Related Links

• Sectoral Determinants of Aggregate Productivity Growth
• AI and Automation Adoption in the Fifth District and Beyond

Transcript

Tim Sablik: My guests today are Pierre-Daniel Sarte and Thomas Lubik. Both are senior advisors in the Research Department of the Richmond Fed. Pierre and Thomas, welcome back to the show.

Pierre-Daniel Sarte: Thank you for having me back.

Thomas Lubik: Thanks, Tim. Pleasure to be with you.

Sablik: Today, we're going to be talking all about productivity. In economics, productivity is a very important concept. A common saying among economists is productivity isn't everything, but in the long run it's almost everything.

I think most of us have some basic idea of what productivity is — how much we can produce with the resources that we have. But to start our conversation, Pierre, could you explain how economists define productivity? What are its components?

Sarte: The most common and intuitive notion of productivity — which we define as labor productivity —refers to the amount of output produced per unit of labor input. That depends, in part, on the machines or equipment which workers have access to for producing goods and services, which is to say the amount of capital per worker. Labor productivity also depends on how well educated or trained workers are.

More generally, economists refer to aspects of productivity that are not already accounted for by inputs, capital, or labor as total factor productivity.

Sablik: How do economists go about trying to measure productivity growth across the entire economy?

Sarte: Well, let's think of it this way. Any good or service that is produced in the U.S. economy is ultimately produced in some industry. In that industry, we can measure the growth in the production of a particular good or service over a period of time, as well as the growth in the quantity of inputs employed in producing these goods or services during that time. Productivity growth in that industry is then measured as a residual or leftover from the growth in the quantity of goods or services that were produced after accounting for the growth in the quantity of inputs used in production.

If we measure productivity growth in this way in every industry, we can then sum across all industries, using appropriate weights, to arrive at a measure of aggregate productivity growth or productivity growth across the entire economy.

Sablik: You recently wrote an Economic Brief examining the history of U.S. productivity growth. What has been the path of aggregate productivity since World War II?

Sarte: Unfortunately, aggregate productivity growth in the U.S. has tended to decline over time. Of course, there have been ups and downs in productivity growth over the post-war period. But, by and large, as we've gotten further away from the end of World War II, the year-over-year growth rate of aggregate productivity has fallen.

For example, productivity growth was higher in the period from 1947 to 1987 than afterward. Starting at a 2 percent growth rate from 1947 to 1948, labor productivity growth peaked at 5.2 percent in 1950 and bottomed out at -0.17 percent in 1982. The average growth rate of labor productivity is around 1.9 percent from 1947 to 2021, notably below the 2.3 percent from just 1947 to 1987.

It is important to understand that the level of productivity is always increasing over the post-war period. Once we've learned to produce something more efficiently or with fewer inputs, we never really unlearn it, so to speak. However, the level of productivity is generally increasing at a lower rate today than it did in the early 1950s or 1960s.

Sablik: I guess it gets harder to find those new ideas that spark faster growth.

You also analyze productivity growth in different sectors of the economy to see what's driving the aggregate. Which sectors have experienced low productivity growth in recent decades and which ones have grown faster?

Sarte: That's an interesting question, Tim, because there isn't a straightforward answer as it turns out. The world is a complicated place and different production sectors have contributed differently to aggregate productivity at different times.

To make things simple, let's think of the U.S. economy in terms of five broad sectors: construction, which produces structures and buildings; durable goods, which includes, among other things, motor vehicles, industrial machinery, appliances, consumer electronics; intellectual property products, which cover research and development and software; non-durable goods, which includes food, clothing, footwear but also gasoline; and finally, services, which cover leisure and hospitality, health, and education.

It turns out that productivity growth in services is generally lower than in all other sectors throughout the post-war period. Productivity in durable goods has generally been healthy and grew rapidly in the 1990s, reflecting, in part, a significant technological improvement in the semiconductor industry during that period. Productivity growth in intellectual property products has also been steadfast for most of the post-war period, while that of non-durable goods was elevated in the early '50s but then gradually declined since then.

I should note that the service sector accounts for roughly 65 percent of GDP from 1947 to 2021. So, at any point in time, aggregate productivity growth reflects, to a large extent, productivity growth in the service industries.

Sablik: Turning to today, many people anticipate that AI will lead to productivity gains across the economy, allowing us to do a variety of tasks more efficiently. It's something we've talked on the show before with you as well. Based on your research, which sectors do you see AI slotting into, and how might it impact overall aggregate productivity growth?

Sarte: My own research, as well as that done by colleagues at the Chicago Fed and others, suggests that the IPP sector — where software lives and intellectual property products — will receive the lion's share of AI adoption. I do not see this as conducive to the high productivity growth estimates that we sometimes see at present. The IPP sector represents, at most, 13-15 percent of GDP, though it is a sector that has steadily grown over the last 70 years.

Now, it may be that AI eventually diffuses to tasks carried out more broadly outside of the IPP sector, but there isn't much evidence for this at present. Moreover, diffusion takes time and I expect the productivity growth implications of AI to stay with us for a while.

Sablik: Thanks very much for that overview.

Thomas, you've been waiting patiently, and now we're coming to your research on how expectations of future productivity growth from AI could already be impacting the economy today. Can you explain how that works?

Lubik: As you pointed out, this is one of the big questions of the day as we are starting to see the impacts and effects of AI all around us.

Let me go back to something that Pierre said. Total factor productivity, TFP, has become over the last 50 years one of the key concepts in macroeconomic thinking. It determines long-run growth. It also is related to how we think about business cycles.

Business cycles are often described as the outcome of shocks being driven by movements in total factor productivity, in TFP. Now what is often criticized about this idea is that, as Pierre said, once we've learned how to do production, we won't forget it. So, can we really explain recessions with declines in total factor productivity?

I would say in the last 20 to 30 years or so, the idea has taken hold amongst economists that TFP is not this exogenous object that moves around. It also depends on expectations about future productivity growth. This has come to be known as news shocks in the literature. Once a new technology is out there, there's news about it and firms [and] households react to this news.

Sablik: Alright, so how do you go about measuring these news shocks about AI?

Lubik: Yeah, that's the $1 trillion question. It's not a trivial thing to do.

The academic literature has taken three approaches to measuring these news shocks on expectations about future productivity. One strand looks at newspapers, websites, the internet in general, and essentially counts how many times a new technology, like AI for instance, is being mentioned. Google Trends, for instance, is one of those great tools for measuring directly how much news is out there.

A second strand is tied to how we think about the process of innovation and diffusion of new technologies. The idea is there's an inventor and this person has an idea. Once this idea is in the process of becoming marketable, you put a patent for it. So, we can measure directly future productivity, or likely future productivity, by counting how many patents have been filed. There's a large strand of literature that looks at patents filing as a direct measure of future productivity.

The third strand of that literature is my research with co-authors and also many other economists that takes a more indirect approach. It starts with a presumption that when we take data on observed total factor productivity, a large component of total factor productivity is driven by these news shocks. Then, we can use statistical techniques to extract news shocks from observed total factor productivity.

Sablik: Okay, so once you've done that and you've got measurements of these news shocks, what impact do you estimate AI will have on productivity?

Lubik: For a given size of an AI news shock, we can identify what the impact is on consumption, GDP, investment, [and] labor supply. In general, they tend to be quite sizable, but they also tend to be very drawn out. This is tied to the notion of diffusion of innovation. At first, once a new technology is out in the open, nothing much really happens in the data, so we don't really see the impact. But once the new technology gets adopted, it diffuses through the economy. Then, we will see — over time horizons of 10, 20 years — a pick-up in economic activity.

Sablik: You also mentioned that firms and households are responding to this news about AI today. Did you also look at how news about AI is affecting decisions to consume or invest right now?

Lubik: Yes, we did. What is a news shock about future productivity, economists like to call this a wealth effect. When we know that in the future we will be much richer than we are today, then households engage in consumption smoothing. They borrow from the future to consume more today. So, the effect of a news shock today is that we borrow from the future, and households consume more because they know they will be richer — all of the robots that will make our work life so much more pleasurable.

Similarly for investment, if firms know that in the future productivity will be higher, they start investing now to take advantage of the higher productivity in the future.

Sablik: So, to sum everything up, as Pierre already alluded to, there have been a wide range of predictions about how AI will impact the economy. No one can know the future for certain, but based on the research that both of you have done, what do you each expect the impact of AI to be over the next five to 10 years?

Pierre, why don't we start with you?

Sarte: Well, if AI contributes an additional three percentage points by itself to annual intellectual property products productivity growth over the next five years, which would be significant, this will only manifest as an increase of 0.4 percentage points in aggregate productivity growth annually over that period.

Optimistic forecasts of aggregate productivity growth resulting from the adoption of AI range from 1.5 percent to 4 percent annually over the next decade. Given that IPP's current share of GDP is at most 15 percent, productivity in the IPP sector would need to be around 10 percent to 27 percent annually over the next decade to reach those kinds of optimistic aggregate productivity forecasts. For context, productivity growth in the IPP sector has exceeded 8 percent only once in 74 years, thus making it highly unlikely that we see the kinds of optimistic forecasts that I've mentioned.

It may be that AI eventually diffuses to tasks carried out more broadly in sectors outside of IPP, such as retail services or health or education. But current work estimates that, at most, 5 percent of tasks across all sectors of the economy will be meaningfully impacted by AI over the next decade. Therefore, even if the productivity growth implications of AI adoption persist over time, I do not see their effects on aggregate productivity as being nearly as large as 4 percent annually over the next 10 years.

Sablik: Thomas?

Lubik: It depends on the size of the AI shock, and this is something that we cannot identify directly. But what I can say is that for a one-percentage point increase in the news component of TFP — and that's the part that I cannot exactly identify — over a time horizon of 10 years, TFP will increase by two percent.

It all depends on the scale. If AI has a really big news component and a really big effect initially, the scale could be very, very high.

Sablik: Pierre and Thomas, thank you so much for joining me today.