How to think about the role of Artificial Intelligence in Operations?
Many people talk up AI, IoT, automation, etc, as the Fourth Industrial Revolution (following steam power, electricity, and computerization). Here is an example. I am more persuaded by the counter-arguments. For example, see a post by Luke Muehlhauser arguing there was only one industrial revolution, because one of the revolutions is substantively larger than, and different from the others, as exhibited in the figure below (data from the site).
The Second Machine Age, by Erik Brynjolfsson and Andrew McAfee, which ironically, I read on paper, explores the effects of the rapid digitization and information technological advances (AI, Automation, etc) on the nature of work, wealth and society. Brynjolfsson and McAfee also seem to adhere to the latter notion that most of the pre-industrial historical productivity growth has been flat, and later computerization advances are smaller. Brynjolfsson (with my Wharton colleague Lorin Hitt), has written fundamental papers on the lack of IT induced productivity.
Hence, the focus of the book is sharper. Are we now living through a slight uptick after the only machine age, or, in a second machine age? The authors argue that we are in the second machine age, as the explosion of growth and challenges parallel those of the “first” machine age, i.e., age of the industrial revolution.
A Story in Three Parts.
The book handles the issue in three themes: (a) explosive growth in technologies (actual description of the technology, models to understand IT), (b) the bounty from the technology and how it has spread (inequality, GDP vs. wages, etc), and (c) policy recommendations.
The first part focuses on the data-driven demonstration of the explosion of technological innovation. While it is fascinating to read, precisely because of the argument that the authors make, this section of the book is going to be quickly outdated. (I am reading this less than two years from the publication, and much of technology has already changed – and the authors have a new book out a few months back).
The second part of the book strengthens the findings from Erik Brynjolfsson’s substantial research into IT infrastructure (the compelling parallels between the growth of electricity vs. information technology, Normal distribution vs. Power Law distribution in incomes), and uses the labor economic research to demonstrate how various bounties either do not show up in productivity gains or are being shared inequitably among the population. The thesis is that the current information growth is substantively creating a winner-takes-all market (e.g. Waze app, Facebook). It is certainly a provocative theory that is being tested out, as we are facing some critical drawbacks in large-scale social networks.
The third part of the book makes six policy recommendations. Some of the policies are along the expected lines often heard from us STEM researchers and Silicon Valley. Understandably, it is certainly the least data-oriented part of the book. The suggestions are: (1) Pay teachers/overhaul education, (2) Restart startups, (3) Match workers to supply (peer networks, 2-sided platforms), (4) Support scientists (through Prizes and innovation contests that I wrote about here on the blog), (5) Invest in infrastructure (Keynesian stimulus) and (6) Tax wisely (Pigovian taxes such as congestion fees, VAT taxes). The authors do suggest exploring long term regulatory strategies that need further exploration such as Universal Basic Income (UBI), and negative taxation (similar to EITC) and taxing superstars.
Fundamentally, note that all the above suggestions could also be argued for non-IT innovations (such as tractors disrupting farming jobs in agriculture). Hence, a challenging question still remains: How can one regulate network externalities specifically created by winner-takes-all type digitization?
—
I flag two issues on the opposite ends of the spectrum of Automation and AI in Operations: Jobs that are surely being replaced, and jobs that are harder to replace.
Technology Training using Cheap Labor
Much of the AI technology requires a fixed cost of investment in “training” the technology (often addressed under the umbrella of “supervised learning” in Machine Learning). Despite the AI can “learn-GO-by-itself” media coverage, a large amount of labor processed data is required for any AI to work well. For instance, AmTurk uses a significant amount of outsourced labor at cheaper marginal costs, to train AI data such as automatic photo tagging products on a webpage. TechRepublic in a recent study reported that the typical weekly wage for AmTurk workers is $79, and 75% of AmTurkers are American. Most tasks are worth a few cents, and as result workers complete thousands of repetitive tasks.
Such repetitive tasks are not dissimilar to old ideas of division of labor on an assembly line. Film aficionados may recall scenes of comic pathos from Charlie Chaplin’s The Modern Times. But the difference is the lack of training data in automation jobs. Compared to the cheap labor that trains it, AI will be superior due to the advantage in speed of execution, and AI will also become available at near-zero cost, once trained (although, it is not clear how true the latter claim on cost is).
Most researchers concur that automation will create some new jobs for labor. However, increasingly, we have begun to wonder about how to formulate the production process behind AI training, and the training labor replacement costs. An excellent recent paper by Acemoglu and Restrepo (2018)1 builds a model on this issue. I will discuss more frameworks in upcoming posts, but the paper recognizes the skill gap that exists between the jobs that train the AI/automation and jobs that are created by AI/automation.
Saved by Baumol Cost Disease
As noted by Brynjolfsson and McAfee, some jobs have remained relatively untouched by the digitization explosion. Many of those jobs such as musicians, gardeners, and critical care providers are Service Operations jobs that face consumers.
These jobs possess aesthetic and interactive components and comprise of valued labor content. The economist William Baumol, who passed away last year, recognized the labor-intensive nature of such jobs and showed that they have low productivity growth. Think of classical music. We have thousands of discs and music files of recorded concert orchestras, but live concerts are superior. Moreover, there cannot be, and indeed there has not been much increase in productivity in concert performances. Here is an analogy originally due to Baumol, from the New Yorker.
When Mozart composed his String Quintet in G Minor (K. 516), in 1787, you needed five people to perform it—two violinists, two violists, and a cellist. Today, you still need five people, and, unless they play really fast, they take about as long to perform it as musicians did two centuries ago.2
Many jobs that are relatively unaffected by AI, are indeed riddled with the cost-disease. Productivity may be lagging in these job sectors, but the cost disease is the saving grace that buys time under the AI onslaught.3
Notes:
- Acemoglu and Restrepo, 2018, Automation, Artificial Intelligence, and Work. MIT Working Paper.
- We can improve productivity by playing up-tempo (try that!) or building bigger orchestra halls (imagine being in the thousandth row). Both produce poorer quality outcomes.
- This quality vs. speed conundrum is addressed in one of the queueing papers.
If you like what you read, I welcome you to subscribe to OWL Mail and have the blog digest delivered to your inbox. It’s free and I respect your privacy (no spam).