The robot economy
The threat to employment from robots is not a figment of the imagination: there is much evidence of technological unemployment. But there are obviously other factors that can explain the disappearance of many jobs in developed countries over the past two decades: the liberalization of trade and offshoring, the way the minimum wage is set, shortcomings in terms of education and the decline in union membership. And even though technological progress has played a key role in this process, robots are only a negligible part of new digital technologies. Why such fear? Because they have always been an object of fascination and because the progress in artificial intelligence is exponential? Because they get excessive media coverage? Or quite simply because the industrial use of robots is booming? According to the International Federation of Robotics, robot sales increased by 12% in 2013. This is a steep enough rise to merit attention.
The definition of a robot is very specific and must meet quite stringent ISO quite restrictive, standards, especially in their industrial applications. The sectors using most of these robots are the car industry, production of industrial capital goods and the chemical industry. Their usefulness is greatest in these manufacturing environments: simple, subject to very few random factors and highly coordinate. Assembly lines or warehouses are very good examples of controlled and highly predictable environments. There is also another dimension that is particularly decisive in the introduction of robotics: the ease with which operations can be programmed. Robots are used more intensively where operations are easy to program and where the procedures are formalized and do not require any kind of inference.
Which jobs are threatened?
The workers threatened by robots are primarily those whose jobs are based on routine and repetitive procedures and operations. The development of means of communication has already completely changed the geographical distribution of the different production stages. Globalization has actually been accompanied by offshoring of “midlevel jobs”, a threat of job losses that could be exacerbated by a more widespread use of robotics. Machines and globalization have quite similar effects: reallocation of labor into other, often less well-paid sectors or net job destruction.
However, an objective look at robots and their potential for replacing humans should not raise too much concern: they are not autonomous and have absolutely no common sense; they do not improvise and – in spite of noteworthy progress in the area of machine learning – they have no intuition. We are therefore still very far away from futuristic images associated with “singularity”, the phase theorized by Ray Kurzweil where robots have their own autonomy and can do without humans (in the film Her, for example).
The vast number of jobs that require flexibility, adaptability, judgment and creativity are likely to be preserved. In this respect, the economist Autor has drawn a distinction between abstract tasks (intuition, creativity, analysis) and those that require continuous adaptation to the environment (services and, more generally speaking, interpersonal relationships or proximity). Based on this distinction, Goos and Manning have talked about the phenomenon of “job polarization” (2003). The table below summarizes these issues. It is based on three simple ideas that can be summarized in three questions:
1- Are technologies a complement or a substitute for the tasks in question? If they complement more than they replace, jobs will be less threatened.
2- Does the fall in goods or services prices they result in generate increased demand? If demand is robust, we can expect a smaller impact on employment – or expect a more pronounced reallocation of employment.
3- Are there entry barriers or not? They can take the form of required skill or education levels, professional degrees or certification requirements. These barriers may limit the impact on job creation in activities where technologies are more a complement than a substitute (tools to help with decisions, management and diagnosis).
The table clearly highlights the risk generated by robots, and automation in general, for midlevel jobs. The progress made in terms of robotics could nevertheless spread to other levels. This should depend on our ability to simplify or model the environment in which robots function, as well as their ability to develop autonomous learning g.
Robots, environment and intuition
As we have seen, the ability of robots to replace humans is highly dependent on their limited interaction with their environment and the simplicity of the framework and the space in which they function. One of the ways to look into robotics is to circumvent the problem: rather than looking at robots’ ability to adapt to their environment, the environment can often be modelled or replicated faster and more effectively. That is the case for example with the Google Car, which can drive thanks to the precision and the permanent updating of digital maps and not because of the car’s ability to adapt to unexpected circumstances. Its “brain” is far more similar to the fluorescent lines and computerized motorways in the film TRON than to our own reality. Even though the integrated sensors, radars and surveillance cameras improve security significantly, the Google Car will remain inert and unable to explore unknown roads without a detailed road map updated virtually in real time.
If robots cannot control their environment, will they be able to develop their intuition and learn to draw conclusions based on inferences and repetitions? Machine learning is based on algorithms that make it possible to solve a problem or to carry out a task by generalizing based on an abundance of examples. The process is fundamentally based on the concept of classification: it involves collecting a multitude of cases in order to arrive at a single result or “exit”. There are many examples of application of the classification principle, from fraud detection, optical character recognition and spam filtering to analysis of customers’ responsiveness to various promotions, manufacturing of medicinal products thanks to the classification of proteins.
Without going into further detail on different types of learning, we can say that they are all more or less based on three factors: a representation (neural network, decision tree), an assessment (likelihood, cost, information gain) and an optimization The hardest task is to go beyond the simple framework of a logical sequence of arguments to introduce associations, or even intuition, in machines. This is a limit to artificial intelligence that Hans Moravec pointed out already in the 1980s: it is easier to reproduce the structured thinking pattern of an adult or a chess player than to instill common sense, intuition and a child's perception into a machine. In other words, the problems that are most difficult to solve are easier to model than the simplest human tasks or reactions, e.g. visual recognition. The human brain’s innate abilities – or those learned from the earliest age – to solve many problems are based on processes that we are not always able to explain.
In addition, there is a fundamental difference between the ability to represent phenomena and the ability to understand them. Being able to represent does not necessarily mean being able to learn or understand what is happening (i.e. give a semblance of “truth”): it is not because beer sales increase when the packs are located close to baby nappies that there is necessarily a cause-and-effect relationship between Heineken and births!
The goal here is not to discuss future advances in robotics, but to point out to what extent their diffusion to all economic activities raises major challenges. The jobs and tasks at the extreme ends will be all the more affected once the two fundamental properties of robots – control of the environment and intuition – are improved. It seems that the most likely path in the short and medium term is the development of decision-support tools: data processing seems to be more advanced than the rare attempts to substitute a robot for a human being in activities with a high interpersonal content – those that require adaptability and flexibility. In such a scenario, the polarization could become more pronounced between local jobs that remain protected and tasks with a high judgment and creativity content, which are already supported by decision-support tools – the growing complementarity of innovation and skills. Conversely, industrial applications, those where the environment is most adaptable to machines, will probably continue to be the area where the use of robots expands the most.
Training and education should be reviewed
It is crucial at this stage to understand that the ability to work with machines will have to be strengthened. Discussing the need to improve training and education makes sense only if teaching does not focus excessively on abstract ideas and formal subjects but rather on apprenticeships and mastery of technological tools. By providing technical support for certain tasks, robots are increasingly playing the role of a complement rather than a substitute for a wide range of professions. Let us take the example of medical assistants, or the large number of craftsmen whose jobs require particularly precise, albeit non-abstract, skills (wiring, installation of electrical systems, car maintenance, carrying out medical tests, etc.).
Robots, economic growth and inequalities
There are few empirical studies that enable us to assess the real impact of robots on economies. According to figures published by the International Federation of Robotics, robot use has been increasing continuously for 20 years. When calculated according to their density, i.e. their number relative to the total number of hours worked, they increased by 150% between 1993 and 2008. As we have seen, the first phase of robotics has mainly benefited three sectors so far: transport equipment, chemicals and metallurgical industries. In these sectors they carry out various tasks such as assembling and welding, data distribution and processing, and process inspection and control. The expansion of robots has gathered momentum thanks to the substantial fall in their prices: -50% in nominal terms; -80% if the improvement in their quality is taken into account. An interesting point is that despite this rapid diffusion, their total share in installed capital goods is very small and far smaller than information and communications technologies (barely 2% versus 11% for the latter).
The works of Graetz and Michaels (1) provide valuable insight into the economic effects of robots. They show that their use enabled an annual increase in productivity of around 0.4% between 1993 and 2007. To give an order of magnitude, this is a contribution similar to that of the steam engine in the United Kingdom between 1850 and 1910 and that of the large US highway program in the 1950s and 1960s. In contrast, the contribution is smaller than that associated with the rapid expansion of information technologies. But the latter lasted only for a short time, i.e. barely more than a decade (1995-2005). The increase in the number of uses for robots should therefore – despite the pessimism of the proponents of the secular stagnation theory – contribute positively to growth in the coming years. It would be pointless and ineffective to fight the spread of robots.
For the authors, a wider use of robots is accompanied by higher productivity gains and wages. But, unsurprisingly, they see a negative impact on unskilled or moderately skilled people. So the overall problem posed by robots remains unresolved, despite these rather reassuring figures. For while the diffusion of robots makes it possible to increase productivity gains and therefore to maintain a positive growth path for the future, the broader issues related to the substitution of machines for humans and the concentration of jobs and wages among the most skilled – or adapted as we saw above – are still not solved.
Robots and redistribution
It is therefore important, at an early stage, to think of redistribution and training mechanisms that are not incompatible with maintaining a high level of innovation. As Benzell, Kotlikoff, La Garda and Sachs have written: “Absent appropriate fiscal policy that redistributes from winners to losers, smart machines can mean long-term misery for all”. Although there are several possible trajectories, their model predicts gradual impoverishment of unskilled and, which is new, skilled individuals. A gradual substitution of machines for the skilled would indeed reduce the share of wages in GDP. In addition to the increase in inequalities, the fall in income would lead to a decline in savings, which would be followed by less investment and therefore, ultimately, lower growth.
Before sinking into pessimism we have to recap the advantages and negative effects of robots as well as their potential for replacing humans. The duo comprising control of the environment and intuition capacity significantly limits the potential for a rapid spread of robots to activities of a non-routine nature that require abstraction or adaptation On the other hand, the increased use of robots in industrial activities is leading to invaluable productivity gains at a time when many economists are concerned that we may be entering an era of secular stagnation or zero growth. It is therefore crucial to guarantee the conditions for robot diffusion while limiting the risk of misery for the people they will replace.
Education has a key role to play, but to do so it must be rethought, with a focus on people’s ability to adapt and communicate, but also on the mastery of technological tools. However, this is unlikely to be sufficient, and redistribution mechanisms must be designed and implemented. The time it will take to introduce such government transfer payments will depend on the impact of robots on our purchasing power and on our ability to rapidly extend their capacities: while the fall in the cost of goods and services produced by robots more than offsets the fall in wages, and while activities with a high level of human density or intuitive content are difficult to replace by machines, we are still far away from a situation where robots have stolen all the work.
(1) Robots at Work, CEPR 10477, February 2015.