AI and Employment: Surprise! So Far, So Good
October 23, 2017, 12 min to read
The evidence with a handy graph, depicting the evolution of the US employment rate over the last 70 years (source: ILO):
Overview: in the United States, since the end of the Second World War, jobs have been created faster than the population has increased. Let us take the opportunity to mention the three major movements that have shaken the labor market: lengthening studies and advancing retirement which have taken individuals out of the labor market, and the massive entry of women into this labor market.
Economic and financial crises (2007-08, for instance) are clearly related to the phases of employment rate decline.
What about unemployment in all this?
You will tell me that the real problem is that of unemployment. And I will reply that we are once again suffering from myopia: the situation in France is not rosy, so we feel that it used to be better in the past.
But over the last few decades, the unemployment rate has not seen an inevitable upward trend, which would suggest the disappearance of jobs for those who are looking for work:
No tipping point is to be found in the sudden emergence of devastating algorithms in recent years, since, beyond the usual cyclical fluctuations, the situation has not dramatically changed since the mid-1980s. It is well-known that the real tipping point was the sudden increase in the unemployment rate at the end of the Glorious Thirty (Les Trente Glorieuses in French).
2 welcome clarifications:
- I am submitting a statistical analysis, therefore aggregated; this cannot be read as an alleviation of the difficulty of unemployment at the individual level.
- This is not an expression of blissful optimism: I am not saying that the employment situation is much better than a few years ago, nor that it will improve. My aim is only to emphasize the errors committed by the oracles of the apocalypse, without relying on established facts.
Above all, the unemployment rate is not necessarily - or even rather not - the best indicator of employment. Even if unemployment speaks to us the most because it is tangible and frightening, it is the employment rate that must be scrutinized rigorously (see our article from last week).
The dividends of technical progress: reduced working time and higher income
The venerable Bank of England publicly released a series of statistics covering an incredibly long period of time - some figures dating back 1,000 years. Of course, such statistics should be taken with a grain of salt - the most distant data are estimates computed by economists, because state statistics are a recent practice. And we found such nuggets only for England. But the study of these series seems to us extremely interesting.
I was interested in the history of GDP per capita, broken down as follows:
GDP per capita = [number of hours worked per week] x [GDP per hour worked] x [labor force participation rate] x [1 - unemployment rate in the labor force]
(NB: the GDP per hour worked is the variable we calculate from the others)
Here is how the English economy has evolved since 1760, just before the Industrial Revolution began:
Verdict: In quantitative terms, technical progress had little impact on labor force participation and unemployment. In the UK, it can be seen that over the past 250 years productivity gains (x23 for GDP per hour worked) have been used to (i) increase real incomes (x14) and (ii) reduce working time (/ 1.6).
Surprisingly, technical progress has left rates of unemployment and labor participation relatively unchanged. In 2016 as in 1760, we still rely on half the population to keep the economy going. Even if the profile of the average worker, in terms of age and sex, has, of course, very different characteristics.
If the AI boom radically changed unemployment rate and / or labor participation rate, it would therefore be a radical revolution in the history of humanity.
The mechanisms of destruction and creation of employment
The debate around employment suffers from an obvious asymmetry: it is much easier to consider which job categories could disappear than to imagine those that could be created, and economic studies focus mainly on the theme of destruction.
Let us go back a hundred years and put ourselves in the shoes of our ancestors, when the majority of the population was working in agriculture. If you were told that a few decades later, farmers would be scarce, and even that salesmen, designers and sophrologists would take their place, you would find it hard to believe. Well, we are the peasants of the 21st century.
How can we consider that, despite obvious productivity gains that allow us to perform tasks in a shorter time, the employment rate can rise, or at least plateau?
Based on an excellent paper written by Deloitte on the mechanisms by which new technologies impact employment, both directly and indirectly, we can distinguish 5 processes:
- Emergence of jobs in the new technology-producing industries (e.g. industrial robot designers, data engineers in machine learning)
- Substitution of workers in certain industries leveraging the new technology - productivity increases but production is left unchanged in the short term (e.g. industrial robots require fewer hours to assemble a car)
- Complementarity of the new technology with other occupations, which increases their demand (e.g. consulting)
- Democratization of the products initially affected by Substitution: productivity gains allow price cuts and ultimately lead to an increase in demand... leading in turn to job creation
- Expansion of extremely diverse sectors of the economy, for which demand is driven by the purchasing power gains made possible by Democratization (e.g. cultural goods)
In one picture:
- Substitution is not the only impact. In fact, it is the only negative impact on employment.
- Through the mechanisms of Complementarity and Expansion, we can see that a new technology generates jobs in sectors far away from its "epicenter". Thus, criticisms of the type "but we will never need 5 million data scientists" are missing most of the employment issue.
- Of course, it will be pointed out to me that the evolution of employment depends on the mix between these different forces. My opinion is optimistic, for the simple reason that if we really wanted to produce the same things with fewer production factors, we could have stopped working a long time ago, with a few farmers and that’s all. This is what the American economist David Autor calls the "never-get-enough principle": if you think that employment is going to vanish, that means you bet that humanity has exhausted its stock of new problems to solve.
Beyond the usual matrix of automation
If you are particularly interested in assessing Substitution risks, whether in your company, in your industry or in the economy in general, we have considered the criteria that can inform you about the possibilities of automating tasks or even entire occupations.
Indeed, this distinction between tasks and occupations is critical: studies often consider entire occupations that may disappear, but each occupation is a basket of tasks. Thus, it is more likely that jobs will be transformed by removing or adding new tasks, rather than simply disappearing (which would imply on the one hand that all their tasks are automated and on the other hand that the employer cannot think of new tasks to entrust to them instead).
An example often used is that of bank tellers: a few decades ago, one of their main tasks was to hand cash to customers. This is now done via ATMs, but tellers have all but disappeared; they have turned to advising and sales tasks.
In the work of economists studying automation, a matrix is almost universally used in classifying occupations or tasks:
- A "routine / non-routine" axis
- A "manual / cognitive" axis
The conclusion being that routine jobs are under threat, that non-routine manual jobs will continue to grow, and that, above all, non-routine cognitive jobs have a bright future.
Nevertheless, this matrix is open to criticism for at least two reasons. Firstly, it is not always easy to distinguish between routine and non-routine tasks. Take the example of bus driving: situations on the road vary from day to day, but the driver's gestures do not change. Or medical diagnosis: all cases are not unique, and at least the major patterns of analysis are relatively stable.
Secondly, the opposition of what is considered as routine / manual on the one hand and non-routine / cognitive on the other is totally reversed in the real world. This is what artificial intelligence researchers call the Moravec paradox (after its author’s name): what seems so simple to us - like grasping objects on a shelf - is difficult to achieve (that is why Amazon makes robotized shelves come to its employees) - and conversely what seems excessively demanding to us - like becoming world champion of chess or go - ends up fully automated.
To get an idea of the potentialities of automating tasks and transforming jobs, instead of trying to classify them objectively, we should better start from the perception of the actors directly involved - employees themselves.
Our evaluation grid of the risk of job automation
Our thinking revolves around 3 successive stages: Opportunity => Will => Capacity of automation. This avoids the hasty identification between risk of automation and automation decision.
And within the Opportunity stage, we are striving to distinguish properly between tasks that could be easily automated, and those to which the saved time could be reallocated. In this "equation", a job can only disappear if its tasks are automatable without new ones being conceivable. Finally, the evaluation of the Opportunity criterion is based on surveys collected from employees - in fact, from what is done in France by the Dares, the equivalent of the US Bureau of Labor Statistics.
1 / POSSIBILITY (gauging the potential of automatable tasks ... and the potential for redeploying labor toward new tasks)
- Constraints on work
Machinery / technical constraints
Strict application of instructions
Inability to organize one's own work
- Skills to be redeployed (only category of our grid that defies automation)
Frequent abandonment of one task for another
Thinking about too many things at once
Insufficient use of some skills
2 / WILL (gauging incentives leading to a decision to automate)
- Competitive incentives
Fear of disruption
Cost / benefit ratio
3 / CAPACITY (gauging the capabilities necessary to implement automation)
- Digital maturity
Use of digital tools
Average size of a firm
Automation index for France
We have made a quick application of this grid to the 10 main French industries:
- excluding the agricultural and financial sectors, which are dealt with specifically in public statistics,
- and without being able to use the criteria "fear of disruption" (it would be necessary to distinguish startup investments by industry) as well as "cost / benefit ratio" (difficult to compute without first looking at numerous companies at an individual level).
It is a simple and relative classification: for each criterion the classification of industries goes from 1 (lowest probability of automation) to 10 (highest probability of automation). Then averages are computed for each of the 3 pillars and for the overall score.
If you wish to consult the raw numbers and build your own index, don't hesitate to send me an email.
One industry stands out from our index: the “transport and warehousing” sector appears to be the most susceptible to job cuts related to automation. For the risks are significant for each of our 3 axes:
- Opportunity: the work there is relatively repetitive, constrained, without a "store" of tasks able to take over from those that will be automated.
- Will: margins are the lowest of our panel, and the industry is rather open to international trade.
- Capacity: digital tools are quite present there.
Let's emphasize the points that we've previously made: we are talking about the mechanism of Substitution here... and the mechanism of Democratization could (more than) compensate the losses. In the specific case of transportation, it seems that in the course of economic history the reduction in cost has made demand grow exponentially (e.g. personal car, mass tourism thanks to airplanes).
Conversely, and by adding more details to the table, it seems that the construction and hospitality industries are relatively more protected from job automation: it would seem the tasks they entail lend themselves fairly well to automation, but their implementation remains centered on the physical world - which is notoriously more difficult to automate - and above all the competitive situation is quite favorable (probably because value is created on a local basis), and digital tools are not widely leveraged within those industries.
And what about job creation?
I pointed out that in the 5 mechanisms by which new technologies impact employment, the 4 generators of new jobs - but not necessarily new occupations - are more difficult to assess. For now, we have yet to found the magic formula, but here are a few suggestions:
- Emergence => tracking job creations in AI startups, and those in the digital giants who invest in these new technologies.
- Complementarity => starting from more narrowly-defined job categories or industries, leveraging the skills indicators to be redeployed as suggested in our grid, so as to identify who would benefit the most from AI tools to increase their productivity.
- Democratization => measuring demand elasticity within industries affected by Substitution: for a 1% price cut due to the reduction in costs enabled by automation, what would be the corresponding increase in demand?
- Expansion => studying surveys of individual values and aspiration - the World Values Survey is a great place to start - in order to consider which industries could benefit from the time and purchasing power gains that AI could deliver.