Scott Sumner  

Are robots taking our jobs?

Schneider's The Deadly Sins... My Writing Tutor: Dan Seligman...

Tyler Cowen and Alex Tabarrok have a new debate on "Econ Duel", discussing whether robots are taking our jobs.

Who's right? I don't think we know. Tyler argues in the affirmative, and suggests that this problem will increase over time. Larry Summers has also made this argument. At least in Tyler's case, he's not using one of those old "lump of labor" type arguments. He understands why fears of the effect of automation during the Great Depression later proved inaccurate, or at least premature. Instead, he argues that automation is reducing the job prospects, including the relative wage level, of certain sectors of the labor force---especially less skilled males. Alternatives such as video games are becoming more appealing. He points to the fact that labor force participation for men has been declining for quite some time. Until the past 10 or 15 years, that trend was covered up by rising participation among women. But that rising trend for women has also ended, and perhaps reversed.

Because of factors such as college education and early retirement, many people like to look at the 25-54 age demographic, which is considered a prime age to be working. If lots of people in this age group are not looking for work, then it suggests something might be wrong. Here's the data:

Screen Shot 2016-10-07 at 10.25.03 PM.png
After rising for decades in response to increasing numbers of women working, the ratio peaked at around 84% in the 1995-2000 period. Then it began falling, to about 81%, before edging up to 81.5% in the most recent reading. Whether you consider that 2.5% decline to be "large" is a matter of perspective. It is certainly statistically significant in a country with a labor force of around 150 million people. And the decline is larger in certain sub-categories. On the other hand it's not so large that it couldn't be due to a variety of factors, ranging from more men with previous prison records to more men on disability. A small portion might even reflect things like policemen retiring at age 50, which most people would not view as a social problem.

What makes this tricky is that these factors may interact. Thus robots might replace some of the jobs that ex-cons used to do. In that case, it's the interaction of automation and incarceration.

I think it's possible that automation plays at least a small role in the 2.5% drop---perhaps by depressing the wages of unskilled workers, and making the unskilled work still available less attractive than other options.

In this area, it's always best to try to approach the issue as unemotionally as possible. It doesn't help to draw sweeping conclusions, such as "it's all about laziness, after all the Mexican immigrants can find jobs" or "it's all about a lack of aggregate demand, after all the big drop occurred during the recession." People are complicated and dozens of factors can interact to produce a given outcome. Look at the comment section after their debate, and you'll see how not to think about this sort of issue.

When I think about causation, I approach it in terms of counterfactuals. Suppose we moved away from free trade? Suppose we discouraged automation. Suppose we reduced the benefits paid to middle age people not working. Each policy counterfactual might or might not have much effect, at the margin. My hunch is that reducing benefits and reducing incarceration would slightly boost the participation rate--but nothing dramatic. (Reduced incarceration may be worth doing for other reasons.) I'm agnostic on the impact of trade and automation restrictions, but if pressed I think automation is probably more important than trade. (Read my earlier post on the steel industry, where the effects of trade were vastly smaller than automation.) I simply don't know if the negative side effects (on jobs) of those restrictions on trade and automation would more than negate any positive benefits on jobs through reduced inequality of wages.

To conclude, economics is not (yet) a powerful enough science to tell us whether automation is costing jobs. If robots are taking our jobs it's not the direct effect, as the lump of labor argument is a fallacy. It's not due to falling AD, as the central bank offsets that factor. It's not due to less income going to labor---their share of national income is almost identical to the level back in 1965. Rather it would be due to increasing wage inequality, which reduces the attractiveness of work for low skilled people---especially men.

Also, even if robots have cost some jobs over the past 15 years, we have no idea whether they will continue to do so.

Screen Shot 2016-10-09 at 10.52.20 AM.png

COMMENTS (17 to date)
TravisV writes:

Prof. Sumner,

There was much much much more technological innovation in the 1990's, but that didn't seem to have much negative impact on male labor participation at all. Can we take that as an indication that, in the past five years, increased automation has probably not been that big a deal?

Scott Sumner writes:

Travis, Yes, I suppose that is one indication, although I don't think it's definitive. My best guess is that automation isn't have much impact on employment, but I'm not certain.

David R. Henderson writes:

What is Alex’s argument? Or is it hard to summarize?

Gordon writes:

"To conclude, economics is not (yet) a powerful enough science to tell us whether automation is costing jobs."

Scott, I'm curious. You used the word "yet" which implies that in the future economics will be a powerful enough science to answer this question. Does this come from a belief in a general progression in economics as a science? Or are there specific trends that you have in mind when you made this statement?

john hare writes:

While not robots, labor saving equipment is a constant item that we keep our eyes out for. A friend of mine that supplies our concrete once told me, "If you can replace a $10.00 an hour man with equipment, the savings will make the payments on $100,000.00." I called him out and he just said run the numbers. $10.00 an hour is $1,600.00 a month plus insurance, tax, and office of 30% takes it to $2,080.00 a month for the employee. At the time I was paying $800.00 a month for a $40,000.00 machine. the math works.

Additionally, when work is slow, your best people can often find better jobs if you don't find something for them even at a loss for the company. The equipment will stick around and work 12 hour days for the same payment when the work comes back. With the equipment, I can pay good wages to good people and ignore the non-skilled by not hiring them.

My take is that technology in general is costing non-skilled people jobs, but skilled people can keep going as they can use the ability to learn to follow the flow of work. My company of 4 people at the moment has more equipment than the company of 15 people I worked for in construction in the 70s, except for vehicles to haul the bodies around. I don't need 5 people on shovels for a week as two people with a laser and grader can do the same work better and faster.

Glen Smith writes:

My experience with automation is that the non-skilled worker has little to fear. Go into most businesses that rely on non-skilled labor and you will usually find that they just decide against filling that job. Most of those automated out of a job were skilled. Most who were unskilled competed against air.

Mark Bahner writes:
Also, even if robots have cost some jobs over the past 15 years, we have no idea whether they will continue to do so.

A couple of responses:

1) "Robots" is a not a good word choice for the current situation. When people hear "robots" they think of C3PO, or at least an industrial robot at an automobile factory. Computers are currently *very* limited in their interactions with the physical word, because their vision, hearing, and touch aren't yet very sophisticated. A better word would be "computers," or something that doesn't imply a body that moves around.

2) The easiest way to see whether computers "will continue to do so" is to look at their capabilities at present and in the recent past, versus their projected capabilities in the near future. If one does that, the answer seems very clear: "You ain't seen nothin' yet!"

a) For simplicity's sake, let's say a human brain is 1 quadrillion operations per second (one petaflop). Ten years ago, there was not a single computer in the world capable of that. In 5 years, a petaflop of computing power will cost about $1000. In 15 years, it will cost about $1.

b) Twenty years ago, it was not possible to buy a 1 gigabyte flash drive for any amount of money. A few days ago, I bought a 128 gigabyte flash drive for $40. In 10 years, that same $40 should be able to buy a 10+ terabyte flash drive.

c) In 1996, a 12 megapixel camera cost about $16,000. Now, one can buy a 12 megapixel camera with a cell phone. Similarly, computer voice recognition and comprehension was terrible 15 years ago. Now people can talk to cell phones. So that means computer vision and hearing (especially with regard to voices) can be coupled with the processing power and memory in items “a” and “b.”

d) Computers are becoming increasingly capable of learning by experience.

I’ve done an (admittedly cursory) analysis of the top 15 job categories in 2012, and their likelihood of being susceptible to computers up to the year 2044 (30 years from when I did the analysis). I see tractor-trailer drivers as being essentially eliminated, and big hits for cashiers, retail salespeople, and others. In fact, I predict that in the top 15 job categories of 2012, more than 50% of the jobs in those categories will be eliminated by 2044.

Note: All these numbers are adjusted for population increase. In other words, if the population increases by 20 percent, but the number of jobs in a category stays the same, that’s equivalent to a 20-percent decrease in the number of jobs in that category. In other words, the number of jobs in that category is 0.80 times what it was in 2012, when adjusted for population increase.

Jobs vulnerable to artificial intelligence

Scott Sumner writes:

David, He uses the fairly standard arguments that almost all economists would have used back in 2000.

Recall that at that time the view that automation takes jobs (in aggregate) was almost universally rejected within the economics community.

Gordon, Actually I'm not not sure, but it's possible that economics will progress over time.

John and Glen, Keep in mind that the key is not whose job is directly affected by automation, rather it was the impact on relative wages. If an accountant with a masters degree is replaced by a machine, they are very likely to find another job.

Larry writes:

Could a mix shift explain something? From the DOL:

"Professional and business services employment rose by 67,000 in September and has risen by 582,000 over the year. Over the month, job gains occurred in management and technical consulting services
(+16,000), and employment continued to trend up in administrative and support services (+35,000).

"Health care added 33,000 jobs in September. Ambulatory health care services added 24,000 jobs over the month, and employment rose by 7,000 in hospitals. Over the past 12 months, health care has added 445,000 jobs.

"Employment in food services and drinking places continued to trend up in September (+30,000) and has increased by 300,000 over the year.

"Retail trade employment continued to trend up over the month (+22,000). Within the industry, job
gains occurred in clothing and clothing accessories stores (+14,000) and in gasoline stations (+8,000). Over the year, employment in retail trade has risen by 317,000.

"Mining employment was unchanged in September after declining by 220,000 from a peak in September 2014.

"Employment in other major industries, including construction, manufacturing, wholesale trade,
transportation and warehousing, information, financial activities, and government, changed
little over the month.

Job growth is happening industries that aren't showing productivity growth. They are also mostly non-tradeable industries, little affected by globalization. By contrast, tech and mfg have high and increasing productivity, but little job growth.

Mark Bahner writes:

Scott wrote:

To conclude, economics is not (yet) a powerful enough science to tell us whether automation is costing jobs.

Gordon responded:

Scott, I'm curious. You used the word "yet" which implies that in the future economics will be a powerful enough science to answer this question.

I look at it from the other direction. I predict that in less than 25 years, it will be obvious to anyone that "automation" (or "computers," or "AI", or "robots") are costing jobs. It won't take an economist to see it, and almost all economists will agree it's happening.

P.S. I even put "costing" jobs in quotation marks, because I think the actual effect will be to dramatically lower the market price for many jobs, such that pay for those jobs will drop below minimum wage (e.g. taxi and Uber drivers, tractor-trailer drivers). But this will be at least partially offset by the prices of goods and services going way down...including prices of services like medical and legal services that have previously been rising.

Mr. Econotarian writes:

Trade can amplify automation. If your factory has no competition, there is likely less pressure to automate.

On the other hand, if you are competing with very cheap labor, you may have to automate just to stay in business.

Neal Reynolds writes:

A huge (in my opinion) point that I've never seen addressed in ANY discussion on whether computers and/or robots are good for the job market is the fact that most computers and robots are made in other countries, and most American workers are made (or at least work) in America.

Thus, even if in theory computers/robots wouldn't cost us jobs, it is quite possible they do simply because they (including the software) are often mostly made outside the U.S.

(If this point catches on in discussions, remember you heard it here first.)

ChrisA writes:

When robots are good enough to take peoples jobs, then we won't need jobs anymore.

In other words, if we reach a perfect automated substitute for human beings that is willing and capable to work for free, then essentially we have no further need to work (except perhaps for the personal pleasure of it). Its hard to think that a society like this would think that not having a job is a problem.

Some people get stuck on the idea that perhaps some evil corporation will keep control of robots and not let the rest of us use them and then somehow compete everyone else jobs out of existence.. But that's a silly paranoid argument. If they did charge for their robot labor, we could simply just keep our existing economy with the robot company providing perhaps some lower priced goods like a cheaper Walmart. If the evil corporation gave away it's robot labor for free, well we are back to the first case of free stuff all the time.

While we wait for perfect robotic human substitutes, there will be no worries about humans not having jobs as automation develops. Specialisation assures this - the way it works is that the more automated the world becomes, the richer people become (by definition) and the more richer that people become the more they wish to improve the quality of their lives (or why bother), which means that the quality of the products and services they consume has to increase, which means that the people providing these products and services have to be more skilled, which means that they must be more specialised. So you end up with people who really know how to make your kids party a success. Of course there are morons who say these are not real jobs, and that only digging coal by hand, or driving a plough horse provides the dignity that a real person needs from a job. But revealed preference says not.

Mark Bahner writes:
When robots are good enough to take peoples jobs, then we won't need jobs anymore.

I agree. In theory. And communism works. In theory. ;-) (Wisdom courtesy of the Sage of Springfield, Homer J. Simpson.)

In other words, if we reach a perfect automated substitute for human beings that is willing and capable to work for free, then essentially we have no further need to work (except perhaps for the personal pleasure of it). Its hard to think that a society like this would think that not having a job is a problem.

Yes, in theory. But the potential problem, as I see it, is that there is some multi-year transition period wherein robots are taking a substantial percentage of jobs (say 25-75%) but people still have bills. For example, say you're a tractor-trailer truck driver who has been doing that for 20 years. You're out of a job, and there are *no* other driving jobs (so you're competing with a bunch of other people who recently lost their driving jobs), but you still have a mortgage and utility bills to pay.

So you end up with people who really know how to make your kids party a success.

Yes, but robots could theoretically do that, too. The problem is that we could have 10-20 years where robots (computers/AI, whatever) go from "can't do many jobs" to "can do most jobs." People who are just starting or are in the middle of their careers will have difficulty switching to new jobs when everyone else in their original career choice is out of a job, too.

P.S. Don't get me wrong. In the long run...say 40+ years out...I think you're basically right. I just think the transition may be much more difficult than you project.

ChrisA writes:


Why won't the early phases of AI just look like regular automation, like we have been having for the last 200 or so years? There has been a continual increase in productivity over time, which is another way of saying that jobs have been continually lost in the existing industries. So people have already had to learn new skills and move to new jobs that they haven't been trained for. Why are you arguing that this is a new thing? Do you think it is the pace of the automation that will increase faster than before (singularity?).

Mark Bahner writes:

Hi Chris,

Do you think it is the pace of the automation that will increase faster than before (singularity)?

Yes, that's exactly what I think. In my previous comments, I noted that 10 years ago it was not possible to buy a one petaflop computer, but that in 15 years, one petaflop of computing power should cost $1.

Another way to look at it is in terms of total worldwide computer power. In February of this year, I re-did a calculation I'd done many years previously, regarding the total power of all computers(/microprocessors) in the world.

Recalculating worldwide computer power

By my calculations, all the computers in the world only equaled 1000 human brains in 2003, and only equaled a million human brains in 2015. But in 2027 they will equal 1 billion human brains, and in 2037 they will equal one TRILLION human brains. So basically, in 3-4 decades, computers will go from being able to do very, very few human jobs to being able to do essentially all human jobs...and for much less. Nothing even close to that has occurred in human history.

And I agree that once robots/computers can do all human jobs, and for much less, things should be fine (as long as the robots/computers don’t decide we’re trouble-makers). But in the multi-year transition period, a lot of people won’t be able to compete with computers, but a majority of society will have a tough time with the idea that they shouldn’t just suck it up and get a new job.

Thinking about this issue should be one of the economics profession's top priorities.

Best wishes,

ChrisA writes:


OK I accept that if we are approaching singularity, all bets are off and we can expect great disruption of society. Indeed I believe that unemployment will be the least of our worries in that scenario, the risk of rogue AI is much more concerning to me. But most of the projections of the singularity expect that it will be a very quick process. If you double machine intelligence every year, for a long period, nothing much happens because you are starting at a low base. A bunch of extra morons, even if they are AI, won't make much difference to the world economy, we already have plenty of the human variety. But then one year you reach human level intelligence and then the next year surpass that, and that when things start radically changing. A couple of years of lots of super intelligent beings will result in many radical changes I am sure. So I guess the difference between us is that I don't see the process as being long and drawn out, I see it as brutally fast.

Comments for this entry have been closed
Return to top