A new working paper from the National Bureau of Economic Research, in Cambridge, Mass., discusses what its authors see as a paradox: although artificial intelligence has made astonishing advances in the last 20 years, taking along with it related technologies and stock prices, real income has stagnated, and productivity seems at best unassisted by these advances.
The first-named author of the paper is Erik Brynjolfsson, of the MIT Sloan School of Management, who is also the co-author of a much praised book on network externalities in microcomputer software.
Brynjolfsson and his associates (Daniel Rock and Chad Syverson) look at a number of possible resolutions of the paradox. They reach their own reflexive equilibrium, though, with the relatively bullish explanation that advances in AI are subject to a considerable implementation lag. Thus, there will be a productivity gain from it all – the gain is on the way – the world simply needs the patience to survive the lag.
Brynjolfsson et al. also include in their discussion some remarks about the national economic statistics. The statistical metrics are outdated and not only don’t fully measure the full benefits of new technologies, some of the measures may “even have the wrong sign,” that is, they may show a negative consequence whereas the reality is positive.
Machine Learning
Machine learning is the AI advance that most fascinates these authors. They write:
Historically, most computer programs were created by meticulously codifying human knowledge, step-by-step, mapping inputs to outputs as prescribed by the programmers. In contrast, machine learning systems use categories of general algorithms … to figure out the relevant mapping on their own, typically by being fed very large data sets of examples.
In this context, the authors refer to ImageNet, a dataset of over 10 million images, as one field for the testing of progress in and by virtue of machine learning. They say that the error rate in computer labelling of the content of such photos was above 30% in 2010, but it had fallen to less than 5% by 2016. Humans also, as it happens, have about a 5% error rate on such a task.
For a practical application in the medical field, these authors observe that “a system using deep neural networks was tested against 21 board certified dermatologists and matched their performance in diagnosing skin cancer.”
In the succinct formulation of Paul Polman, the CEO of Unilever, “The speed of innovation has never been faster.”
Productivity Growth
So, what has happened to productivity? One might have expected the unprecedented “speed of innovation” to speed up the growth of labor productivity. The amount of useful stuff that a given worker can get done with the tools available should almost definitionally improve as those tools improve. And should explode as the value of those tools explodes.
But though productivity is still growing, it is growing at ever slower rates in both the developed and the undeveloped countries of the world. Or, as Brynjolfsson et al put it, “Aggregate labor productivity growth in the U.S. averaged only 1.3% per year from 2005 to 2016, less than half the 2.8% annual growth rate sustained from 1995 to 2004.” The unweighted average annual labor productivity growth rates across a group of 29 countries, (excluding the US, and excluding emerging market nations) tracked by the Organization for Economic Cooperation and Development (OECD), were 2.3% from 1995 to 2004 but have been only 1.1% since 2005.
Further, this deceleration has not been a consequence of the global financial crisis of 2008. It is older than that. In the U.S., productivity growth rates turned down in the late 1990s. In the other mature economies, they turned down in the mid ‘90s.
Optimism
As noted, these authors are optimistic that the new technologies will eventuate in important and positive effects on welfare and aggregate statistics – though perhaps more on welfare than on aggregate statistics, because the latter may be poorly designed for measuring the former in the digitized world. The world is now experiencing a lag.
Why the lag? Brynjolfsson et al posit two distinct components: on the one hand, the time it takes to “build the stock of the new technology to a size sufficient to have an aggregate effect,” and on the other the time it takes to discover and develop complements to the new technologies and their implementation.
If these authors are right, (though they do not put it this crudely) then the ‘lag’ is also from a global macro point of view … a buying opportunity.