The fear that technological innovation will increase unemployment is not new, and various theories in response suggest technology does not necessarily pose a threat to jobs. This column goes one step further, arguing that because automation requires rising wages and that requires increasing the set of labour tasks, innovation should ultimately reduce unemployment.

Many innovations come in the shape of machines that replace workers. We hear of cars that drive themselves, of robots that perform more and more tasks, and of how artificial intelligence can replace smart jobs. These technological developments cause alarm among many, and this has intensified since the last recession that began in 2008. The recovery from the recession has been slow, and especially in creating new jobs. That is why many have called it ‘a jobless recovery.’

Interestingly, this fear of ‘technological unemployment’ is not new and has surfaced many times since the Industrial Revolution, as workers feared that new machines might drive them out of jobs. This led to the rebellion of the Luddites in 1811-1817. They were artisans who viewed in awe how the mills of the textile industry threatened their existence. They started a mutiny and the British government had to send a large army to oppress it.

The issue came up again often, especially in periods of rapid technical change. On 26 February 1928, Evan Clark wrote an article in the New York Times, titled “March of the Machine Makes Idle Hands”. He claimed that “the onward march of machines into every corner of our industrial life had driven men out of the factory and into the ranks of the unemployed”. At the time, the US rate of unemployment was only 4.2%. Around then, Keynes (1930) coined the term ‘technological unemployment’ in his famous essay “Economic Possibilities for our Grandchildren”.

The experience of the past two centuries has shown that this fear did not materialise, but the issue keeps returning to public discussion. This is partly because our memory is short and partly because some claim that ‘this time is different’ and that the risk of automation is greater. A recent publication by the World Bank discusses this claim seriously and offers ways to reduce the risk of unemployment (Chuah et al. 2018). Throughout the years, many economists were more optimistic. Keynes (1930) suggested that future technologies will increase leisure and thus will increase employment. Zeira (1998) has shown that machines that replace labour in some tasks make labour in other tasks more productive, so the demand for labour increases. In a sequence of papers, Acemoglu and Restrepo (2018a, 2018b among others) show that if technology creates new tasks in addition to automation, it increases the demand for labour and keeps the rate of unemployment from rising.

Our recent paper takes this line of research further and makes a bold contribution to the debate (Nakamura and Zeira 2018). We argue that the rate of technological unemployment will decrease in the future and will converge to zero in the end. As we show below, the mechanism we describe is different from those suggested both by Keynes (1930) and by Acemoglu and Restrepo (2018a, 2018b).

Our analysis of automation and unemployment uses a theoretical framework of production by tasks, in which technical change involves two processes. One is automation, namely shifting existing tasks from production by labour to production by newly invented machines. The second process is adding new tasks. Unlike Acemoglu and Restrepo, we do not impose any assumption on the rate of creation of new tasks, and this rate can be zero as well. As shown below, the dynamics of automation are what drives the creation of new tasks.

To explain the main result of our paper, first note that the dynamics of automation can fall into two main cases. The first is that automation does cover all tasks and stops at some finite level. In this case, the rate of automation must decline as it gets close to the upper bound, and it converges to zero over time. Since the rate of technological unemployment is proportional to the number of new automated tasks relative to the total number of labour tasks, it follows that the rate of unemployment converges to zero as well. Clearly, this is a simple case, in which our main result holds.1 The more interesting case is when automation grows without bounds, which we discuss next.

Automation grows only if producers adopt it, i.e. if they prefer capital over labour for the next tasks. We assume that newly invented machines or robots are more expensive, as we first invent the easy machines and then over time invent increasingly more complicated ones. This is a reasonable assumption that is supported by empirical studies. Producers prefer a machine for a certain task to labour only if it costs less. This means that in order to continue to adopt the ever-costlier machines, wages must rise as well. This is required to keep automation going.

In our model, the wage level depends positively on the number of labour tasks and negatively on the share of labour in output. The reason for the first positive dependence is straightforward – the larger the number of labour tasks, the smaller the number of workers who perform each task. Hence, their marginal productivity in the task is higher and this raises wages. The negative effect of the share of labour on wages is harder to explain intuitively. If wages are higher, the tasks performed by labour are more expensive and producers will buy less of them. In order to keep output at the same level, producers purchase more of the automated tasks. This increases the share of capital in output and reduces the share of labour in output. Hence, the share of labour has a negative correlation with the wage rate.

Have you read?

As a result, if automation is unbounded wages rise continuously, and that can happen in one of two ways: either the number of labour tasks increases rapidly (actually faster than automation) or the share of labour goes down continuously all the way to zero. If we rule out the case that the share of labour converges to zero, then the number of labour tasks must increase continuously. This means that the rate of unemployment is decreasing and is converging to zero, since it is proportional to the number of new automated tasks divided by the number of existing labour tasks. If the denominator – the number of existing labour tasks – increases, the rate of unemployment falls and converges to zero.

Hence, our study shows that the dynamics of the rate of unemployment depend crucially on the dynamics of the share of labour. Many empirical studies, the most famous of which is Kaldor (1961), have actually found that the share of labour in output is very stable, at around two-thirds, both across countries and over time. In recent decades, the share of labour has experienced a slight decline, but it is still around 60% of GDP and in any case, it is far from going down to zero. Hence, the assumption that the share of labour does not converge to zero is fully in line with the well-known stylised facts of modern economic growth. This supports our result that the rate of automation unemployment should decline over time and converge to zero.

Our paper therefore presents a strong result. Help is on the way and automation unemployment will not rise – on the contrary, it will go down to zero at the end of times. Nevertheless, we should treat this prediction with some caution. In general, economic models are better at explaining processes and mechanisms than at predicting the future. Our paper should be treated accordingly. Its main message is that although automation causes unemployment by turning labour tasks into machine tasks, it might also ignite a mechanism that reduces unemployment. Automation requires rising wages, and that requires increasing the set of labour tasks. This increase reduces the rate of unemployment. Hence, automation leads to a continuing reduction in unemployment by its own adoption mechanism. This is an important point to bear in mind when we consider the effect of automation on unemployment and on the labour market in general.