Every year, the Council of Economic Advisers — the White House's internal team of economists — prepares a document known as the Economic Report of the President, reviewing the past year's economy and making projections for the future. It's often a pretty dull affair without much news, but as a few outlets have noticed, this year's ERP contains a striking prediction about the effect of robots and automation on the job market:
The team at CEA used a 2013 paper's estimates of how likely certain jobs are to be automated in the near future, and matched them up with each job's median hourly wage. That let them calculate how likely low-paying, medium-paying, and high-paying jobs are to be automated.
The results are striking: Low-paying jobs (those paying less than $20 an hour, or under $40,000 a year for full-time workers) have an 83 percent chance of being automated. Medium-paying jobs ($20 to $40 an hour, or $40,000 to $80,000 a year) have a 31 percent chance, and high-paying ones (more than $40 an hour, or more than $80,000 a year) have only a 4 percent chance.
This may seem obvious. There are a whole lot of low-paying service jobs you can imagine being automated out of existence. Better Roombas could reduce the need for janitors, self-checkout machines are already replacing cashiers, etc.
But there are also high-paying professions that intuitively appear at risk. Just see this vintage 1998 Atul Gawande article about how artificial intelligence was already better than experienced cardiologists at interpreting EKGs. Radiologists, who spend much of their time visually interpreting test results, are also at risk. So are lawyers who formerly could spend hours scouring paper documents during discovery, charging the client throughout, and now are threatened by "e-discovery" software that makes those files easily searchable.
What the CEA estimates suggest is that these jobs are a minority, and that the overwhelming share of high-paying jobs aren't at risk of being automated out of existence. Or, alternately, the jobs could likely be automated but not entirely so. Maybe radiologists just get better computer tools that increase their productivity and wages without actually rendering the whole profession obsolete.
The CEA estimates also suggest that the "job polarization" hypothesis won't hold going forward. A number of economists, notably MIT's David Autor, have floated the idea that technology has gutted medium-paying jobs, leaving only poorly compensated service jobs (like food prep or janitorial work) and highly compensated, high-skilled jobs (like computer programmers and creative professionals). There's a lot of debate over whether and how much this has occurred, and if it happened in the 2000s, but if the CEA estimates are right, then automation won't cause polarization going forward. The middle class won't suffer the most; the working poor will.
Why the estimates might be wrong
It's important not to take these estimates as definitive. The study it relies on, by Oxford researchers Carl Frey and Michael Osborne, uses a Department of Labor database that lists characteristics of 702 different occupations, and checked each occupation for characteristics (like "originality," "social perceptiveness," and "manual dexterity") that would be hard to automate. Since the database wasn't prepared with automation risk in mind, the paper supplements this with the researchers' own subjective impressions of how likely jobs are to be automated.
It's not a bad methodology, but it has its limitations. As Miles Brundage noted at Slate when the study originally came out, AI researchers have been notoriously bad at predicting what artificial intelligence will and won't be able to do in the near future, casting the study's subjective estimates in doubt. And the DOL database doesn't include skills that all humans have in its occupational descriptions. "Requires basic common sense and the ability to talk language" is a bit redundant when comparing human jobs.
It's also worth wondering just how soon the automation revolution is coming, and how fast it'll be if it gets here. It's true that deep learning technology has notched some major achievements of late, notably beating the world's best Go player just this month. But it's also true that productivity growth in general has been slow in recent decades, casting doubt on the idea that this is a time of unusually huge innovation.
Further, past big technological jolts — the Industrial Revolution in the late 18th century, electrification, the Green Revolution in agriculture, etc. — haven't led to mass unemployment. People with automated jobs (like textile artisans put out of work by industrial looms) were temporarily put out of work, but in the long run new jobs were created, the labor market adjusted, and employment levels stayed roughly constant.
But how fast that transition happens — and how many are displaced in the meantime — remains a crucial question. "As an extreme example, if a new innovation rendered one-half of the jobs in the economy obsolete next year, then the economy might be at full employment in the 'long run,'" the CEA writes in this year's report. "But this long run could be decades away as workers are slowly retrained and as the current cohort of workers ages into retirement and is replaced by younger workers trained to find jobs amidst the new technological opportunities."