences coming from sort of the marginal good weather conditions or extreme weather conditions that makes them a discard protest we see a pattern that's very similar to the os version where a high amount of protest uh occurrence uh it indeed seems to lead to more ai investments by the local government okay this is to say that local government demand more ai when there's unrest or sort of threat from risk happening in in the previous period now do those ai investment actually help local government to to to suppress the subsequent unrest and that that's the second question now this to answer this question purely is is actually not super straightforward because there's going to be a lot of autocorrelation across protest occurrence over time so just looking at overtime variation uh we i just show you that the protests you know local governments are responding to high protest occurrence in the previous period and high protest occurrence is going to going to correlate over over time so this is going to this is going to complicate uh uh how we'll look at sort of how the pro the aim investment is changing subsequent protest occurrence so we're going to come with two similar but but different strategies to look at the you know the effect of ai investment on protests the both are related to whether a investment is tempering protest occurrence with respect to a range of factors the first one we could look at is whether you know we see a possible relationship between good weather condition and protest occurrence protests are more likely to occur when the weather is good than when you when you're more likely to to to encourage people have have large gathering we can ask the question does the presence of high investment in public security ai temper this possible issue between whether good weather condition and protest to make it sort of less responsive to good weather we can we can do this to to do a similar exercise with respect to whether protests are are more are less likely to spread uh when the when when a certain place has has ai presence you know there's going to be a generic uh a pattern that protests tend to spread across prefecture to prefecture potentially because they share sort of uh latent political demands and we can ask whether a particular place with high level ai investments by the local public security of the government when surrounding areas are protesting uh this the the high aim you invested to the local uh local prefecture uh uh has has higher has a fewer smaller chance of of being spread over uh by the near by the by the protest in the surrounding region okay so so this is what the the regression results that that that that that sort of answer the the the exercise that i just described the punchline here let me let's focus on on panel uh a which is looking at the elasticity of protest occurrence to uh to the local weather condition we see a very strong positive coefficient on good weather across any kind of local political unrest or breaking it down to various different type of events such as protest demands and threat and so on we see a possible user suggesting indeed sort of the good weather on the margin is more likely to induce a occurrence of those events but the interaction between good weather and ai is negative that suggests while the ai investment is not killing off all the possibility between between weather and protest occurrence that really should become milder in this in the sense of protesters are less likely to occur even if the weather condition is favorable uh in the in in in a particular region okay we see similar kind of tempering effect when we look at the protests spread over space where when a nearby space nearby prefectures protesting uh even though the places are more likely to also protest in in in the same period but but this sort of influenced by the nearby region is becoming significantly less likely to happen when a particular destination prefecture has a high amount of public security ai investment again suggesting this is pointing the direction where when local government the local public security on the government is putting in a lot of investment in in fish recognition ai it actually helps in the sense of you know very very rapidly in in in in in in quarter to quarter we start seeing protest occurrences start to be tempering down and presumably this is exactly what the local government wanted when they put in that investment to to generate more political stability okay finally it's just this is quite quite suggestive but we see some evidence that the investment of ai technology by the local public security uh arms of the government actually shifts the technology of social control and public security broadly speaking before we look at the hiring of the local police we see that for the places that have high level of public security ai investments in the next period and here at period is is is is the next quarter local governments are hiring less police forces so in some sense the ai is substituting away from the from the ground force of the police and among the police that they're hiring they're more likely to hire office police rather than for the police on on the ground again seems to shifting away from from the putting a lot of the police forces on the on on the on the streets versus with this more technologically aided uh kind of police operation we also see that those public security ai investments is leading those local that those prefecture to have less political prisoner in the next period suggesting a more targeted sort of uh crackdown or or political oppression there's not such a sort of fewer political crackdown but just potentially sort of a a more targeted crackdown uh potentially aided by again by technology and this is very particular in the in the in the political domain and we don't see sort of a a a such an active region between ai investment but local public security with respect to economic prisoners in the subsequent period okay so that's that's the first part of of the equilibrium do autocrats do local governments in china actually benefit from ai the answer seems to be yes they demand more of ai when there's pending political unrest uh and and such ai sort of helps them to to achieve better political stability in the subsequent period now what happened to those ai firms who who you know help the local government or or supply ai software to the local governments uh do they do better in terms of their innovation to to answer this question i'm going to use a a triple diff triple difference in differences in pure personality would would that boils down to is that we're going to compare the cumulative software releases of ai firms before and after they receive their first politically motivated sort of public security contracts with the local government relative to those who received a less political motivated contract okay so let me be more specific here so the so the the three differences that we're going to exploit is that we're going to compare a form a firm before and after so we're going to put in firm fixed effects so so there's always going to be the same firm and we're going to look just a quarter before and a quarter after the first contract arrives to the extent a different firm received their first contract with the local public security government at a different time we can also control for the overall secular trend of this industry and explore that the first contract took place at different calendar times for different firms and finally to the extent that political motivated contracts might differentially benefit from certain local uh local ai firms because if that might involve sort of you know the local government and providing more data to the to the to to the firm we're gonna look at sort of the comparing the the the the contracts that are issued by a local government that had just experienced local unrest in the previous quarter uh or local government that that has a lot of a relatively high share of weaker minority in the population relative to working with local government that hasn't recently experienced any local political unrest so or or any kind of generic public security uh contract in that sense okay i want to show you walk you through this first show you're just sort of what would a average public security contract do to the ai firm i'm importing the government plots of different categories of the software right after the firm receives uh its contract so so i'm going to start by by looking at sort of this is not actually ai software this is just data complementary software that allow the the the firm's releases and development of such software that allow themselves to store and transmit and calculate and process large amount of data this will suggest the firm probably received you know to extend the firm need to develop new software to store data that suggests the firm received a large amount of data that previously hadn't had access to okay so the way to read this figure is that what important you hear is within a particular firm relative to the period before relative to the period where the firm received this first public security contract with the local government which is the red vertical line here the number of software that has released this every quarter leading up to the contract arrival and happened and after the contract okay so what you can read here is that there is no particular trend leading up to the contract arrival but as soon as as soon as the first quarter of the firm has working with local government uh public security armed government the firm starts to release more software that help them store data and and transmit it suggesting this there seems to be a quite a large amount of government data exchange tenants implicitly uh because of because presumably because the firms are asked to process the data for them for the government okay now if we look at not turn our attention to ai software the software the the ai from developed for the government such as traffic control population control uh publication accounting kind of software recognition software for the for the street uh look at that segment of the software i could do the similar exercise look at the firm's releases of the software leading up to the first contract arrival and then for each of the quarter after the contract arrival what you see here is that almost immediately after contract arrived the firms start to produce more for the government but this is largely potentially as you would expect mechanically because the firm just are contracted to produce something for the government because of the because of the service provision and the firm is probably pro you know literally producing those software that the government is is contracting and asking for okay now what makes it a bit sort of surprising and quite important from economic perspective is that not only did the increase in ai software production or innovation happening in the government domain it also happened almost simultaneously in the commercial sector the government contract it wasn't it wasn't exclusively contracting about okay so immediately after the first quarter of the of the public security contract uh uh took place uh happening to the to the air from the firms are also producing more for the commercial sector such as again supermarket checkout booths uh crowd control customer recognition for for for retail space and so on this is suggesting a fairly high degree of economy of scope of whatever the input of the government data the firm has received access to from working with the government the firm almost simultaneously are using that to either train a different algorithm or train an improved algorithm that can then take also to the commercial sector that benefit as a wider market on top of the top of the government okay the final thing i want to show you on the slice is that if we zoom into the to the to the to the subcategory of the ai software that is from developing that involves some degree or some element of civilians that such as monitoring and so on we see an increase of the number of software that involves civilians that these firms are are producing in the ai space this is sort of one one of the the evidence that we have suggesting that while the firms are becoming broadly more innovative and producing more and productive more productive after working with the local government there is a sense of repivoting and potentially distortion from a normative perspective that there are there are more civilians related to the product coming out presumably because the initial data that they receive the initial contract they receive is is meant for civilians and and you know if not for the you know part of the contractor but if not for the for the government contracting the firm for the commercial sector at least might have produced less civilians sort of focused uh ai goods than what they currently uh produce okay the final final results i want to show you before i open up to some of the questions it is that now that was for a a average contract that the firm received they seem to do better uh after they receive the contracts the question that i initially asked is that now what happened to those politically motivated contracts the contracts coming from a local government that just rece just experienced a lot of local unrest or coming from local government have higher share of legal population for example uh which which potentially the investment though of those ai contracts are actually sort of motivated by the exclusive demand for for recurring political unrest what we see here is that uh we want to focus on the 16 quarter after interacting with the unrest coefficient which is suggesting that all the positive effects that you show in the previous slides amplifies for those particularly motivated contracts so the firm who work with a government who are politically motivated when they sign those ai contracts differentially benefit from those contracts by producing more for the government producing more for the commercial sector and marginally pivoting a bit more towards the civilians kind of product okay so so to tie all this to to to together what i show you is is is a sort of several pieces of empirical evidence that that that that sort of uh uh suggests uh this this this alignment uh between and i would argue a quite deep alignment between autocross demand for political stability and ai sectors aim for data intensive innovation autocross benefit from from from from certain kinds of ai product and ai sector in particular the commercial sector ai benefit from working with the autographs because of the government data they receive generate large economies of scope that allow them to produce more for the various kinds of markets on top of on top of the direct demand coming from the government this will suggest the stable but but i should be careful here to say that when you know this is a stable equilibrium but this equipment could be distorted you know this this alignment between autocross and and ai sector can overcome a lot of the autocad frictions on innovation that the literature has been pointing to for the for the last decades uh because of the because of the additional data that the the the firm's able to receive and to the extent that the underlying economy of scope from government data is sufficiently large the autograph's demand for ai may generate greater growth for the sector uh and to the extent that the ai sector is important becoming increasingly important part of the economy it might also generate greater greater growth to the entire economy but however the direction of innovation is going to be distorted uh and and the citizens welfare uh could be reduced uh because of because of such uh distortion okay so i'm out of time unfortunately so i want i want to go actually show you the sort of the results uh for the for the for the on trade and indicators of china has a complex advantage in uh a a and china become a dominant forces in terms of exporting ai technology to the to the rest of the world uh and i i can save that for either for the q a or for for another talk thank you so much for for your attention ray thank you very much david since you're on this point actually one question is along a similar line um which is um do you think different political system in the us compared to that in china put the us in the disadvantaged position in the ai competition it's not exactly um what you have been presenting but it might be relating to your last part do you want to just venture to um answer this question yeah so so i think there are two important points to to to to make regarding sort of comparing you know china or sort of any sort of similar country with with similar political sort of uh institution uh with the us there is some commonality uh between them in a sense of you know across the board and throughout history government is always playing a very important role in collecting data and that's not uniquely true for china so when government needs to collect tax taxes or government provide any sense of sort of proper security large amount of data would be collected one way or the other and to extend that ai firms are are starving for data for the innovation working with any government is going to potentially beneficial to the to the ai firm to the extent that they receive access to this data now there is a differential effect for autocracy in a sense on top of the generic public security or public services local government providing they might have additional political incentives to use ai services because of her political sort of civility uh incentives and reasons that's gonna push from the chinese government or chinese local government in in the case of my talk uh having a greater demand than say the us local government uh uh and that that means more data potentially generated in this course now to the extent that there is a differential demand of the local government from china in demanding air services and providing more data uh you know from a pure theoretical perspective that that that is going to suggest a company advantage of countries such as china relatives to the to the u.s now if actually let me let me actually just uh share my screen i will show you the the the the the two slides i i i skipped in in the end is that in our in our next project we're looking at sort of is that actually true in the sense of from economic perspective the most direct way to look at this is whether there is a current advantage to look at who is dominating global trade in ai china or u.s or any other country who

are we think are dominating the frontier attack so we sort of collected sort of the original sources on ai trade and we see that china is by and large dominating global trade that involve ai product 50 of the global transaction uh cross-country transaction ai has china as a country of exporting and this is very different from all the other frontier attack uh that we're familiar with such as biotech and so on so ai seems to be china it's very advantage uh in in in in this sector and compare this to the u.s the u.s is exporting quite a bit about about this ai but it's much less so uh than china and again this is very unique to to the air sector and not the case for for other furniture attack where where u.s on that is is is is exporting more so if you take this as a as as a sort of uh uh uh evidence for company advantage then yes indeed seems to be that china has a common advantage in yeah because it is a net exporter of this technology and we don't see this in any other frontier attacks well we look forward to your next presentation on those results definitely um but this is uh what you just reply uh is related to another question the question is your assumption is that ai firms want to have contract with the government because of access to the government data but in the case of china in fact private company like tencent and all of that they actually have a huge amount of data have you ever considered that there's also a um incentive for the government to work with these firms so that they can get access to those data that they do not have yeah so that's that's a great ques that's a good question so i'm very much so far uh abstract away from the existence of private data and focus on sort of you know to the extent that working with the government has this benefit from a local from an ai from perspective access to government data the private data was was a bit on the background this is also an important factor related to the first question that you know compared to say this to the us the relative share of the government data versus private data in the in the economy or in the ecosystem uh could be could be very good we don't have a good estimate of you know who has the line share of the of data and who has more but probably the certainty is an important factor here now the two things i want to i want to emphasize when we think about private data is that there even to the extent attention and a lot of the large private firms is is certainly big players in collecting data there are certain data that only government has and that government has monopoly over for example this is one of the reasons why i focus on faith recognition and facial recognition air sector almost at least in the foreseeable future local governments and to extend sort of the central government of china and this is going to be the case for most of the governments regardless of whether it's china or not it's going to have control over who owns the the street service surveillance camera and who owns the video stream coming from the civilian camera so even for all powerful sort of private firms such as tencent and facebook they will not in the foreseeable future have direct access to ownership over such data and if they if if working with street civilians data is important part of how they develop some parts of the software they have to work with the government uh to to to to obtain that data because there's no other way because you can't put up a civilian's camera by a private firm now on the other hand the government is certainly also picking big firms and big players when they ask them to supply their software so there are especially for those high profile local government who potentially have a lot of pressure for for local political stability they're working with some of the best firms uh in fish recognition sector because they probably want the best kind of uh uh software so they're definitely pick uh you know when i say procurement there's a bidding process in the end no bidding is not very transparent but there definitely is a picking in the selection of some of the largest and most capable firms uh that's not going to be directly affecting some of the results that show this we're controlling for firm fixed fragments this is within being the top firm and you're still additionally benefiting almost immediately after working with the government but the background selection that i sort of skipped over is that these firms who end up working with the government which is overall is about 20 15 to 20 of the firms in china are 15 to 20 of the largest fish recognition firms uh small firms don't get to work with local government uh on average at least um there's a question from um bill xiao you have done a very interesting um quantitative study of the political economy between ai firms and auto credit governments and there is strong spill for effects on the commercial side um so um the question is um that um he's reminded about the spiritual effects of u.s military funded research have you compared um the demand in supply of chinese military research and this bill for effect so is it something that you see a parallel that yeah so empirically we we unfortunately don't i mean even when i say our our procurement contract database is very close to the universe i don't think we'll have nearly close to the universe or if any uh of the lion's share of the of the military-related procurement those are just not in the community so extend that empirically can we can we say something about the military side unfortunately so far we're limited these are local police department local public security armored government that could include local prisoners and so on when they're purchasing the app but it's it's not it's not very different from military now the the the big picture question that that this racist is is a very important one which is to the extent that working with the local government that generates spillover from the government sector of the commercial sector a very close analog of that is indeed military space racing would be another example where you know in those cases typically think it's human have to carry over certain ideas of how to do certain things and that when you will help the government to build a rocket that teach you a little bit of how to build an aircraft and 10 years later you work for boeing and then and that's go over to the commercial sector and we see that sort of you know in several sectors uh in in the last century or so now ai sector in that sense because of data is an input and data has this economy scope feature which means that once you take the data from the government or access the data from the government that same data allows to do things above and beyond what government directly wants you to do is and that generally this this sense of spillover that's quite mimicking the spillover com coming from space race or working with the military so so in essence when we think about sort of the direction between private firms and government in the ai sector uh we're back into the in the 50s and 60s when we think about some of the headline sectors when they're going to benefit tremendously from working with the state okay so the three questions from bill alfred um first of all again he complimented you on a very impressive presentation and work his first question is are there indirect ways in which the autocrats may help ai companies for example as in precluding development of new laws that would protect privacy more strongly or in providing other benefits not directly related such as loser application for wreck of regulation access to finance the second question which is related is um does association between ai companies and the autocross hampered the ai companies to raise social cap to raise capital outside of china third question which is also a question another person have asked is do local governments publicize their ai acquisition um the other person's question is almost the publication of this acquisition is a way to deter uh protest already so yeah so these are all great questions let me try to touch upon a couple of things that that's that's that's that's raising them um no first one is uh sort of privacy and and norms and regulation regarding privacy is an important one especially if you think conceptually that data is a key input into the sector uh and what governs sort of the supply of this input in the economy is coming through parts at least part of it the people's norm towards privacy that governs how much data people are willing to give to the government open to the private sector and the regulation that that that government states so this is sort of again advertisements of work in the progress we are we're in a process i've tried to having sort of various sort of experimental uh tools to elicit people's privacy preferences around the world in particular for the for the ai producing countries and see is this actually indeed the case uh as as sort of a lot of commercial wisdom suggest countries such as china have a lower privacy uh uh preferences by the better citizens but we actually don't don't know access to whether that's that's the case at least suggestively as well but they want to see very see rigorously is the case that people are less concerned about privacy uh uh in countries such as china now of course that lower concern could be indulgence the government's responses the autocratic government might have incentive to lower people's privacy concerns because that helped him collect more data and survey better uh we're also looking at sort of very systematically the global regulation in concerning privacy and data collection over the last 30 years and that's actually painting a bit more sort of a complicated picture where if you look on paper china has more strict regulation on privacy than the us uh by and by and launch no enforcement is a separate issue but at least for you know this there's a interesting interaction between norms regulations and how that governs the supply of data which which very much is what we're currently being working on the the you know i think the the next the second question about sort of ai uh the hottest ad from the racist funds from from from globally i think you know so far we have one firm who are undergoing ipo uh the first regulation which is also the first firm ai firms in the world who's going to look on ipos we're waiting and yet to see how the global financial market is receiving uh et cetera firm there's certainly i think increasing amount of scrutiny uh by the global community regarding how the firms who receive the data and which kind of government services it provides so i think this is very much evolving but i think i i think it's it's slightly less relevant i think from the from these firms perspective in terms of the source of funding because they the the ecosystem of the venture capitalists and and and and the capital market uh for these startups are quite healthy and if anything sort of you know uh over supplying what the funding that they're there it's not in the situation where their funding is constrained and they really need the funding from the global market but on the other hand sort of how do we think about to the extent many of these firms are indeed global firms and they're going to become global firms through the global financial market or through exporting their product think about trade related responses and regulations it's going to be quite important uh and this is exactly where we're going with with our trade projects uh going forward i don't have an answer yet but this is very much on our mind uh when we think about sort of the globalness of this phenomenon the final thing about sort of you know to the extent of local ai firms are publicizing their workings the government i think a lot of it you know you know they put this on on the website many of them suggesting you know working with with the government as i mentioned in the earlier question is a signal for the quality local governments especially sort of the shanghai local government is working with the best ai firms that's a good indication of quality uh to the various market with that i'm working with that government so a lot of ai firms are not shy about advertising that local governments sometimes also advertising their ai equipment as as bill said potentially for insurance uh in fact there are reports out there that say that you know a lot of the ai cameras are not actually air equipped but looked ai equipped just again as as deterrents but but but that's on the on on the background there you know local governments are oftentimes quite quite open uh about their about their uh their ai investment i should say that this this last point is actually about and this is related to sort of the the citizens compliance and and how they receive their investment is that you know both part of this ai investment is not going for political suppression directly it's it is about providing public security to to a certain extent to the extent that you know if the local citizens do do perceive ai uh equipped camera is providing a safer screen that does not generate a lot of cognition support now of course part of that is also going to go for a sort of suppressing political unrest or deterring political unrest but that may not be on top of people's mind for average citizen uh in china and and that's potentially suggestive why local government is not particularly hiding the fact that they're using a lot of ai well thank you so much we have a few more minutes uh and uh let me just pick one question from bobby phong um his question is um what might be the conditions or circumstances that you will see the equilibrium to be uh breaking up and uh how much do you think that the equilibrium i'm paraphrasing is stable and what might be the forces that would actually break it up yeah so the so so that's a deep question uh i don't have a perfect answer but i think what i'm if anything i try to convince you from the talk is that the raw ingredients that we need to sustain equilibrium exist that means you have benefits coming from both directions equilibrium wouldn't exist if only one party benefited from the other and the other party issues loser or at least don't benefit that would be a fragile one now so far it seems like both parties both party meetings the government and the ai firms benefit in both directions now what will change the equilibrium is going to be that if the ai firms start to benefit less from local from the government because for example if the privacy norm or privacy regulation becomes strict enough that the ai firms couldn't use government data and the government had become less useful for commercial sector that means that there's going to be less incentive for the local ai firms to benefit from the government and work with the government that will tilt the equilibrium towards shifting you can also imagine that if the if the local governments are having even better technology that's different from ai that that makes them sort of sustain political stability even more cheaply uh you know we don't know whether the technology will exist or when will that happen but to examine that might happen um you know in the next decades or a couple of decades that would be a force coming from the government side they're pushing away this program but at least as the current understands this is not a statement about about sort of numbers or quantity but at least in the current standing in terms of the the benefits going from both directions i think it is it is quite uh stable and i don't see a particular sort of trend in the media future where it might break this equilibrium and that's what i mean by when i say there's alignment between autocracy and ai it's a quite deep one and it can actually overcome a lot of the sort of headline frictions that would typically associate with autocracy discouraging innovation uh because on net there seems to be this the this alignment and benefits right well thank you very much um it's time and there's still a number of questions in the chat box and i encourage you to write to david directly i think i would just speak for day with them sure they will mind um and thank you very much for this very innovative work and i'm impressed with the massive amount of data that you have put together from so many different sources that allow you to answer new questions and i also see that um from your talk there are a number of uh mixed questions that would fill out that would be a resulting from this current research um and we look forward to welcome you back again to um share your work and we also look forward to having you work more closely with the center and different aspects of it so thank you everyone for attending and until next time bye-bye thank you so much for having me you

2021-03-04 18:30

Other news