Will AI take your job?

 By RP Siegel

It started with the Automated Teller Machines (ATMs), which are now everywhere. They do not provide the same service as a teller (they don’t answer questions, and you can only get $20 bills from them) but they are open all the time and more importantly, they save banks lots of money. Thanks to ATMs, one bank saw its teller transactions decrease by 25 percent between 2014-16. The Bureau of Labor Statistics expects the number of tellers to decline by 8 percent by 2026…

After these developments came automated toll booths and cashier checkout lanes. These are all the result of information technology. Before that, of course, we had jobs displaced by simple machines, as commemorated American folk song John Henry.
Now with recent surges in artificial intelligence (AI) and machine learning (ML) we can expect to see more jobs replaced. Autonomous cars and trucks and perhaps even aircraft (drones) will eventually result in pink slips being handed out to drivers and pilots alike. Of course, there will be new jobs created to produce and maintain these things, but will they match the numbers?
To get a handle on this question, researchers Tom Mitchell, a machine learning expert at Carnegie Mellon University and Erik Brynjolfsson, an economist at MIT specializing in the digital economy and author of The Second Machine Age, got together to assess the trend. The pair defines machine learning (ML) as a subset of artificial intelligence that answers the question, “How can we build computer programs that automatically improve their performance at some task through experience?”

Tom Mitchell from Carnegie Mellon University and Erik Brynjolfsson from MIT

Machine learning is based on a core technology known as neural networks or deep learning networks. These are computer algorithms that can be trained to discover an implicit relationship between inputs and outputs given a sufficiently limited context to operate in. A network could, for example, be developed to see an image (input) and identify it by its name (output).
The biggest breakthrough that machine learning has brought to the world of information technology is that it gets around the need codify knowledge into explicit mathematical formulas. Instead, a machine learning network is initialized with a set of training data, from which each node is adjusted as it contributes to right and wrong answers. In essence, the network learns by experience. But experience in a computer can happen millions of times faster than human experience, making these algorithms a quick study indeed. In time, the very structure of the network comes to embody the implicit input-output relationship, though the learning never stops.
This is the process that lies at the heart of deep learning. Improvements in computer power, as well as innovations in training algorithms, now connect ever-deeper networks to ever-larger training sets capable of addressing increasingly sophisticated tasks.
Mitchell and Brynjolfsson undertook a study that evaluated the work content of jobs, to assess which ones were suitable for machines learning (SML) and to what extent. Using the O*NET taxonomy of jobs used by US Department of Labor (DOL), they obtained a breakdown of 964 occupations into 18,156 distinct tasks which were then amalgamated in 2,069 direct work activities (DWA).
They then developed a set of criteria that could be used to evaluate how suitable a given DWA might be for machine learning. These criteria could then be used to identify SML tasks based on the fact that they:

  1. Involve learning a function that maps well-defined inputs to well-defined outputs. (e.g., classification of image data)
  2. Can access large data sets that already exist or can be easily created
  3. Provides clear feedback (e.g., optimizing traffic flow in a city) of success or failure
  4. Involve relatively simple logic and planning
  5. Require no explanations of how a decision was reached
  6. Have a reasonable tolerance for error (Can’t be 100 percent correct)
  7. Will be applied to a relatively stable process
  8. Require no specialized physical skills or dexterity

They then went to the crowd-sourcing platform Crowdflower (now Figure Eight), a site designed to provide human input to improve machine learning tools, to ask people to rate the DWA tasks against these criteria. Then they folded those direct work activities and their scores back into the occupations, based on the DOL definitions.
What they found revealed a couple of interesting points. First of all, there were no occupations that were either extremely suitable or extremely unsuitable for machine learning. On a scale of 1 to 5, with 5 being most suitable, the highest score was 3.78 which went to both clerks and credit authorizers. Both of these occupations have aspects that are already being performed by computers.
The lowest score was 2.78, which went to massage therapists, no doubt because of the extremely “hands-on” nature of the work. Yet the score was far from the lowest possible. Why? Well, you can already purchase electronic massage chairs that do a decent job and are only likely to get better.
The reason that most of the jobs were grouped fairly tightly in the middle range is because of the difference between tasks and jobs. AI or machine learning can already perform a large number of tasks and will likely conquer many more as time goes on. But most jobs consist of a variety of tasks, some of which can be easily automated, while others can’t.
Those jobs that consist of a single task, such as digging ditches, or collecting tolls or repetitively assembling products, have already experienced substantial declines in their numbers. Still, even those jobs have not been eliminated entirely, since there are certain aspects of the job that still require human oversight, or perhaps, a human touch.

SoftBank Corp's human-like robot named "Pepper" introducing Nestle's coffee machines at an electric shop in Tokyo

In short, all jobs will have some elements that are suitable for automation and others that aren’t. The report concludes:

  1. Most occupations have at least some tasks that are SML
  2. Few occupations consist entirely of tasks that are SML
  3. Any implementation of ML in the workplace will require some level of redesign of job responsibilities

This suggests that the conversation needs to shift away from job replacement and focus instead on job redesign. This, the authors suggest, can be accomplished by re-bundling high and low SML tasks within jobs. Some jobs will likely be lost. Those that remain might have less repetitive content, and then, of course, new jobs will also be created, as change always brings opportunity.

Cover image by: Mike MacKenzie, Flickr

READ MORE: AI and the Generosity Culture by Paola Arpino

about the author
RP Siegel
Skilled writer. Technology, sustainability, engineering, energy, renewables, solar, wind, poverty, water, food. Studied both English Lit.and Engineering at university level. Inventor.