Wednesday, January 17, 2018

Automation and labor: insights from the American Economic Association meetings last week

As briefly mentioned in my previous post, the American Economic Association held its annual meetings last week in Philadelphia. While the panel on gender bias was not webcast, several other lectures and discussions are available online. Another session that caught my eye was on automation and the future of labor, seeking to answer: what are the projected effects of automation, artificial intelligence, and robotics on labor share, wages, and the nature of work?

Daron Acemoglu presented a theoretical paper, co-authored with Pascual Restrepo, that provided a framework for understanding the relationship between artificial intelligence and labor share. He breaks down the impact of artificial intelligence into two countervailing effects: a displacement effect and a productivity effect. The displacement effect is the inevitable displacement of the labor force that takes place when firms substitute machines to complete specific tasks previously done by labor. The reason there is a displacement effect at all is that the capital is cost-saving for firms. One result of this cost-saving displacement is that there may be an increase in productivity associated with the firm's output. This productivity effect will lead to a demand for new skills and new job creation. But the question posed by Acemoglu and Restrepo is how large is that increase in productivity associated with employing machines instead of labor and will it lead to large enough job creation that it will balance out the displacement effect?

An example provided in the paper of these two effects comes from Bessen's (2016) analysis of the introduction of ATM machines. The paper found that the introduction and wide dispersal of ATM machines, a technology that took over many of the existing tasks of bank tellers (notably many existing tasks that were performed more expensively by bank tellers), allowed banks to cut costs. This cost saving allowed them to open more branches, which in turn led to an increased demand for bank tellers who could then focus on more specialized skills that the ATMs did not have. I don't review that paper here, but I note that it is contentious in its isolation of the causal effect of ATM machines on the banks' decisions to open new branches. The example, however, illustrates the mechanism by which the displacement and productivity effects work according to the paper (some bank tellers in existing branches displaced and bank tellers in new branches added).

The model in Acemoglu and Restrepo indicates that the effect of automation on labor share is unambiguous (labor share will decrease with the displacement effect holding productivity constant) but if the productivity effect leads to new job creation (demand for new skills leads to new job creation) then it has the ability to lessen the inevitable job displacement associated with AI. The productivity effect from employing ATM machines arguably led to an increase in the number of bank branches employed and the number of bank tellers employed who then needed to have new skills in the tasks that the ATM could not complete. I don't think that the "new skills" required in the bank teller positions are necessarily a good example of the demand for new skills modeled by Acemoglu and Restrepo since the new jobs created are effectively the same as the old jobs being displaced at existing branches but it's possible they may require improving on some existing skills in order to better advise the client on the different transaction opportunities available to them or bringing in new clients.

There are a couple of takeaways I think are important here:
  • Will jobs be created at all?: 
    • The paper highlights the case in which firms adopt technologies that are only marginally more efficient than labor at performing the same task ("so-so technologies"). The adoption of these technologies leads to few productivity gains and as a result lesser job growth through new skills. But the displacement effect will still be resounding and Acemoglu and Restrepo argue that it is these marginally more efficient technologies that will be the most harmful to the labor force since they don't lead to productivity gains. 
    • In the case of the bank tellers, what if instead of investing in new physical branches (requiring employment of bank tellers), banks invested in improving their mobile and online infrastructure to better serve clientele online? Firms' productivity may be growing but the productivity gains do not necessarily translate into job creation at the same rate (creates jobs for those tasked with updating and maintaining the online infrastructure but would this be comparable to creating jobs for a new set of tellers at new locations?).
  • Address inequality implications: This leads to the next point. It is clear that the jobs that are created through the demand for new skills will not employ the same skills as the jobs that are displaced (see the example of investing in new physical branches versus investing in a better online infrastructure and the skills needed to maintain each of those). Which raises the question of whether income and wealth inequality will be exacerbated by rising automation if the jobs that are displaced disproportionately impact those at the lower quintiles and the jobs that are created disproportionately require skills that those at the lower quintiles do not possess or cannot reasonably acquire. Perhaps anticipating the impacts on those at the lowest quintiles of the income ladder several prominent tech executives, including Elon Musk, have advocated for a universal basic income that they claim will be the only way to address the widespread job loss associated with automation. 
  • Identify type of jobs created: The response from Ben Jones in the discussion directly following Acemoglu's presentation raised an important point: the model appears to assume that all of the new tasks that are created based on the productivity effect are essential, i.e. there would be no output if the task were not completed. How likely is it that the new jobs created in the aftermath of technology adoption would be essential jobs (essential to production)?
    • If, as Jones hints at, the new jobs are non-essential, it is also likely that they may be of lower quality. Quality of employment is particularly important given the rise of the gig economy and trend towards temporary and part-time employment that offers fewer benefits and protections to workers. 
  • Prepare for "new skills": In order to preempt the potentially negative implications on labor share and inequality the key would be to identify the kinds of new skills that will be most valuable in a future with automation and how governments, policymakers, and educators can effectively plan for such a future by preparing students for those skills. Furthermore, they would want to be able to prepare those outside of formal education (those who are not in schools, universities, or training programs) for retraining and lifelong learning so that they can better adapt to changing conditions in the labor market. 
  • Identify market failures contributing to "excessive" automation: In their paper, Acemoglu and Restrepo outlined the phenomenon of "excessive" automation that is only marginally more cost effective than labor and that leads to few productivity gains and little job creation. They provided a few reasons for the "excessive" automation, one being that capital is potentially over-subsidized through the tax system which in turn encourages firms to automate.


  1. Very interesting and pertinent topic. I like the way you framed the takeaways.

    Re: Addressing the inequality implications - this reminded me of recent articles on racist robots used in the US criminal justice system. Machines basically learning systemic racism due to the data they are fed, their creators, etc.

    Also Did you see that Finland recently tested Unconditional Basic Income? I think the results were promising but I haven’t kept up with the story much.

  2. Right, I remember reading about the "racist robots" in this article a while back: One of the things that struck me about the issue you brought up is that the regulation and policing of these issues requires significant statistical/machine learning technical know-how (to even identify that there is an issue with how these AI programs work and that they are discriminatory). This is also raised in the Guardian article that you posted in that not only does it take a high degree of understanding to parse out how the AI works but also a battle to acquire the transparency to investigate it in the first place.