1 Introduction
2 Methodology
2.1 Theoretical Approach
2.1.1 Countervailing E ects
2.1.2 ImpedingE ects
2.1.3 The Basic Formal Model
2.1.4 Technology and Labor Demand
2.1.5 Discussion
2.2 Empirical Approach
2.2.1 Methodology and Data Sources
2.2.2 Results
2.2.3 Discussion
3 Policymaking
3.1 The U.S. approach
3.2 Policy Responses to AI and its E ect on Employment and Wages
3.3 Discussion
4 Further novel research
5 Conclusion
6 Appendix
6.1 Table and Figures
AI has been a buzzword since its beginning in the late 1950s. However, the recent robust and developing focus on four self-enforcing trends take the notion of disruptive technologies onto a higher level. These are statistical and probabilistic methods, the abundance of increasingly large amounts of data, the accessibility of cheap, enormous computational power, and the transformation and adaption of more places into IT-friendly environments (e.g., Smart Cities and IoT).
These fundamental elements of AI can be found in many applications. Boston Con- sulting Group (2017) states some of these applications as follows: Marketing and Sales with personalized services and goods, Research and Development with aggressive forms of data collection and automating previously outsourced service tasks in large companies by combining AI with robotic processing automation1.
AI’s unprecedented ubiquity among these areas and more has underlined the feasibility, importance, and scalability of AI. Consequently, critical voices raised about how this enormous disruptive technology should be regulated and adjusted into the economy. As a response to the ethical, social and economic impact of AI, in October 2016, the White House O ce of Science and Technology Policy (OSTP), the European Parliament’s Committee on Legal A airs, and in the UK, the House of Commons’ Science and Technology Committee released their initial reports on how to prepare for the future of AI.
Furthermore, in contrast to the presumptions about the adverse e ects of AI on wages and employment, which had led the discussion into a false dichotomy, this paper introduces a more distinguished analysis. In the final analysis, the authors suggest that AI reduces wages and demand for jobs which can be replaced by AI. However, and this is the missing element causing the false dichotomy, Artificial Intelligence and automation induce countervailing e ects which increase the demand for traditional non-automated tasks. Also, AI helps to create new labor-intensive tasks which in turn increase labor share and thereby counterbalance the displacement e ect of automation in the aggregate.
Nevertheless, automation decreased wages and amount of jobs for low-skill and medium- skill occupations and caused an increased inequality and employment polarization. (e.g., Autor et al. (2003), Goos and Manning (2007), Michaels et al. (2014)). The IFR esti- mates that current industrial robots operations amount to 1.5-1.75 million, while Boston Consulting Group (2015) states that this figure could go up to 4-6 million by 2025 in the U.S.
Although automation is abundant, yet, it seems unlikely that we know much about the impact of technologies, especially of robots, on the society. Nonetheless, there are estimations about the scale of AI’s impact. Based on the tasks employed, over the next two decades, Frey and Osborne (2017) estimate that 47% of US workers are at risk while McKinsey (2017) estimates the number at 45%.
Albeit these incredibly high numbers, there is no guarantee that firms will e ectively automate. According to the profit maximization calculation, they would only automate if it is beneficial to substitute machines for labor and consider how much wages change in response to this threat.
Moreover, not only the place of automation is decisive, but also how other sectors and occupations react to it. It could be that other industries soak up the freed workers, or even automated industries could expand employment due to higher productivity.
Going beyond speculative figures mentioned above, Acemoglu and Restrepo (2017) estimate the equilibrium impact of one type of automation technology: Industrial robot: “an automatically controlled, reprogrammable, and multipurpose machine.” Most closely to this work is the paper by Graetz and Michaels (2015). Their work variates the data across industries in di erent countries and estimates that industrial robots increase productivity and wages while the employment of low-skill workers decreases. The data used in Acemoglu and Restrepo (2017) and the above-mentioned paper use are the same, but the empirical method variates and allows Acemoglu and Restrepo to observe cross- country, cross-industry comparisons, exploit exogenous changes in the spread of robots, and estimate the equilibrium impact of robots on local labor markets. The Micro-data mentioned above also allows them to control for detailed demographic and compositional variables when focusing on commuting zones and study the impact of robots on industry and occupation-level outcomes.
This thesis provides a framework to think about AI and how it a ects and shapes the society. The primary references used are as follows:
- Acemoglu and Restrepo (2018)
- Acemoglu and Restrepo (2017)
- Cath et al. (2018)
Finally, this thesis is structured as follows. First I will start summarizing the The- oretical and Empirical Approaches by Acemoglu and Restrepo in stating the e ects of Automation on wages and employment. Afterward, I will state Cath et al.’s review of recent policymaking decisions regarding AI and how to “make a Good AI Society” by the U.S., EU and UK. Furthermore, after each section, I am going to discuss each paper respectively and supplement them with my findings. Finally, I will state further novel research which could foster implementing a society in which AI does more help than harm and conclude afterward.
In this section, I will respectively summarize two papers with a theoretical and empirical approach. The theoretical paper: “Artificial Intelligence, Automation and Work” and the empirical paper: “Robots and Jobs: Evidence from US labor markets” by Acemoglu and Restrepo (2018) and Acemoglu and Restrepo (2017) respectively. While the former provides insights into the implications of both Artificial Intelligence and Automation, the latter one focuses on Industrial Robots. The reason for choosing di erent technological focal points is that there was no other recent empirical data regarding the e ects of Artificial Intelligence on wages or employment. Although acknowledging the technological di erences between Artificial Intelligence and Industrial Robots, I nonetheless consider these both technologies to be similar concerning economic and societal implications as we shall see in this section.
In this section, I summarize the findings of Acemoglu and Restrepo (2018). The following subsection is structured as follows: After introducing and explaining the occurring e ects of AI to revisit the dichotomy verbally, I will afterward present the formal implications.
ProductivityE ect: Reduces the cost of production and thereby increases the demand for labor in non-automated tasks. As cheaper capital substitutes for labor in specific tasks, the prices of the goods and services will decrease and make the households e ectively wealthier which ultimately increases the demand for such goods and services. There are two complementary ways through which the Productivity E ect works.
The first point is labor demand which might expand in the same sectors undergoing automation by reinstating workers in tasks which can’t be automated by AI. This is epitomized by the rapid spread of ATMs (Automated Teller Machines). ATM as a new technology displaced bank tellers which were more expensive for the banks. However, at the same time banks were enabled to open more branches as costs sharply fell and consequently employ more bank tellers who did jobs which ATMs could not automate Bessen (2015).
Secondly, AI causes higher real incomes by reducing the cost of production and consequently the price of consumption. This, in turn, causes higher demand for all products which triggers a higher demand for labor in other industries due to the increased supply of goods. Exemplary is the mechanization of the agriculture. Due to lower food prices, mechanization enriched consumers who then demanded more non-agricultural goods as suggested by Herrendorf et al. (2013). These examples reveal a crucial implication of automation as stated in (e.g.,Brynjolfsson and McAfee (2014)). The danger, thus lower wages, and higher unemployment seem not to be resulting from highly productive technologies which a ect other industries positively, but rather from mediocre technologies which are barely productive enough to be adopted, hence cause displacement.
CapitalAccumulation: Arises from the fact that automation binds high capital (Allen (2009), Olmstead and Rhode (2001)). The high demand for capital triggers all the more capital accumulation, for instance by increasing the rental rate of capital. The higher capital, in turn, raises the demand for labor.
Deepening of Automation: Di erent from the former described Productivity E ect which a ects the jobs extensively2 and thereby displaces workers, here the intensively deepening of automation instead raises from within the productivity of already automated tasks without causing the loss of jobs.
An illustrative example of this can be found in the replacement of horse-powered reapers and harvesters by diesel tractors (Olmstead and Rhode (2001); Manuelli and Seshadri (2014)).
All in all, these countervailing e ects reveal, that the displacement e ect by automation is not necessarily accurate in all scenarios. Disruptive technologies a ect workers in two respects: While some abolished tasks can cause the workers to be unemployed in one sector, other sectors could be triggered to raise their demands for new labor arising from the increased productivity and higher real incomes.
However, one e ect of automation cannot be counteracted upon: The increasing substitution of capital for labor which lowers the share of labor in national income. The tasks, the labor will execute, could be shredded down to a smaller set. Nevertheless, the ever-existing automation in history has yet to exhibit a decline in the share of labor in national income. The following headline New Tasks suggest the reason why we did not see such a development.
New Tasks: Historically, periods of intensive automation were always accompanied by the creation of new job descriptions. For instance, Acemoglu and Restrepo (2016) find that between 1980 and 2010, half of the employment growth in the USA resulted from the creation of new job titles. Acemoglu and Restrepo refer to this as the Reinstatement E ect, which creates new tasks to oppose the Displacement E ect, and thus can lead to a balanced growth path.
AI could help to create new tasks in two ways, as emphasized by this paper. Rapid automation may endogenously generate the incentive to establish new labor-intensive tasks. Such that if automation goes ahead of new tasks, the wages will decrease and further automation is less profitable. So, the firm will profit more if it introduces new labor-intensive tasks.
Moreover, Automation Technology Platforms such as AI could create new tasks. Future job descriptions could be called “AI Trainers, AI Explainers, and AI Sustainers” (Accenture PLC).
Finally, AI could be applied to enable the individual education for disabled students so they can take part in community life on equal terms. A policy Germany wants to establish in the form of an inclusive education system Werning (2014).
The above-described e ects so far explained how the negative aspects of AI could be counterbalanced, at least in theory. In practice, however, ensuring the smooth transition of reinstating displaced workers can be a tricky endeavor.
Firstly, Acemoglu and Restrepo are pointing to the changes like existing jobs. The process enhances both that workers find new jobs and learn the needed skills. New tasks tend to require new skills. If the education system cannot provide the necessary skills, the adjustment of new technologies can be hampered. Deloitte suggests that if specific technologies need a parallel execution by human beings, the lack of the necessary skills will impede the potential productivity outcomes of such technologies. Therefore, introductions of new technologies need time to exhibit positive e ects on the economy. Before wages and labor demand can increase, wages could stagnate, poverty expand, and living conditions worsened.
Intimately connected to the first point, the second argument concerns the excessive automation, meaning faster automation than socially desirable. As a result, excessive automation creates ine ciencies by wasting resources and displacing labor and thereby impeding the transition process Acemoglu and Restrepo (2018).
The basic model introduced in this paper is the simple model of a task-based framework used in Acemoglu and Restrepo (2016). In contrast to a standard aggregated production function, T1 considers tasks as the central unit of production, instead of a good or service such that
Abbildung in dieser Leseprobe nicht enthalten
where Y denotes aggregate output, and y(x) is the output of task x. The tasks run only from N-1 to N so that we can examine new tasks without changing the whole aggregate.
Each task can either be executed by human labor l(x) or machines m(x), depending on whetherthetaskhas been technologicallyautomatedor not. The output of task x
Abbildung in dieser Leseprobe nicht enthalten
considers all x œ [0 , I ] to be automated, thus to be executed by either machines or labor depending on the productivity “ M and “ L respectively. I denotes the threshold of tasks which can be automated with current technological knowledge and the range x œ (I, N ] those ones which cannot. Hence, the upper function of y L (x) is the output of both labor and machines while the lower function states the output by labor only. Furthermore, a realistic representation of human-machine relationship is considered in the assumption that “ L (x) / “ M (x) increases in x. The given intuition here is that labor has a learning curve with an upward slope, whereas on the other side machines are considered to have a constant learning curve without marginal gains. Hence, implying that labor has a comparative advantage over machines in higher indexed tasks. Moreover, it suggests consistency when looking at the assumption that only human labor can execute tasks in the higher ranges, namely x œ (I, N ].
The demand for labor can be expressed as a function of Wage (W = the marginal product of labor) equaling labor (N ≠ 1) times the exponent of labor in the aggregate productionfunctionas follows
Abbildung in dieser Leseprobe nicht enthalten
Transforming the equation we receive the share of labor in national income
Abbildung in dieser Leseprobe nicht enthalten
Both equations decrease in I, indicating that by increasing automation wages and consequently the share of labor in national income declines.
Following types of technological change induce di erent kinds of e ects on employment.
- (Extensive) Automation expands the set of automated tasks represented by I
- Labor-augmenting technological advances increase “ L (x)
- (Intensive) Deepening of automation increases “ M (x) for x Æ I 3
- Creation of new tasks increases the tasks (N) in which labor has a comparative advantage over machines
While the last three technological changes do not necessarily displace labor, the first one exhibits a direct displacement e ect, per results of this paper. To show how the displacement, arising from new technologies, is counteracted I will now state formally the countervailing e ects.
Examining formally, one can see two opposing implications when automating a task as follows.
Abbildung in dieser Leseprobe nicht enthalten
The Productivity E ect increases both the demand for labor where automation is taking place and in other industries where automation is not that prevalent. Both e ects sum up to the productivity e ect seen in T5. However, if the productivity e ect is limited, automation will always reduce labor demand.
This equation implies that if new technologies are only productive enough to replace one specific task but not highly productive enough to impact other tasks within an industry positively, these “so-so” technologies will exhibit displacement e ects.
Formally, one can see that when “ M /R = “ L/ W, the “so-so” techs will have a displacement e ect. If on the other hand, “ M /R > “ L/ W, the productivity is su cient to raise the demand for labor and wages. This is consistent with the hypothesis that only technologies which reveal comparative advantages over traditional labor will be applied.
However, thederivativeof T4
Abbildung in dieser Leseprobe nicht enthalten
shows that regardless of the productivity, the share of labor in national income will always decline, because automation always increases the productivity more than the wages4.
Apart from the Productivity E ect, there are two other counteracting forces, as already mentioned in Section 2.1.1, which rather mitigate the displacement e ect by a ecting the Productivity of the new technology.
Capital Accumulation: Acemoglu and Restrepo (2018) initially assumed that the supply of capital is fixed, such that investments in new machines increase the demand for capital and thus increase the rental rate R. The other more realistic possibility is that the newly bought machines do have a positive impact on the costs of production. And thus, as machines and labor are complements, an increase in the capital stock with a constant share of labor will rather increase the wages and reduce the rental rate. The productivity e ect is as a result of this even more significant, as can be seen in T5.
Interestingly, however, T4 still applies and continues to decrease the share of labor in national income. The intuition is again that the productivity gain arising from automation is always higher than the increase in wages.
Deepening of Automation: A further powerful e ect is the improvement of already automated tasks which doesn’t make the worker obsolete but slightly increases its produc-tivity. Translated into the model, this would mean an increase of “ M (x) for tasks x Æ I. If we further assume that “ M (x)= “ M (for all x)5, we can conclude that the Deepening of Automation tends to increase the labor demand and wages as stated by the total derivative of wage
Abbildung in dieser Leseprobe nicht enthalten
New tasks and the share of labor: More potent than the countervailing forces is the establishment of new tasks in which the labor has a comparative advantage over machines. In the formal perspective, this would translate into an increase of the set of tasks (N) as follows
Abbildung in dieser Leseprobe nicht enthalten
In contrast to the e ects as mentioned above, T7 reveals that the increase in new tasks does not induce a displacement e ect. Instead, we can see two e ects going in the same direction, namely the positive e ects of Productivity and Reinstatement. Moreover, we can also see thattheshareof labor in national income does indeed increase as such
Abbildung in dieser Leseprobe nicht enthalten
This is intuitive because the share of labor increases by new tasks: I has to grow the same amount as N, then and only then equilibrium wages grow proportionately with productivity and the labor share s L remains constant.
Mismatch of new technologies and skil ls and Inequality: Even if there are newly established and applied tasks, there is still a potential mismatch between the required skills of the workforce and the requirement of the new technologies or tasks. Let’s consider the following parameters in the next model – For brevity, I will only state equations for comparative reasons. Full values can be looked upon the Online Appendix:
Abbildung in dieser Leseprobe nicht enthalten
Higher S implies that there are plenty of tasks the low skilled employees can perform, while a lower S implies that there are only a few tasks left that low-skilled workers can perform. Thus, N ≠ S can be seen as the mismatch.
Hence, the impact of automation on inequality is the wage premium between high and low skilled workers.
Abbildung in dieser Leseprobe nicht enthalten
This inequality shows that the wage premium increases in I. In other words, the wage gap increases when the non-automated set of tasks decreases. This is intuitive because if S is close to I, the low skilled workers will be squeezed into only a small range of tasks which they can perform. This will lead to decreasing wages as the supply of workers will concentrate on the small number of left jobs. If S is big, however, the low skilled workers will have a more extensive set of tasks to perform and consequently won’t concentrate on one particular small set of tasks. This raises an interesting point mentioned before in the text: If automation does not go hand in hand with new job descriptions, the share of labor or low-skilled workers in national income will shrink.
Abbildung in dieser Leseprobe nicht enthalten
Abbildung in dieser Leseprobe nicht enthalten
In order to prevent inequality due to automation, decision-makers need to foster an environment in which the supply of skills goes hand in hand with new technologies. See for this policy-making in Section 4.
Impeding e ects slow down the transition process and reveal a conflict of objective between encouraging the workforce to acquire specific skills in AI and excessively investing in this specific technology. On the one hand, to solve the first point, we want to accelerate the learning process by concentrating the resources on Artificial Intelligence. On the other hand, we see that excessively using these resources for one specific technology causes ine ciencies.
Acemoglu and Restrepo (2018) try to find an equilibrium model concerning the grade of Artificial Intelligence and the e ects on employment and wages, as we saw in the formal section. The above-described conflict of objective could be solved when the optimal intensity of resources for the development of AI could be found. The ine ciently used resources, on the other hand, could then be used for di erent technologies in order to mitigate the short-term adverse e ects of displacement by creating tasks unrelated to AI, making the transition process smoother.
The problem with the current discussion in new technologies and labor is that the discourse fails to anticipate the balance between automation and the creation of new tasks.
[...]
1 Observing the worker’s performance and learning by repeating the steps successively to automate the tasks.
2 The word ‘extensive’ indicates that the new technology is unrelated to the old one, thus is considered exogenous, while ‘intensive’ means that old technologies are being improved.
3 The assumption is that “ M (x)= “ M for all x. Thus, the Deepening of automation tends to increase labor demand and wages instead of displacing labor.
4 dln(Y /L) / d I > dln W/ d I , as inferred from T5.
5 Assuming that for di erent figures of task indexes the marginal productivity does not change.
