♦ ♦ ♦
First of all, in terms of the hotly-debated impact of AI on employment, Massachusetts Institute of Technology (MIT) Professor Daron Acemoglu et al. present the basic framework (theoretical model) for observing the effects of automation through robots and AI on wages and employment.
This model emphasizes that while the displacement effect, which is generated as automation replaces labor in tasks that it used to perform, certainly reduces employment and wages; it is counteracted by a positive effect—namely, the productivity effect in which the cost savings generated by automation expand the economy, and the demand for labor in non-automated tasks is increased.
In addition, automation entails additional capital accumulation, which, in turn, increases the demand for capital and the demand for labor. Moreover, it could have the effect of deepening automation, i.e., improving the productivity of tasks being done by existing machinery. However, with such countervailing effects alone, increased output per worker exceeds wages per capita, and, accordingly, a reduction of the share of labor in aggregate cannot be avoided.
On the other hand, the most emphasized countervailing effects of automation, in Acemoglu et al., are the reinstatement effects whereby new labor-intensive tasks are created, and labor is reinstated in new activities. These effects tend to increase the labor share ratio.
Historically, for example, in 19th century England, when new industries emerged, new jobs were created including engineers, machinists, repairmen, conductors, back office workers, and managers. In terms of AI, Acemoglu et al. also point out that entirely new categories of jobs are emerging in firms using AI. Specifically, these jobs include "trainers" to train the AI systems, "explainers" to communicate and explain the output of AI systems to customers, and "sustainers" to monitor the performance of AI systems.
Acemoglu et al. claim that the application of AI systems to education and medicine/nursing care may also result in employment opportunities for new workers. For example, the application of AI will enable individualized education programs customized to the needs of each student, which currently are prohibitively costly. In the process, they predict the creation of new jobs such as the development and implementation of such individualized education programs.
However, the economy and the labor market will not necessarily adjust to automation as quickly as the above example. The reallocation and transfer of workers in the labor market could be slow and painful. Acemoglu et al. point out that such delays in adjustment will weaken the effect of productivity gains.
Specifically, first of all, if a mismatch occurs between the requirements of new technologies and tasks and the required skills of the workforce, it would slow down the adjustment of labor demand, contribute to inequality among the workers, and reduce the productivity gains. Consequently, the acquisition of new technology and complementary skills will become essential, and, in this sense, the role of the educational system cannot be over-emphasized.
The second is the case of excessive automation. Excessive automation not only creates direct inefficiencies but may also hold productivity growth down by wastefully using resources and displacing labor. Acemoglu et al. claim that the dismal growth in U.S. productivity, despite the development of information and communications technology (ICT), may be attributed to the fact that new jobs are not necessarily being generated due to this type of excessive application of technology.
The above discussion does not differentiate between robots and AI as machinery that replaces labor. However, in order to fully measure its implications, a discussion focusing on AI is essential. What clearly distinguishes AI from other forms of automation is its ability for machine learning (including deep learning) utilizing massive volumes of data including sensors, images, videos, textual information, etc.
♦ ♦ ♦
University of Toronto Professor Ajay Agrawal et al. claim that machine learning is more suitable to tasks that involve prediction. With machine learning having become cheaper, it is suitable for tasks that complement and automate prediction. Machine learning can cover a wide range of tasks, jobs and industries, such as automobile driving (predicting the correct direction to turn the steering wheel), diagnosing illnesses (predicting causes), and recommending merchandise (predicting customer preferences).
In addition, machine learning has been designed to encourage self-improvement over time. For example, an algorithm for machine learning (information processing procedures) can, assuming a large sample, autonomously find the coefficient that links a given set of input with a given set of output. Voice recognition that converts the recorded voice into text is one such example (See Table).
Table. Examples of Machine Learning Systems
|Historical market data
||Future market data
|Drug chemical properties
|Store transaction details
||Is the transaction fraudulent?
||Future purchase behavior
|Car locations and speed
|Source: "The Business of Artificial Intelligence," Erik Brynjolfsson and Andrew McAfee, Harvard Business Review, July 2017
Prediction using machine learning has an interesting application in public policy. Stanford Graduate School of Business Professor Susan Athey presents a case in which machine learning is predicting the probability of fire law violations being detected upon inspection when allocating fire equipment inspectors in New York City. Athey also presents an actual case in which a 30% to 50% improvement was seen in the number of violations found per inspection by health inspectors of restaurants in Boston through the use of such methods in the allocation of health inspectors.
If the essence of machine learning is understood to be prediction, then its risks will also become apparent. MIT Professor Erik Brynjolfesson et al. point out that it is difficult to explain the results generated by machine learning because AI will not tell you how it arrived at such a conclusion.
This could also lead to the following risks. First, the machines may have hidden biases. If a machine is fed data reflecting a set of decisions by a person, such human bias could also enter the machine. Second, with AI, it is impossible to prove with complete certainty that the system will work in all cases. Therefore, it cannot be used in critical decisions, such as life-or-death decisions. Third, a machine learning system will inevitably make errors, and avoiding such errors or diagnosing and correcting exactly what went wrong could be difficult.
♦ ♦ ♦
Finally, we must also not forget the gap that exists between prediction and decision-making. For example, as Agrawal et al., in the world of medicine, AI is beginning to take the place of radiologists. Using IBM's AI, "Watson," it is now possible to detect not only lung nodules and fractures but also pulmonary embolisms. Predictions are being made by statistically showing the probability of certain causes.
While greater precision in these predictions would reduce the number of biopsies performed, which are highly invasive, the decision of whether to perform such a procedure or not is still left to the judgment of the radiologist. We must not forget that the role of weighing the theories and the causes and effects, and making the final decision still remains with human beings.
>> Original text in Japanese
* Translated by RIETI.