John Lewis.
My earlier post arguing that robotisation wouldn’t destroy jobs, slash wages or drastically shorten the working week prompted many thoughtful responses. Richard Serlin and others countered, arguing that if automation affects all sectors, then displaced workers may have nowhere to go. Others asked if the sheer scale, speed and scope of robotisation might make it much more disruptive. Or if wages fall, who will be able to buy the extra output? And Noah Smith raised the prospect that robotisation might eventually differ from earlier waves of innovation by replacing rather than complementing human labour. This post attempts to respond to those points, expand on the original post and explain why I’m still relatively relaxed about robots.
Will broad based technological change leave workers with nowhere to work?
My previous post used Krugman’s hot dog parable of technological change, to make the point that if one sector experiences a productivity innovation, labour moves away from that sector to other sectors. That reallocates jobs rather than destroys them. But Richard Serlin, Mårtin Blix and others countered that robotisation affects all sectors, leaving unemployed workers with nowhere to go.
My point about hot dogs was about the relative allocation of labour when a relative productivity shock hits. But economic theory also has something to say about the case when a shock hits all sectors. The easiest way to think about this is to assume just one sector, so there’s nowhere for the displaced labour to go. Since production functions can be quite dry, perhaps a parable might help…
Suppose that firms own nets (capital) which they give to identical fisherman (labour) to go out and get fish (output) from an ocean teeming with seafood. Adding an extra fisherman to a given amount of net yields extra fish, but that extra output declines for each extra fisherman added. In a competitive market, economic theory says the wage for all fisherman is equal to this marginal output. Similarly, if I give extra netting to a fixed amount of fisherman, more fish are caught, but this extra amount gets less each time.
Now there’s a technological innovation— firms come up with better netting. They can now catch the same amount of fish with less labour. The innovation is “labour saving” in that sense, but does it create unemployment? No. Firms use it to catch more fish. And because the marginal productivity of labour is raised, wages go up too, so there’s someone who can afford to buy them. This is the theory supported by the historical data on wages and output that I showed in the previous post.
What if robotisation means everyone gets hit at once?
Another concern raised was around the short term dynamics— what happens if there is a sharp and synchronised shakeout? In the past specific occupations were rendered obsolete on an idiosyncratic basis – water carriers were made redundant by centralised water systems, switchboard operators by the digital telephone exchange etc. Perhaps with robotisation this shorter term pain comes all at once and there is a “big bang” of resource reallocation with widespread synchronised unemployment spells?
But I’m sceptical that automation represents such a shock. Previous economy wide transformational changes didn’t happen in a short space of time. The industrial revolution, although drastic in the broader sweep of human history, took at least 50 years, the so-called “Technological revolution” around 1900 took around 40 years . If an underlying innovation occurs in a single “big bang” like the railways, or electricity, it can take time for the implications to fan out to the wider economy, even if particular industries are affected much quicker.
And on the scale of change, it’s worth remembering that innovations which lead to stratospheric efficiency gains in a particular task, may have much more mundane effects on economy-wide productivity. Consider the effect of the microchip: since the mid 1950s, computing power has increased approximately a trillion fold in line with Moore’s Law, yet US total factor productivity hasn’t even doubled over that same period.
In a world of mass unemployment and/or low wages who will buy all those driverless taxi rides?
In comment threads on Bank Underground and elsewhere, several people argued that impoverished workers couldn’t afford to purchase the extra output that robotisation might make theoretically possible, raising the possibility of a permanent demand deficiency.
But workers aren’t the only people in the economy. Someone or some company must own the robots, and they must do something with the income— spending on the stuff they like, or adding to GDP via investment. Or perhaps they just save it— but then these savings can be reinvested via the financial system, and if the change is truly transformational there should be plenty of opportunities for that investment. That might not happen automatically, but if desired savings are out of whack with desired investment, monetary policy can usually be adjusted to bring the two into line and get the economy to potential output.
What if robots replace rather than complement human labour?
In the past, new technology has complemented rather than replaced human labour. But in a great reply, Noah Smith raised an important point:
“Economic assumptions are right, until they’re not. The future isn’t always like the past. Sometimes it breaks in radical ways…There is no fundamental law of economics that says that technology must always complement human labour.”
Perhaps automation will be qualitatively different from the technological change we have seen up to now because robots will have a cognitive function. Last year Brad de Long argued that, for most tasks, humans have already been replaced as sources of energy and as sources of manipulation by machines by earlier innovations. He suggested the lack of substitutability between labour and capital is because—up to now— humans have been the sole source of intelligence, sensory perception and reasoning. If robots did one day challenge humans’ cognitive monopoly, then perhaps labour might be eclipsed as an input into production in the same way that horses were in the early 20th century.
I find this a theoretically plausible explanation for how technological progress, at some point in the future, might start to displace labour, because it doesn’t commit the lump of labour fallacy, and it’s consistent with previous waves of technological innovation failing to drive down wages or create widespread unemployment.
But the million dollar question is whether this change in substitutability will occur, and if so, how fast. Neither Noah or Brad believes a drastic change is underway right now, or is imminent, and I’m inclined to agree…
Is substitutability between machines and people low, and will it stay that way?
The evidence on wages in the Robot Macroeconomics post suggests that substitutability has been low. But there’s some evidence that the US labour share has declined a bit over the past decade— from about 66% to 60%. That’s consistent with a bit more substitutability- but also with many other explanations. For example, that US labour has been substituted not with machines but with Chinese labour— a process that might reverse as China becomes richer.
But even if there is some replacement of jobs going on, I remain sceptical that this will lead to mass unemployment, slashed wages or more dystopian outcomes.
First up, even if many of today’s jobs can be entirely replaced by machines, technology can also create new roles. If I wrote down a list of “jobs” at the end of the 19th century, when half the US workforce was employed in agriculture, most of them would have either been rendered obsolete by technical change or have a drastic decline in the number of people doing them. But in that time a whole raft of new occupations – electrical engineer, computer programmer, etc – have been created.
Second, within a given organisation the dynamic response can be different to the static one— jobs can and do get reconfigured in response to apparently fatal technological innovations.
Consider bank tellers vs the cash machine (ATM). A great example of a technological innovation entirely replacing human labour for a particular task. This led to a massive fall in the number of bank tellers right? No. Between the 1970 (when American’s first ATM was installed) and 2010 , the number of bank tellers doubled. As James Bessen notes, reducing the number of tellers per branch made it cheaper to run a branch, so banks expanded their branch networks. And the role gradually evolved away from cash handling and more towards relationship banking.
So will robots redefine the economics of technological change?
So I remain unconvinced that we are on the cusp of a sharp break with the past. Substitutability is subtler than lists of tasks or obsolete jobs. And substitutability aside, technological changes tend to have gradual and less dramatic effects at the macro level.
Perhaps in the distant future robots might displace large swathes of human labour. But unless there are rapid advances in medicine or time travel, I fear I won’t be there to see it…
John Lewis works in the Bank’s Research Hub.
If you want to get in touch, please email us at bankunderground@bankofengland.co.uk. You are also welcome to leave a comment below. Comments are moderated and will not appear until they have been approved.
Bank Underground is a blog for Bank of England staff to share views that challenge – or support – prevailing policy orthodoxies. The views expressed here are those of the authors, and are not necessarily those of the Bank of England, or its policy committees.
It is very true that we can easily be amazed at the way jobs change when they get disrupted, so it is very difficult to anticipate the future. I guess we would be impressed by the number of average hours spend by a human being on intellectual actikvity (including watching video) between now and the 19th century.
To me the fundamental question is not whether robots are gonna be the next catastrophy but what are we going to do with them.
How can we use them to create a better society, and not only to feed the greed of corporations. If robots are to replace humans on more and more non-relationship based tasks, how do we manage that these relationships are actually bettered and not only terrible experience in a call center.
How do we get better teachers, nurses, doctors to make a better society. Not really your usual question to an economist, yet a much relevant one.
The interesting question is: What kind of intelligence will the robots replace? The “raft” of new occupations listed are electrical engineer and computer programmer. The people capable of training for those occupations will certainly succeed. But, what happens to the people with IQs below about 110? Will inequality just get worse? The rise of oxycontin dependency is an example of the consequences of having a state-supported unemployable group.
For me this perspective is much too limited in timescales and on the incremental widening process of automation in any particular sector. An example for me is the author’s strikingly time-limited perspective on ATMs. The polemic put, is, “This led to a massive fall in the number of bank tellers right?”. That’s in order that the author’s immediate riposte of ‘No’ can by put. In fact the process of automation has proceeded increasingly apace following the introduction of ATMs. I’d posit that the (relatively short-lived) expansion of the number of bank branches in the USA had far more to do with the wider, ‘exuberant’, trends that took us to the 2007/08 crash – rather than to do with the introduction of ATMs. In the UK our socio-economic landscape has been riven with the closure of literally thousands of local banks (and with them the tellers’ jobs). Automation is now increasingly at a possibly faster rate with the advance of digitilisation (hence, ‘nowhere for displaced tellers to go?). We now encapsulate part of the range of this systemic and pervasive change in the term fintech.
there are flaws with your labor argument. One, the value of labor is driven entirely by demand. In your fishing analogy, there is some cap on total demand for fish. Adding new technology into fish production reduces the cost of production. But it does not raise demand. If the price falls, more people may be able to afford fish, and you may maintain equilibrium of demand to labor, but at some point, total demand will be satisfied by less labor. that is the outcome of all productivity gains. As the fishnet building process improves, the marginal cost of fish production will drop towards zero. This will happen until, hypothetically, a single operator could catch all the fish necessary to satisfy total fish demand. This is the essence of robotics. A single operator could manage the entirety of production. this hypothetical is catastrophic for jobs in fish production, but it’s wonderful for the fish consumer, because the cost of fish is near zero.
the underlying point being it’s the cost of goods and services relative to the demand that has impact on standards of living and survival. As prices drop towards zero, less labor is used in production and conversely less wage earning is needed for purchasing power.
Your argument contains an implicit assumption: that labor has a zero cost. It does not. Human labor always comes with a training cost. this training cost occurs at large as “public” education, but it also occurs locally in firms. the costs of human training cannot drop towards zero because of the rate of human learning. A new employee takes time to train. Automation has a substantial effect on labor because of this cost of training.
In existing firms or existing jobs, the cost to automate must be some factor less than the cost to continue to do business as is. Which means we don’t see automation effects in existing enterprises until there is something like a 2-10x improvement productivity or cost reduction. the risks of automation must be weighed against potential gain. However, in new jobs, there is no inherent risk, because new jobs require developing and implementing training + the costs labor. If automation is 1:1 to the costs of of labor, the automation costs are competitive with new job creation, because the training cost is zero. Thus we should see automation occur the most in new firms. And that is exactly what we see. New firms take advantage of automation when the price of automation is equivalent with the cost of labor because it reduces training costs to zero. and because these firms do not have job-replacement costs.
With a general purpose robot, such as Baxter, the training cost of one robot can be spread across a million robots with zero cost. But more importantly, it can be replicated across a million robots in near zero time. The advantage human beings have relative to computational solutions, like robots, is that they are trainable in some reasonable amount of time. Automation will continuously improve against the time to it takes to train a person as computational power continues to improve. Automation solutions which can be implemented more quickly than human training will come to dominate jobs. This has a cascading effect. Because training time is the primary source of labor wage disparity. If a doctor could be trained in months not years, there would be many more doctors, and hence health prices would fall. If some of the jobs doctors do now can be implemented in software, then they can be distributed to everyone with a cell phone. This very thing is happening now.
The cost to automate will continue to fall against the cost to train. Increased automation will tend to have a cumulative effect. At some point in the future, nearly every new job that requires training which could be automated will be automated before it can turn into a job listing, because that will just be easier.
As far as I can tell, there is only one kind of job human beings do that could never be automated by some kind of system. And that is human learning. Human teaching of course can be highly automated and cheaply distributed, and we are seeing that very thing.
Your fish parable assumes a large (infinite) supply of fish, and expanding market demand. In my world this is unlikely.
there are flaws with your labor argument. One, the value of labor is driven entirely by demand. In your fishing analogy, there is some cap on total demand for fish. Adding new technology into fish production reduces the cost of production. But it does not raise demand. If the price falls, more people may be able to afford fish, and you may maintain equilibrium of demand to labor, but at some point, total demand will be satisfied by less labor. that is the outcome of all productivity gains. As the fishnet building process improves, the marginal cost of fish production will drop towards zero. This will happen until, hypothetically, a single operator could catch all the fish necessary to satisfy total fish demand. This is the essence of robotics. A single operator could manage the entirety of production. this hypothetical is catastrophic for jobs in fish production, but it’s wonderful for the fish consumer, because the cost of fish is near zero.
the underlying point being it’s the cost of goods and services relative to the demand that has impact on standards of living and survival. As prices drop towards zero, less labor is used in production and conversely less wage earning is needed for purchasing power.
Your argument contains an implicit assumption: that labor has a zero cost. It does not. Human labor always comes with a training cost. this training cost occurs at large as “public” education, but it also occurs locally in firms. the costs of human training cannot drop towards zero because of the rate of human learning. A new employee takes time to train. Automation has a substantial effect on labor because of this cost of training.
In existing firms or existing jobs, the cost to automate must be some factor less than the cost to continue to do business as is because of hidden costs and risks in changing business processes. Which means we don’t see automation effects in existing enterprises until there is something like a 2-10x improvement of productivity or cost reduction. the risks of automation must be weighed against potential gain. However, in new jobs, there is no inherent risk, because new jobs require developing and implementing training + the labor costs. If automation is 1:1 to the costs of of labor, the automation costs are competitive with new job creation, because the training cost is zero. Thus we should see automation occur the most in new firms. And that is exactly what we see. New firms take advantage of automation when the price of automation is equivalent with the cost of labor because it reduces training costs to zero. and because these firms do not have job-replacement costs.
With a general purpose robot, such as Baxter, the training cost of one robot can be spread across a million robots with zero cost, the marginal cost of training drops to zero. But more importantly, it training can be replicated across a million robots in near zero time. The advantage human beings have relative to computational solutions, like robots, is that they are trainable in some reasonable amount of time. Automation will continuously improve against the time to it takes to train a person as computational power continues to improve. Automation solutions which can be implemented more quickly than human training times will come to dominate jobs. This has a cascading effect, because training time is the primary source of labor wage disparity. If a doctor could be trained in months not years, there would be many more doctors, and hence health prices would fall. If some of the jobs doctors do now can be implemented in software, then they can be distributed to everyone with a cell phone, so the marginal cost for those particular jobs which doctors perform is zero. This is already happening.
The cost to automate will continue to fall against the cost to train, and increased automation will tend to have a cumulative effect. At some point in the future, nearly every new job that requires training which could be automated, will be automated, before it can turn into a job listing. Because, automation will be faster. New jobs are constantly being created. The question that matters to firms is, can the new jobs be automated in the same or less time than it takes to train a person ? If that answer is yes, automation will race ahead.
There is one kind of job human beings do that could never be automated by some kind of system. And that is human learning. Human teaching of course can be highly automated and cheaply distributed, and we are seeing that very thing.
“but if desired savings are out of whack with desired investment, monetary policy can usually be adjusted to bring the two into line and get the economy to potential output”
Interest rates have been on a downward trend for several decades with unconventional measures like QE being shown to not be very effective and have diminishing returns.
What if in a more automated, more productive world, the Wiksellian rate is deeply negative? You can’t just assume that monetary policy can fix everything anymore.
It should be considered the biggest crisis in macroeconomics.
This won’t do. “I won’t be there to see it” is not a reason to dismiss a looming problem.
Remember the riddle of the lily pad, since you like stories, and apply it to the spread of AI.
And try broadening your historical reading. A better example than the hoary ATM one is the spread of the tractor in the USA and elsewhere in the late 20s and thirties, and the resulting transformation of employment. How easy and peaceful was that? Without a war, the world may have remained stuck in a permanent depression.
(Note that although ATMs in the 1980s resulted in an increase in points of presence for banks, internet banking is now in the 2010s being cited by banks as the reason for closing branches. removing ATMS, and reducing employment in the remaining branches. At least in my country.)
I have now had a chance to read your post carefully:
If you look at my post in response to your first post (http://richardhserlin.blogspot.com/2016/09/ai-and-krugmans-hot-dogs.html), which is much more complete and polished than my comments, I think you’ll see I’ve covered your points, and why I think they haven’t been enough to stop a plummeting in income and employment security for low-skilled workers over the last two generations. And, why there is a substantial probability that AI and robotics will be catastrophic to the employment and wages of the unskilled over the next generation or two.
But, I’d like to note some specifics here.
You write:
“The easiest way to think about this is to assume just one sector, so there’s nowhere for the displaced labour to go. Since production functions can be quite dry, perhaps a parable might help…
…Now there’s a technological innovation— firms come up with better netting. They can now catch the same amount of fish with less labour. The innovation is “labour saving” in that sense, but does it create unemployment? No. Firms use it to catch more fish. And because the marginal productivity of labour is raised, wages go up too, so there’s someone who can afford to buy them.”
As I noted in my post, there are possible problems with this. One is, if there are more than two products, say, you can hit inelasticities of demand. People can and will only eat so much fish. The price of fish can plummet, and so if the workers only have the skills to produce fish, and it’s too expensive and/or difficult to obtain new skills, then they will be forced to accept plummeting wages. And the wage can easily drop below subsistence.
Now, you may say, at least they will be able to eat fish, which would be as cheap as their wages, but:
(1) Having fish, but nothing else makes for a short harsh life, as you don’t have a home, medical care,…
(2) Even adequate fish might not be affordable, as low-skilled labor is not the only input necessary to get fish. You also need relatively rare and expensive high-skilled labor, and raw materials to make the boats, refrigeration, radar, etc., to run and manage the enterprise.
And so again we see the problem. If robots and AI’s can now do 90% of the unskilled work, so that you would need 10% as many low-skilled humans supplied to keep the market wage the same, you can’t just say ok we’ll produce and eat 10 times as much fish. There are bottlenecks in high-skilled workers, raw materials, and other capital. And there is an inelasticity of demand. The people with money will shift to other products with their increased real income, and those will tend to be higher prestige goods, that may tend to require a greater percentage of skilled labor, not lesser. Supply of low-skilled labor may go up 10-fold or more, with the influx from Robotia and the AI Republic, but demand will not keep pace, and so the market wage will plunge.
You acknowledge, “That’s consistent with a bit more substitutability- but also with many other explanations. For example, that US labour has been substituted not with machines but with Chinese labour” So this is key. If it can happen for low-skilled Chinese labor, then it can happen when robots and AI’s get equivalent abilities to low-skilled Chinese laborers, and perhaps at a far lower wage than even they have. And with much higher quality and reliability. So, this is the key, how good will machines get at substituting for low-skilled human labor (higher skilled too, but from my study, low-skilled is the far bigger employment threat over the next 10 – 30 years).
You write, “Perhaps in the distant future robots might displace large swathes of human labour. But unless there are rapid advances in medicine or time travel, I fear I won’t be there to see it…” But based on the age implied by your CV, and your photo, you should have at least 40 years. I’m stunned you could be as confident as you sound about how good – or actually not that good – AI and robotics will get over the next 40 years. Have you studied this technology extensively to be as confident as you sound?
I have read thousands of pages from top experts, including AI textbooks. My opinion is that there is a very substantial probability that AI’s and Robots will substitute for the vast majority of low-skilled human labor over the next 40 years. And, in fact, I think, there is a significant chance that substitution will be wide scale as soon as 10 years from now, with a far greater affect than that of Chinese labor.
So, what evidence can I provide for this from my thousands of pages of reading and study over the last few years? A lot is in this post, which was in Mark Thoma’s links:
http://richardhserlin.blogspot.com/2015/11/robotai-revolution-decimating.html
But I’ll add something here. There are five major methods of machine learning (which basically is artificial intelligence), and all five can be powerfully combined: (1) Statistical modeling and methods, (2) Symbolic logic, deduction and induction, (3) Analogy (4) Neural networks, (5) Evolutionary learning, where algorithms compete for the highest utility score, and mutate, and may even reproduce sexually, mixing parts of the algorithm of each parent.
Now, I’d like to focus here on (2) and (3), as our human brains use all five of these methods also to create intelligent thought, either implicitly or explicitly.
Consider the computer driven car. When will we have reliable effective computer driven cars, as safe as a human, as fast as a human, as reliable and competent, as inexpensive as a human driver – or much more so for all of these things?
What single group of people would know best the answer to this?
The top experts working on it, or in related areas.
And who are the best people who fit this bill? People who work at companies on projects to develop AI driving commercially; ultra-smart, ultra-educated specialists who have worked massive hours on this, lived it, for years.
And what are they saying, where it really counts, with their money? And when their mouth being widely off can cost enormous money to their company, and personally to their career.
Key companies are spending billions, and they would not do this if success were a maybe, maybe, it will happen in 80, 100 years. We all know how powerful the time value depreciates far future money at the very high discount rate you would apply if you thought this was very unlikely and risky.
And yet, these companies are spending billions, tens of billions jointly, per year, and making projections for commercially available, affordable, mostly or completely autonomous cars in 5-20 years. Many things that could be cited to show this, but given time and space, I’ll just start you out with this one:
http://www.motherjones.com/kevin-drum/2016/08/heres-how-driverless-cars-will-happen
Now let’s use, (2) Symbolic logic, deduction and induction, and (3) Analogy and categorization.
If an AI can drive a car, on dense, unpredictable, urban roads, faster, better, and cheaper, than a human, with the incredible sensory ability and judgement, and flexible thinking, that takes, then those same skills are basically what it takes to do almost any manual labor. It’s very little stretch to then have a robot navigate a fast food restaurant, a home as a housekeeper, a yard as a landscaper, a grocery store as a stocker, and to watch and identify correctly what shoppers put in their carts, so no need for a human checker, …
And as far as the dexterity of human hands, we’re not that far from that now.
Now, there’s far more evidence and logic than this that makes me think that it could be within 10-30 years before most low-skilled human labor is substituted for, but as this is already the longest comment in history, I’ll leave it at this. Please feel free to let me know if you’d like me to provide more.