Unpacking the AI-Productivity Paradox

We see the effects of transformative new technologies everywhere except in productivity statistics. Systems using artificial intelligence (AI) increasingly match or surpass human-level performance, driving great expectations and soaring stock prices. Yet measured productivity growth has declined by half over the past decade, and real income has stagnated since the late 1990s for a majority of Americans.

What can explain this paradox?

Our close examination of recent patterns in aggregate productivity growth highlights the apparent contradictions. Examples of potentially transformative new technologies that could greatly increase productivity and economic welfare abound, as noted in the 2014 book The Second Machine Age. For instance, consider the recent progress in areas such as machine image recognition (see “AI Versus Human Image Recognition Error Rates”). At the same time, productivity growth has been historically slow over the past decade.

And the sluggishness is widespread, occurring not only in the United States but also in other nations of the Organisation for Economic Co-operation and Development (OECD), as well as among many large, emerging economies.

Some pessimism about future technological progress has spilled over into long-range policy planning and corporate strategy plans. The U.S. Congressional Budget Office, for instance, reduced its 10-year forecast for average annual labor productivity growth from 1.8% in 2016 to 1.5% in 2017. Although modest, that drop implies U.S. gross domestic product will be considerably smaller 10 years from now than it would if productivity simply continued to grow at the same rate as before — a difference equivalent to almost $600 billion in 2017.

Nevertheless, when we review the evidence, we come to a different conclusion and take a more optimistic view. The recent productivity slowdown says nothing about future productivity growth and is no reason to downgrade prospects. In fact, history teaches the opposite lesson. Past surges in productivity were driven by general-purpose technologies (GPTs) like electricity and the internal combustion engine. In turn, these technologies required numerous complementary co-inventions like factory redesigns, interstate highways, new business processes, and changing workforce skills before they truly fulfilled their potential. Importantly, these co-inventions took years or even decades to materialize, and only then did productivity improve significantly.

We believe that AI has the potential to be the GPT of our era. And like earlier technologies, it requires numerous complementary innovations — including new products, services, workflow processes, and even business models — that are often costly and time-consuming to develop. The low productivity growth of recent years may partially reflect these costs and may also be a harbinger of significantly higher growth once necessary co-inventions are put in place.

Accordingly, we see no inherent inconsistency between forward-looking technological optimism and backward-looking disappointment. Both can simultaneously exist. Indeed, there are good conceptual reasons to expect them to simultaneously exist when the economy undergoes the kind of restructuring associated with transformative technologies. Future company wealth and historical economic performance show the greatest disagreement precisely during times of technological change. Our evidence suggests that the economy is in such a period now.

Four Explanations for the Paradox

Specifically, we found four possible reasons for the clash between expectations and statistics: (1) false hopes, (2) mismeasurement, (3) concentrated distribution of gains, and (4) implementation lags. While a case can be made for each of these four explanations, implementation lags are probably the biggest contributor to the paradox. In particular, the most impressive capabilities of AI — those based on machine learning and deep neural networks — have not yet diffused widely.

The first three reasons rely on explaining away the discordance between high hopes and disappointing statistical realities rather than accepting both of these sets of claims. In the misplaced optimism scenario, it’s the expectations of technologists and investors that are off base. The mismeasurement explanation means that it is the tools we use to gauge reality that are inaccurate. And in the concentrated distribution theories, private gains for the few don’t translate into broader gains for the many. The technological promise is there, but dissipative activity prevents the technological returns from reaching most people.

The fourth explanation allows both halves of the seeming paradox to be correct: There is good reason to be optimistic about the productivity potential of new technologies while also recognizing the recent slowdown. It takes considerable time — more than is commonly appreciated — to sufficiently harness new technologies.

There are numerous cases where we see a lag between tech achievements and economic impact. Retailers’ recent experience with e-commerce is a good example. The e-commerce excitement of the 1990s was prophetic, but it took nearly two decades — until 2017 — for online business models to approach 10% of total retail sales. The sector as a whole required the build-out of an entire distribution infrastructure. Customers had to be “retrained” to buy online. Organizational inertia held back innovation in business processes, supply chains, and product selection. None of the needed changes happened overnight, even though the potential of e-commerce to revolutionize retailing was widely recognized, and even hyped. The actual share of online commerce was a miniscule 0.2% of all retail sales in 1999. Only now are companies like Amazon.com Inc. having a first-order effect on more traditional retailers’ sales and stock market valuations. Self-driving cars, medical applications of machine learning, and many other AI breakthroughs will likely follow a similar trajectory.

AI Ramping Up to Its Full Potential

As a GPT, AI will ultimately have an important effect on the economy and public welfare. At the same time, profound and far-reaching restructuring requirements will continue to prolong how long it takes to see the full impact on the economy and society.

What’s more, and what business leaders may find most relevant, is that the required adjustment costs, organizational changes, and new skills can be modeled as intangible capital. A portion of it is already reflected in the market value of companies. However, going forward, national statistics will need to be reinvented to measure the full benefits of the new technologies and their true value.

Realizing the payoffs of AI is far from automatic and will require more fundamental changes than many executives typically imagine. We predict that the winners will be those with the lowest adjustment costs and the right complements. Companies that can best adapt to the new environment will find great opportunities, while competitive pressure awaits those that do not respond nimbly enough. With historical perspective and a realistic road map, we will all be prepared to share in the eventual benefits.