Erik correctly reports that there is no correlation between productivity growth rates in successive decades, and that consequently the slow productivity growth of the previous 2010-19 decade has no implications for the current decade of the 2020’s. In guessing what the past means for the future, “it’s a coin flip.” Lacking guidance from the past, he asserts that we must look at the specific technologies being invented and how they are being deployed. Two of these are artificial intelligence (AI) and machine learning.
The problem is that Erik has been preaching the “staggering potential” of AI and machine learning for almost ten years, since our 2013 TED debate and the 2014 publication of his book The Second Machine Age. AI has developed and matured since then and yet productivity growth has not responded -- the annual rate of productivity growth between the early 2013 date of our TED debate and the latest data for early 2022 is an underwhelming 1.1 percent per year in the U.S. private business economy, the slowest for any decade in American history.
Why aren’t robots and AI having a greater impact on productivity growth? One reason is that they are surprisingly unimportant forms of business investment. Even though the number of U.S. robots has doubled in the past decade, annual spending on robots is less than one percent of the more than $1 trillion of annual U.S. business equipment investment. In 2019 there was only one robot per million hours of work, and 75 percent of those robots were in two sub-industries within manufacturing that account for barely 5 percent of GDP. Similarly, annual worldwide AI spending is about one percent of total equipment investment.
Simply put, AI has so far done a poor job of replacing the human brain. “Deep learning” is what delivers facial recognition, language translation, and Amazon book recommendations. Super-fast computers are fed millions of images of cats and dogs (that is the “deep” part), and this allows them to look at a photo and conclude “this is a cat” or “this is a dog.” But that doesn’t allow them to know anything that humans instinctively recognize about the behavioral differences between cats and dogs.
Autonomous vehicles, a potentially fruitful application of AI, have been just around the corner for years but they still haven’t arrived. This has turned out to be a tougher problem than facial recognition, because computers must teach cars not only to recognize objects in the field of vision but to predict where they are going to move next. And even deciphering fixed objects is imperfect – in an example displayed at a Congressional hearing, three white strips placed on a red stop sign caused a car’s computer to misinterpret the stop sign as displaying “45 MPH zone.” As one expert has concluded about autonomous vehicles, “the first 90 percent was easy, the last 10 percent is 10,000 times as hard.”
Erik has done impressive research on the value to consumers of free internet services. Yes, this means that GDP growth is understated. But this omission of the “consumer surplus” of new inventions has always been true. Consider some of what I call the “Great Inventions” of the 1870-1930 era. Consumer welfare leapt ahead when electric light and appliances replaced candles, kerosene lamps, iceboxes, and scrub boards. Health and well-being surged forward when motor vehicles replaced horses and their manure strewn over the streets. Running water and sanitation infrastructure allowed a transition from outhouses and pails of carried water to kitchen sinks and indoor bathrooms. Social networks and free GPS are not the first instances of GDP mismeasurement and will not be the last.