Could AI be the answer to our productivity woes?
Regardless of the technology hype, we’re in a productivity slump worldwide. Will it rebound as robotics and AI deliver a boost?
A couple of weeks ago I attended the OECD Global Forum on Productivity in Sydney with the not so snappy title “Keeping pace with technological change: The role of capabilities and dynamism”. My main take outs from the forum were lack lustre productivity performance across the OECD and speculation about whether the current malaise is just the lag between investment in new technology now and the pay-off to come.
First, some tech and productivity history. This graph tells the story of US productivity over the last 70 years.1 (some other advanced economies had a similar experience although the ICT productivity boost, in particular, was the most pronounced in the US.)
The big question is what to expect next? John Fernald, described by The Economist as “the foremost authority on American productivity figures” asked in a recent paper “Is Slow Productivity and Output Growth in Advanced Economies the New Normal?” He noted that
productivity growth has not yet shown signs of recovery …The biggest wild card for productivity growth is the future of technology, and the degree to which the gains from artificial intelligence and robots will spread more broadly across the economy. It could happen. We just do not know when.
Others point to the similarities between the present situation and the 1980s, when Robert Solow, observing the computer revolution, noted that “a technological revolution, a drastic change in our productive lives” had curiously been accompanied by “a slowing-down of productivity growth, not by a step up”. This combination of widespread productivity-enhancing technology and slow measured productivity growth became known as the “productivity paradox”. It was not resolved until the mid-90s, after a productivity step-up appeared in the data.
Yet the step-up lasted for only a decade, supporting the arguments of Robert Gordon and others that ICT has somewhat less dramatic effects that the technology that fuelled prior industrial revolutions.
MIT’s Erik Brynjolfsson and his colleague Chad Syverson at the University of Chicago both presented at the Sydney forum.2They argued that we’re in the midst of a modern “productivity paradox” and moreover, that such paradoxes are just an example of a more general phenomenon when a general-purpose technology diffuses through an economy.
General-purpose technologies are the “defining technologies of their times” and can radically change the economic environment.
They can be products, processes or organisational systems but fit four criteria:
- they are a single, recognisable generic technology;
- there is much initial scope for improvement, but they come to be widely used across the economy;
- they have many different uses; and
- they create many spillover effects.
Electricity is a good example of a general-purpose technology – it became pervasive in homes and factories and successive improvements reduced its cost and improved reliability. And there were spillover benefits for households and workers in the form of safer cleaner environments. The modern washing machine, for example, replaced my grandparents’ “copper” which had a habit of exploding. (Once it spectacularly destroyed the wash house and left both my grandparents with serious burns.)
AI looks on face value to have the characteristics of a general-purpose technology, but it will not be clear for some time – perhaps decades – whether we’ll get a productivity boost from AI; or indeed if history will classify AI as a general-purpose technology.
Some characteristics of general-purpose technologies help to explain delayed productivity effects from their introduction:
- R&D costs are very high in the early stages of their development. Of course, much R&D expenditure goes into a black hole, as firms follow technological leads that don’t turn into profitable products or services. But successful R&D projects lead to patents, trademarks and copyrights. Initially, this shows up in successful firms’ market valuations, but not in their measured profitability or productivity. Products and profitability come later.
- General-purpose technologies tend to spawn new applications as their price falls and knowledge spreads about how to utilise them. But to get the full value from these innovations, adopting firms need to create new business processes, develop managerial experience, train staff, install new equipment etc. Such “complementary investments” are essential to widespread productivity improvements from the new technology.
Brynjolfsson and Syverson say that this pattern of early investment and later rewards creates a productivity “J-curve”. Measured productivity growth slumps in the investment phase, rising as the technology spreads to other firms and transforms production and economic performance.
This curve can stretch out over many years. For example, the technologies driving the British industrial revolution led to “Engels’ Pause,” a half-century-long period of capital accumulation and industrial innovation but also prolonged wage stagnation. Electrification is another general-purpose technology that radically changed production processes, but it took decades. The IT revolution took roughly two decades.
Where are we, technologically speaking? Brynjolfsson and Syverson think that AI (or, more specifically, machine learning) is the next general-purpose technology. Yes, the robots really are coming, but not just yet. The most impressive capabilities of AI are yet to diffuse widely. Like other general-purpose technologies, their full effects won’t be realised without waves of complementary investments by adopting firms. And those firms face adjustment costs, organisational change, and further skills development.
If Brynjolfsson and Syverson are right they could have hit on the explanation for the decline in measured productivity growth in many countries over the last 15 years. But we’ll will only know that for sure with hindsight. We certainly can’t know where the world’s economy is on the J-curve exactly. Brynjolfsson and Syverson are saying it’s somewhere on the downward slope before we get the step-up in productivity on the back of the wide diffusion of the new general-purpose technology.
What does it all mean for us down here at the bottom of the world? Here’s my take:
- While Brynjolfsson and Syverson might have explained poor productivity growth in the rest of the world, their theories don’t adequately explain New Zealand’s particularly poor performance.
- We need to look at how well we’re investing in the things that will allow us to be fast adopters of new productivity-enhancing technology. That’s because research by Nobel Laureate William Nordhaus found that almost all of the benefits of new technologies spill over to consumers and technology adopters rather than the researchers and developers of general-technology R&D. (Using US data for 1948-2001, Nordhaus estimates that innovators captured only 2.2% of the total social surplus from their innovations.)
- If we want to enjoy these benefits, policy in New Zealand would best be directed at removing potential barriers to new technology adoption. Without aggressive adoption, NZ may miss opportunities to complement the skills of our workforce, find better ways of doing things, create new products, capture new markets and improve our wellbeing.
Will New Zealand have too many robots? My worry is that we will have too few.
1. The 70 year productivity graph is based on John Van Reenan, which in turn is based on data from John Fernald
2. The presentations by Chad Syverson “The Productivity J-Curve: How intangibles complement general-purpose technologies” and Erik Brynjolfsson “What can machine learning do and what does it mean for the economy?” at the OECD Global Forum on Productivity, Sydney 20-21 June 2019 should soon be available on the Global Forum’s website.
- Acemoglu, Daron, & James A. Robinson. (2013). Why Nations Fail: The Origins of Power, Prosperity, and Poverty. Broadway Business.
- Allen, Robert C. (2009). “Engels’ Pause: Technical Change, Capital Accumulation, and Inequality in the British Industrial Revolution.” Explorations in Economic History 46 (4):418–35.
- Bresnahan, Timothy F., & M. Trajtenberg. (1995). “General Purpose Technologies ‘Engines of Growth’?” Journal of Econometrics 65 (1):83–108.
- Brynjolfsson, E., D. Rock & C. Syverson. (2018a). “AI and the Modern Productivity Paradox” in Agrawal, Gans & Goldfarb (ed.s) Economics of AI, An agenda, NBER.
- Brynjolfsson, E., D. Rock & C. Syverson (2018b). “The Productivity J-Curve: How intangibles complement GPTs” NBER working paper. https://www.nber.org/papers/w25148
- David, Paul A. (1990). “The Dynamo And The Computer: An Historical Perspective On Technology. American Economic Review 80 (2):355.
- Nordhaus, William D. (2004). “Schumpeterian Profits in the American Economy: Theory and Measurement”, NBER Working Paper No. 10433. https://www.nber.org/papers/w10433
- Fernald, John G. (2016). “Reassessing Longer-Run U.S. Growth: How Low?” Federal Reserve Bank of San Francisco. Working Paper 2016-18. https://www.frbsf.org/economic-research/files/wp2016-18.pdf
- Fernald, John G. (2018). “Is Slow Productivity and Output Growth in Advanced Economies the New Normal?”. International Productivity Monitor 35 Fall 2018, pp.138-48. http://www.csls.ca/ipm/35/Fernald.pdf
- Gordon, R. (2018). “Why Has Economic Growth Slowed When Innovation Appears to be Accelerating?” NBER Working Paper No. 24554. https://www.nber.org/papers/w24554
- Lipsey, Richard, Kenneth I. Carlaw & Clifford T. Bekhar. (2005). Economic Transformations: General Purpose Technologies and Long Term Economic Growth. Oxford University Press. pp. 131–218.
- Solow, Robert. (1987). “We’d Better Watch Out.” New York Times Book Review, (July 12).
Have your say
No one has commented on this page yet.
RSS feed for comments on this page | RSS feed for all comments
AdbotDog 6 Jul 2019, 01:58 (4 years ago)
Very good post. I definitely appreciate this site. Continue the good work!
Superb stuff, Appreciate it!
Robert 5 Jul 2019, 19:45 (4 years ago)
Hi Judy. True, A&R use a broad brush, and predicting subsequent effects is difficult. But it is a useful to ask whether introduction of a particular new technology is expected to increase productivity or keep it at about the same level because you are filling a labour gap. The purpose of a futures project is to explore the second, third, and higher, order consequences of change, so I’d be surprised if the Future of Work project doesn’t explore that.
Editor 8 Jul 2019, 10:47 (4 years ago)
Thanks Robert for your insightful comments and conversation. Even for a specific application of AI, the effects might play out differently depending on the final product market. Take your kiwifruit picking robot. NZ is a very large kiwifruit producer (approx 43% of world exports), so a productivity improvement could lead to a world price reduction and thus an increase in demand for NZ kiwifruit, potentially increasing overall local employment. Whereas a mandarin-picking robot (which could be the same machine with a software tweak) would give a productivity boost in NZ without affecting the world price, as we're not a significant exporter. This example shows why it is very difficult to make generalisations. Dave will explore this further in a forthcoming post. - Judy
Robert 5 Jul 2019, 10:54 (4 years ago)
Having more robots (or machine learning, etc) isn't a magical solution to improving productivity. Acemoglu & Restrepo write about "the wrong kinds of AI" - https://economics.mit.edu/files/16819 - where human labour is automated but without much change in productivity, or creation of additional jobs for people. Kiwifruit picking robots in NZ may be a technological marvel, but they aren't adding value to the product, they are just filling a gap in labour. So reducing barriers to technology adoption needs to be considered alongside social outcomes. Improving productivity, along with economic and social conditions will be as much about our mental models as about software models.
Editor 5 Jul 2019, 11:57 (4 years ago)
Thanks Robert. I don't find Acemoglu & Restrepo's distinctions very useful. It may be possible to determine the difference between "right" and "wrong" applications of AI in retrospect, but the complexity of the second and third order effects (see Dave's post https://www.productivity.govt.nz/blog/reading-past-the-job-loss-headlines-0) means that evaluating them in advance is highly problematic. And it gets even harder should you want to predict social outcomes in addition to net employment effects. Judy