Tesi etd-12232021-184704
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
BORDOT, FLORENT
URN
etd-12232021-184704
Titolo
Robotics and artificial intelligence : what impacts on employment and inequalities?
Settore scientifico disciplinare
SECS-P/06
Corso di studi
Istituto di Economia - JOINT PHD IN ECONOMICS
Commissione
relatore Prof. MINA, ANDREA
Parole chiave
- Robotics
- artificial intelligence
- technological unemployment.
Data inizio appello
20/06/2022;
Disponibilità
parziale
Riassunto analitico
The purpose of this thesis is to analyze the effect of robots, AI and, more broadly, automation on employment and inequality. Chapter 1 is dedicated to an empirical analysis that aims to answer a very simple question: is there any statistical evidence that robotics and AI technologies could increase unemployment? The results obtained on a sample of 33 OECD countries between 2005 and 2017 allow us to answer in the affirmative: a 10% increase in the stock of industrial robots is associated with a 0.42 point increase in the unemployment rate; and the AI variable constructed from patents is also positively correlated with the unemployment rate, but the relationship is less robust than for robots.
We continue the analysis by regressing the robots and AI variables on the unemployment rates differentiated by education level, and the results indicate strong heterogeneity between the different groups. For example, the effect of robots is strongest on the unemployment rate of people with upper secondary or post-secondary non-tertiary education, followed closely by the unemployment rate of people with below upper secondary education, and then the weakest effect is on the unemployment rate of people with tertiary education. This first result supports the thesis of a polarization of the labor market induced by automation. Finally, we continue the disaggregation of the unemployment rate by analyzing the effects of these technologies on the unemployment rates by level of education and age groups. Here again, the results highlight strong heterogeneity within a group with the same level of education. For example, the positive correlation between robots and the unemployment rate of individuals with below upper secondary education is found only for young people (25-34 years old), and no statically significant effect is found for other age groups with a similar level of education. These results contribute to the literature by providing additional evidence of the potentially harmful effect of AI and robots on employment at the aggregate level, while highlighting effects that turn out to be positive depending on the education level and age of workers.
In the second chapter, we focus on the theoretical mechanisms explaining the polarization of the labor market. This phenomenon, observed in many developed countries, is characterized by an increase in the share of employment and wages of low- and high-skilled workers, and a decrease in the share of employment and wages of medium-skilled workers. The theoretical literature explaining the link between automation and labor market polarization is based on the routine-biased technical change hypothesis (RBTC), which postulates that automation mainly targets routine tasks performed mostly by medium-skilled workers. The hypothesis is intuitive at first sight, but it also tends to condition the result obtained: if we postulate that automation targets routine tasks, which are themselves performed by medium-skilled workers, then we indirectly postulate that automation targets medium-skilled workers first, and the dynamic can only lead to a polarization of the labor market. This raises questions about the relevance of this assumption, and about the possibility of generating labor market polarization from less restrictive assumptions.
To answer this question, we have created and programmed a multi-agent model allowing us to simulate the effect of automation on employment and wages. This model is populated by heterogeneous agents, each endowed with different skills that they improve during their professional activities, and that deteriorate during periods of inactivity. Evolving in an uncertain environment that prevents the resolution of an intertemporal maximization program, the agents in the model adopt simple adaptive behaviors that allow them to fulfill an objective: to increase their profit for firms, and to increase their wage for workers. To produce one unit of a good, firms combine different types of jobs following a Leontief-type production function. Agents can only apply for a job if their skills match those required for the job in question. The model includes six different skills rated between zero and seven, the latter being the maximum level of expertise. The wages and skills required for each of these occupations are initialized from U.S. data obtained by aggregating information from the O*net database and the Bureau of Labor Statistics. Finally, companies seek to improve the efficiency of their production process by hiring engineers who conduct research and development activities to improve the level of productivity embodied in capital and to automate certain tasks.
First, we test a scenario without automation. The model reproduces certain stylized facts, such as endogenous real GDP growth, heterogeneity in the level of productivity of firms, and a positive correlation between the unemployment rate and income inequality. This first scenario leads to wage polarization, linked in particular to a competitive labor market that offers higher wages to the rarest skills, but the distribution of the demand for skills remains unpolarized. Empirically, labor market polarization is characterized by a deformation of the distribution of wages and the skill level of jobs; so this first scenario is not able to fully reproduce this phenomenon.
In a second step, we test a scenario with automation. Unlike the RBTC explanation, we do not make any assumptions about the types of tasks targeted by automation; we simply assume that firms automate certain tasks to reduce their production costs. This scenario reproduces the stylized facts listed above, as well as the erratic nature of automation, which materializes as irregular automation, characterized by spikes. Keeping the parameter values the same as in the first scenario, the introduction of automation leads to a polarization of wages and of the distribution of skill demand. This result is explained by the deskilling aspect of automation: when a task is automated, the skills needed to perform it are no longer required, and consequently workers no longer develop this skill. The working time saved is reallocated to other non-automated tasks, allowing the development of skills related to these tasks to be accelerated, because the learning mechanism is based on experience effects. The result is that workers are more specialized but, on average, less qualified because their skills that are no longer mobilized have depreciated, leading to a loss of versatility. In this model, automation generates technological unemployment, with an increase of 13.6 points in the unemployment rate between the two scenarios. Finally, we test an intermediate scenario in which automation is slower, and show that the degree of labor market polarization and the magnitude of technological unemployment are proportional to the pace of automation.
After having studied the empirical link between robotics, artificial intelligence and unemployment (chapter 1), and after having highlighted the theoretical mechanisms that can explain the emergence of technological unemployment and labor market polarization (chapter 2), we focus in chapter 3 on policies that can counteract the negative effects of automation. Technological progress is, historically, a vector of economic development; it is therefore not a matter of trying to slow it down, but rather of deploying policy tools to accompany the changes it brings about in the labor market. To do this, we extend the model developed in chapter 2 to test four policies: regulation of layoffs and resignations to stabilize the labor market, a minimum wage, an unemployment insurance system and a training system.
The results indicate that these policies have heterogeneous effects that are sometimes complementary, such as the regulation of labor contracts and the introduction of a minimum wage, which reduce wage polarization, while the training system reduces skill polarization; and sometimes opposite, such as the regulation of labor contracts, which negatively impacts productivity, while the minimum wage and unemployment insurance increase it.
Some of the policies tested appear to be effective in reducing the polarization of either the wage distribution or the demand for skills, but none of them manages to act on both simultaneously. In a final scenario, we test all four policies together, and the results are very positive, with a reduction in wage and skill polarization, an increase in the median real wage, in the general level of skills, in labor productivity and in real GDP, which leads to a decrease in the unemployment rate. These results show that technological unemployment and the induced labor market polarization can be effectively tackled with appropriate policies; and that policymakers need to think about public policies as a whole and not separately, at the risk of missing potential positive synergies.
We have shown in Chapter 1 that there is a statistically significant link between robots, AI and unemployment. In Chapter 2, we explained the theoretical mechanisms that can explain the emergence of technological unemployment and labor market polarization. In Chapter 3, we tested different policies to combat these two phenomena and showed that they can be mitigated through public intervention. These three chapters contribute to the literature on the link between technological progress, employment and inequality by covering both the empirical, theoretical and policy aspects of the subject; but they also suffer from several limitations.
In Chapter 1, the effects of robots on unemployment are easily interpretable, while the effects of artificial intelligence are less robust and more difficult to interpret. While for the "robots" variable we use data on the stock deployed in firms, for the "AI" variable we construct the data from patents related to artificial intelligence. Consequently, our data on AI are more a measure of innovation in this field than a measure of the stock of software and algorithms used in production processes, a discrepancy that may explain the lack of robustness of the results related to this variable. The best solution would be to have data on the stock of AI-related capital, as for industrial robots; but such data are not, to our knowledge, available at the macroeconomic level.
In chapter 2, firms compete in a single market by producing a homogeneous good. Although the heterogeneity of wages and skills required for the same type of occupation already emerges from the properties of the model, it would be interesting to observe whether the dynamics of these two variables differ in a multi-sector model. Indeed, consulting the Bureau of Labor Statistics data indicates that there are large disparities in wages for the same job between different sectors; and it would be interesting to be able to study whether the heterogeneity of technological trajectories between sectors can explain these differences. Finally, a sectoral approach would make it possible, thanks to an input-output matrix, to include intermediate consumption in the production function to be able to study the inter-sectoral compensation mechanisms. Indeed, automation in sector A leads to productivity gains and thus to a decrease in unit production costs and hence in prices. This fall in price is also reflected in the other sectors that use the output of sector A as an input in their production function, allowing firms in these sectors to be more competitive, to gain market share, to increase their production and, finally, to create jobs. Finally, opening the model to international trade could mitigate the extent of technological unemployment and labor market polarization generated by the model. By lowering costs, automation allows domestic firms to reduce the price-competitiveness gap with countries where labor is cheap, thus preserving jobs.
The last limitation raised for chapter 2 can also be applied to chapter 3, since the combination of the four policies tested has the effect of increasing the median real wage by 35 percent, which is not a problem in a closed economy, but could be detrimental to the competitiveness of domestic firms in an open economy model. The net effect on employment, which is slightly positive in the case of a closed economy, would then be much more uncertain and could even become negative.
These limitations provide avenues for future research. The combination of more precise data and more developed models constitutes a fertile field of research for a more detailed analysis of the transformations of the labor market induced by technological progress.
We continue the analysis by regressing the robots and AI variables on the unemployment rates differentiated by education level, and the results indicate strong heterogeneity between the different groups. For example, the effect of robots is strongest on the unemployment rate of people with upper secondary or post-secondary non-tertiary education, followed closely by the unemployment rate of people with below upper secondary education, and then the weakest effect is on the unemployment rate of people with tertiary education. This first result supports the thesis of a polarization of the labor market induced by automation. Finally, we continue the disaggregation of the unemployment rate by analyzing the effects of these technologies on the unemployment rates by level of education and age groups. Here again, the results highlight strong heterogeneity within a group with the same level of education. For example, the positive correlation between robots and the unemployment rate of individuals with below upper secondary education is found only for young people (25-34 years old), and no statically significant effect is found for other age groups with a similar level of education. These results contribute to the literature by providing additional evidence of the potentially harmful effect of AI and robots on employment at the aggregate level, while highlighting effects that turn out to be positive depending on the education level and age of workers.
In the second chapter, we focus on the theoretical mechanisms explaining the polarization of the labor market. This phenomenon, observed in many developed countries, is characterized by an increase in the share of employment and wages of low- and high-skilled workers, and a decrease in the share of employment and wages of medium-skilled workers. The theoretical literature explaining the link between automation and labor market polarization is based on the routine-biased technical change hypothesis (RBTC), which postulates that automation mainly targets routine tasks performed mostly by medium-skilled workers. The hypothesis is intuitive at first sight, but it also tends to condition the result obtained: if we postulate that automation targets routine tasks, which are themselves performed by medium-skilled workers, then we indirectly postulate that automation targets medium-skilled workers first, and the dynamic can only lead to a polarization of the labor market. This raises questions about the relevance of this assumption, and about the possibility of generating labor market polarization from less restrictive assumptions.
To answer this question, we have created and programmed a multi-agent model allowing us to simulate the effect of automation on employment and wages. This model is populated by heterogeneous agents, each endowed with different skills that they improve during their professional activities, and that deteriorate during periods of inactivity. Evolving in an uncertain environment that prevents the resolution of an intertemporal maximization program, the agents in the model adopt simple adaptive behaviors that allow them to fulfill an objective: to increase their profit for firms, and to increase their wage for workers. To produce one unit of a good, firms combine different types of jobs following a Leontief-type production function. Agents can only apply for a job if their skills match those required for the job in question. The model includes six different skills rated between zero and seven, the latter being the maximum level of expertise. The wages and skills required for each of these occupations are initialized from U.S. data obtained by aggregating information from the O*net database and the Bureau of Labor Statistics. Finally, companies seek to improve the efficiency of their production process by hiring engineers who conduct research and development activities to improve the level of productivity embodied in capital and to automate certain tasks.
First, we test a scenario without automation. The model reproduces certain stylized facts, such as endogenous real GDP growth, heterogeneity in the level of productivity of firms, and a positive correlation between the unemployment rate and income inequality. This first scenario leads to wage polarization, linked in particular to a competitive labor market that offers higher wages to the rarest skills, but the distribution of the demand for skills remains unpolarized. Empirically, labor market polarization is characterized by a deformation of the distribution of wages and the skill level of jobs; so this first scenario is not able to fully reproduce this phenomenon.
In a second step, we test a scenario with automation. Unlike the RBTC explanation, we do not make any assumptions about the types of tasks targeted by automation; we simply assume that firms automate certain tasks to reduce their production costs. This scenario reproduces the stylized facts listed above, as well as the erratic nature of automation, which materializes as irregular automation, characterized by spikes. Keeping the parameter values the same as in the first scenario, the introduction of automation leads to a polarization of wages and of the distribution of skill demand. This result is explained by the deskilling aspect of automation: when a task is automated, the skills needed to perform it are no longer required, and consequently workers no longer develop this skill. The working time saved is reallocated to other non-automated tasks, allowing the development of skills related to these tasks to be accelerated, because the learning mechanism is based on experience effects. The result is that workers are more specialized but, on average, less qualified because their skills that are no longer mobilized have depreciated, leading to a loss of versatility. In this model, automation generates technological unemployment, with an increase of 13.6 points in the unemployment rate between the two scenarios. Finally, we test an intermediate scenario in which automation is slower, and show that the degree of labor market polarization and the magnitude of technological unemployment are proportional to the pace of automation.
After having studied the empirical link between robotics, artificial intelligence and unemployment (chapter 1), and after having highlighted the theoretical mechanisms that can explain the emergence of technological unemployment and labor market polarization (chapter 2), we focus in chapter 3 on policies that can counteract the negative effects of automation. Technological progress is, historically, a vector of economic development; it is therefore not a matter of trying to slow it down, but rather of deploying policy tools to accompany the changes it brings about in the labor market. To do this, we extend the model developed in chapter 2 to test four policies: regulation of layoffs and resignations to stabilize the labor market, a minimum wage, an unemployment insurance system and a training system.
The results indicate that these policies have heterogeneous effects that are sometimes complementary, such as the regulation of labor contracts and the introduction of a minimum wage, which reduce wage polarization, while the training system reduces skill polarization; and sometimes opposite, such as the regulation of labor contracts, which negatively impacts productivity, while the minimum wage and unemployment insurance increase it.
Some of the policies tested appear to be effective in reducing the polarization of either the wage distribution or the demand for skills, but none of them manages to act on both simultaneously. In a final scenario, we test all four policies together, and the results are very positive, with a reduction in wage and skill polarization, an increase in the median real wage, in the general level of skills, in labor productivity and in real GDP, which leads to a decrease in the unemployment rate. These results show that technological unemployment and the induced labor market polarization can be effectively tackled with appropriate policies; and that policymakers need to think about public policies as a whole and not separately, at the risk of missing potential positive synergies.
We have shown in Chapter 1 that there is a statistically significant link between robots, AI and unemployment. In Chapter 2, we explained the theoretical mechanisms that can explain the emergence of technological unemployment and labor market polarization. In Chapter 3, we tested different policies to combat these two phenomena and showed that they can be mitigated through public intervention. These three chapters contribute to the literature on the link between technological progress, employment and inequality by covering both the empirical, theoretical and policy aspects of the subject; but they also suffer from several limitations.
In Chapter 1, the effects of robots on unemployment are easily interpretable, while the effects of artificial intelligence are less robust and more difficult to interpret. While for the "robots" variable we use data on the stock deployed in firms, for the "AI" variable we construct the data from patents related to artificial intelligence. Consequently, our data on AI are more a measure of innovation in this field than a measure of the stock of software and algorithms used in production processes, a discrepancy that may explain the lack of robustness of the results related to this variable. The best solution would be to have data on the stock of AI-related capital, as for industrial robots; but such data are not, to our knowledge, available at the macroeconomic level.
In chapter 2, firms compete in a single market by producing a homogeneous good. Although the heterogeneity of wages and skills required for the same type of occupation already emerges from the properties of the model, it would be interesting to observe whether the dynamics of these two variables differ in a multi-sector model. Indeed, consulting the Bureau of Labor Statistics data indicates that there are large disparities in wages for the same job between different sectors; and it would be interesting to be able to study whether the heterogeneity of technological trajectories between sectors can explain these differences. Finally, a sectoral approach would make it possible, thanks to an input-output matrix, to include intermediate consumption in the production function to be able to study the inter-sectoral compensation mechanisms. Indeed, automation in sector A leads to productivity gains and thus to a decrease in unit production costs and hence in prices. This fall in price is also reflected in the other sectors that use the output of sector A as an input in their production function, allowing firms in these sectors to be more competitive, to gain market share, to increase their production and, finally, to create jobs. Finally, opening the model to international trade could mitigate the extent of technological unemployment and labor market polarization generated by the model. By lowering costs, automation allows domestic firms to reduce the price-competitiveness gap with countries where labor is cheap, thus preserving jobs.
The last limitation raised for chapter 2 can also be applied to chapter 3, since the combination of the four policies tested has the effect of increasing the median real wage by 35 percent, which is not a problem in a closed economy, but could be detrimental to the competitiveness of domestic firms in an open economy model. The net effect on employment, which is slightly positive in the case of a closed economy, would then be much more uncertain and could even become negative.
These limitations provide avenues for future research. The combination of more precise data and more developed models constitutes a fertile field of research for a more detailed analysis of the transformations of the labor market induced by technological progress.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|