DAILY DISRUPTION FEATURE
In our experience academic papers seem to fall into three broad classifications:
- Those that give “answers”. Researchers run an experiment, show the results, and describe what they mean.
- Those that summarize what we know today. Meta-analysis (looking at many different studies around a topic and cross-checking the results) is a popular form of this approach just now.
- Those that outline a useful framework for a large-scale problem.
Today we have an excellent example of the last one, titled “Toward understanding the impact of artificial intelligence on labor”. The paper is the product of 13 researchers from various parts of MIT as well as the Kennedy School at Harvard, UNC-Chapel Hill, Arizona State and Northwestern’s Kellogg School. Bright minds, in other words, and the work is currently in the top 20 “Most Read” publications on the Proceedings of the National Academy of Sciences website. So the “right people” think this is an important contribution to the discussion.
The authors begin with the point that there’s nothing new under the sun on the topic of technology and how it might disrupt labor markets. Plato fretted over mass adoption of writing and how that might affect human memory and knowledge. Robert Kennedy worried about automation in 1964. In between the two you had the Luddites of the Industrial Revolution, of course. Nonetheless, labor markets and humanity as a whole have found their way through all these challenges.
In terms of the current debate over artificial intelligence and its potentially disruptive impact on global labor markets, the researchers outline 3 commonly held views:
- The Doomsayer: Numerous studies put the percentage of the global workforce at risk from “technological unemployment” at anywhere from 9% - 60% depending on the geography and sort of work being displaced.
- The Optimist: The idea here is that labor markets evolve in response to technological change. History is on the side of the optimists, but AI doomsayers still worry that the technology is unlike anything that has come before.
- Those looking for a “Unifying Perspective”. The paper’s authors put themselves into this camp and argue that we lack the data necessary to know if the Doomsayers or the Optimists are correct.
To develop a framework on how AI will affect labor markets, the paper’s authors first posit that, “occupations are best understood as abstract bundles of skills and that technology directly impacts demand for specific skills instead of acting on whole occupations all at once”.
That makes a lot of sense, but it sets up several messy problems if you want to use this tidy paradigm to assess AI’s future impact on the global workforce. According to the paper, here are three notable issues:
#1. There’s not much data on exactly what goes into that “bundle of skills” for a specific occupation. We can back into whether these jobs are highly paid or not and even develop a coarse list of attributes (educational attainment, interpersonal skills, etc.). But none of this is granular enough to determine the level of disintermediation that might occur when artificial intelligence comes online.
The authors’ best idea on how to crack this code: LinkedIn (although they do not mention it by name) is a granular-enough dataset, even with its obvious geographic and occupational limitations.
#2: Every labor market is different when it comes to how it interacts with disruptive technologies. The authors cite the well-known example of ATMs, which were supposed to put bank tellers out to pasture. For a raft of reasons the opposite actually occurred and now there are many more bank tellers than before ATMs existed. As it turned out, tellers were still needed to address ever more stringent KYC/AML requirements, small business growth, and product cross selling.
This is a good example of how fuzzy the AI-labor market intersection really is, and the paper basically says the best way to navigate it may be (no surprise) with artificial intelligence itself. Given enough data about workers/their skills and job openings/requirements, an AI system might be the most effective way to match the two.
#3: Geography matters. Urban centers, for example, have been major labor market beneficiaries of disruptive tech-driven growth over the last 20 years. But as AI starts to take programmer jobs, for example, that may change. And as with points #1/#2, there is no granular skill-specific/city-level employment market data available to track the impact of AI as it grows.
Summing up the authors’ central message: traditional methods of assessing labor markets are woefully inadequate when it comes to either measuring the impact of AI or helping researchers/policymakers address any disruption as it develops. There is a historical analogy here: the US only started measuring “unemployment” during the 1880 census, long after the Industrial Revolution. The modern definition of the term – looking for work but not employed – dates to the latter part of Great Depression (1937).
Measuring labor markets always lags the disruption that makes its measurement important in the first place.
So whenever you hear someone spout off a statistic about how AI will change labor markets, the right retort is a simple “very interesting… but how do you know?” The MIT et al paper makes it crystal clear that no one except perhaps Microsoft’s LinkedIn asset has the data to even begin to measure AI’s impact. No, that’s not a comforting message. But it is an honest one.
Source paper: https://www.pnas.org/content/116/14/6531