Tyler Cowen on the AI gains GDP misses
More free time, better services and fewer wasted purchases can make us wealthier without showing up in the headline numbers.
Like most AI-pilled commentators, I believe that artificial intelligence will boost GDP significantly by increasing production in virtually every area of the economy. It could drive the creation of more startups, more efficient factories, better governance systems and, eventually, more and better medical cures and treatments. But my concern in this essay is different. I’m interested in the ways that AI could help increase our individual and collective economic well-being that GDP alone doesn’t capture.
GDP, by definition, can be measured. It gives us a first-order approximation of the economic gains from artificial intelligence. Many of the effects of AI may not show up in measured GDP, but can boost human welfare nonetheless. We all know that GDP is an imperfect measure, not just for human welfare but also for the even narrower question of how much an economy is enjoying gains in productivity.
In this piece, I will thus consider non-GDP-measured gains from AI. These are important because they show that the practical gains from the technology are higher than our current numbers are indicating. AI will have some economic benefits that are not so easily measured or verified, even though anecdotally, we are pretty sure they exist. This in turn helps us perform a more complete, and I would suggest more positive, analysis of the benefits from AI.
Among the more hidden economic benefits AI could bring, I will consider leisure time, the gains from better matching of consumers to products and the gains from higher-quality services. I will also consider the related question of whether we are undervaluing the quantity of investment, due to the special nature of how AI services are produced. This, too, will have implications for our economic future, relative to what the normal economic data otherwise suggests, and it points to a greater importance for America’s tech sector in the economy at large.
Leisure time
Many of the gains from current AI usage spring from additional leisure time.
Many of the gains from current AI usage spring from additional leisure time. Remember that work memo that used to take you half a day? Now you can finish it in half an hour. Ask Claude or ChatGPT to do the first draft, make some edits yourself and then smooth out the parts that sound “too much like AI.” Voilà. Some people object to the slop-like elements of this process, but often the results are better than what humans would have done on their own.
Conversations at the frontier, in your inbox
SubscribeYou now have extra time, which you might spend chatting with colleagues at the water cooler, resting or watching short videos. In any case, you are having more fun, though those gains are not registered in GDP. The same memo ends up being written, but it took you less time than it did in the past.
In these scenarios, the short-term, medium-term and long-term effects of more AI use differ in meaningful ways.
In the short term, measured GDP is flat, and workers have more leisure time and more fun. Right now, most of us are living in the short term.
In the medium term, bosses realize that workers are using AI to write memos, thereby saving time, and thus the bosses assign the workers more tasks. The workers still probably have more leisure time, but not as much as before. And now the company actually is getting some more concrete work done. GDP rises eventually. But since GDP numbers have yet to boom, we are at best just now starting to live in this medium term, even though most programmers and the AI companies themselves have been living in it for a while now.
In the long term, production is reorganized entirely, most likely beginning in new startups. Perhaps that memo-writing job does not exist in the first place, because AIs are doing that work with only minimal human supervision. The human workers who used to write the memos are now on the phone talking with customers, or perhaps they are part of some new AI-enabled venture altogether, such as testing a new drug suggested by AI biomedical models. GDP is higher still. Leisure time might be higher because the world is wealthier and some individuals might retire earlier than they otherwise would. But many of the first-order positive effects on leisure time would vanish. In fact, due to higher wages and higher potential returns, a lot of individuals would probably end up working harder, and being happy to do so.
We are not, however, living in this long run—at least not yet. We remain in the early stages of powerful AI, as evidenced by the fact that labor market churn, circa mid-2026, is at historically typical levels. It is frequently debated how long it will take to get us to this long run, but there is no simple way of settling the issue other than to sit back and watch the data come in.
I started with the example of memos, but the operative principles extend to other activities. If you are working with data, AI agents can help you clean, organize and present your data much more easily than before. Those are usually time-intensive tasks, and furthermore, they require expertise and intelligence. The opportunity cost of doing that tedious data work is high, since the people who used to do it typically had significant skills in other areas, including coming up with new creative hypotheses.
The leisure time benefit can take many forms. For instance, my wife and I have a dog, and once we saw that the dog had a rash. In the old days, we would have taken the dog to the vet for a diagnosis, which might have cost us hundreds of dollars. It also would have taken us time, involved some stress and, of course, the dog would not have enjoyed the trip either.
Instead, we took a photo of the rash and asked our friendly LLMs whether we needed to take the dog to the vet. They hemmed and hawed a bit, not wanting to incur legal or reputational liability for a mistaken answer. But basically, they said no, suggesting the rash was a pretty normal and temporary development. We did not take the dog anywhere, and he remains fine.
In the short run, GDP might even go down because of an episode like that. The vet’s bill for the dog would have been recorded as a component of GDP, since it was the sale of a service. But that does not mean that GDP would be permanently lower. We will either spend those funds elsewhere or keep them in the bank, which will lend them out to businesses or to consumers for a mortgage or credit purchase. A bit past the short run, GDP would go up again, and economic output would end up being less wasteful and better targeted than if we had made the mistake of taking the dog to the vet.
Or consider the use of AI as a therapist or conversation partner. Without AI, people would spend more money on therapists. Even if you think human therapists are better than the AIs (debatable), a lot of people simply cannot afford good human therapists, who may cost hundreds of dollars an hour. AI serves as a useful substitute. It is not that AI therapy always works out for the best, but it does seem to date that millions benefit from it, and it would likely pass a cost-benefit test.
In the short run, as with the vet, there is no transaction, and GDP might fall. Many people are better off as a result of having AI as an option, however, and soon enough, GDP would recover as the saved funds are deployed elsewhere.
Leisure time also fits into this analysis. The costs of a therapist include not only the upfront fee but also the time and aggravation costs of getting there. In many American cities and suburbs, traffic congestion is a problem. A therapist does not have to be that far away for the trip to take 30 to 40 minutes or longer, given that you should add in a cushion not to be late. In contrast, if you are using an AI therapist, you can do that from home, and the traffic costs are zero. You can spend that extra time exercising, which by some accounts can be just as effective as a therapy session.
Better matching
Some of my biggest personal benefits come from very basic, “low-level” uses of AI. For instance, I often use AI instead of traditional search. I took a summer vacation in Europe and asked some AI models which concerts I should go see.
I was led to a Grigory Sokolov piano concert in Wiesbaden that I otherwise might not have found. ChatGPT Pro told me I should go! And I did. It was one of the best classical music concerts of my life, and the tickets were not expensive. My consumer surplus, to use the term from economics, was very large. The experience was worth at least $1,000 to me, but the ticket price was a mere fraction of that.
In a world without AI, I likely would have seen a good concert, but probably not nearly as good as Sokolov, whom I had been wanting to hear for decades (he does not tour America). The ticket price would have been roughly the same, but my consumer surplus lower. In the AI world, I matched to what I wanted far more effectively. The AIs know my musical taste, and furthermore can search a wide variety of websites far more quickly than I can.
It is not just concert tickets. I use AI to select my hotels, often my meals and even the destinations for my vacations. For instance, I was recently staying in a town near Heathrow Airport outside of London, and the AI models told me to go see St Albans Cathedral. It was wonderful, and the visit was absolutely free and thus not registered in GDP (though my Uber trip will be). I know of many other people who do the same with their travel, and indeed with many of the other goods and services they buy.
These days, you can use AI to improve the quality of virtually everything you purchase, and that represents a real gain in human well-being.
These days, you can use AI to improve the quality of virtually everything you purchase, and that represents a real gain in human well-being. To be clear, GDP statistics do adjust for the quality of customer experience in a number of ways. For instance, if cars become safer due to better airbags, that typically will register as a quality improvement; in time (with possible lags), that will make its way into GDP. (In essence, we adjust price indices so that improvements in quality count as real improvements in the economy for the purposes of calculating GDP.)
When it comes to gains in consumer matching, however, we have no such comparable process. There is no one in the economic statistics office who can register how much I enjoyed my purchases or not. One way to think of these matching gains is to say that AI helps consumers get hold of more customized products. The effective, available degree of product variety rises sharply due to AI. But again, that effect is not captured by typical, quality-based adjustments in how GDP is calculated.
Note that new, customized products typically involve higher gains in consumer welfare than older, established products. If there is a new product, or at least a product newly available to you (perhaps for informational reasons), ex ante the effective price of that product to you was infinity. You simply could not get it. But when that product becomes available, typically it does so at a pretty normal price. That is, in essence, a pretty large price decline, and if you value the new product a lot, as I did my Sokolov concert, the gains can be large. It is more difficult to reap comparably large gains by simply purchasing more of an established product that has been in front of you as a buying opportunity for a long period of time.
The superior search services provided by AI also probably induce service providers to boost the quality of their offerings, again in directions that GDP adjustments do not readily pick up. During one of my recent doctor visits, I let it be known that I was running the data from my blood tests through advanced AI models and comparing their outputs to what the doctor told me. In other words, I told my doctor that I was doing a more intense and better job of monitoring the quality of his services and his attentiveness. Presumably, at least in some cases, the doctor who knows he is being “watched” pays more attention and does a better job. If my doctor told me some questionable information about my blood test, I could inquire as to why he was making that (possible) mistake. In these cases, we can expect quality to rise, but once again, without direct impact on GDP figures.
If a plumber comes over to fix your sink, ask Claude in advance what the proper procedure is. That will make it harder for the plumber to rip you off and do excess work for an excess price. Again, the economy is made more efficient and better for consumers. If you wish, you can think of this as another instance of matching—in this case, matching yourself to better-quality plumbers and doctors.
Data and investment
A focus on GDP also overlooks the fact that millions of people, knowingly or not, are investing their time and energy creating data for the AIs.
I believe that we are also undervaluing investment in an AI-rich economy. A focus on GDP also overlooks the fact that millions of people, knowingly or not, are investing their time and energy creating data for the AIs. The simplest example of this is people writing for Wikipedia or typing Reddit threads, both well-known sources for AI models. We do not usually think of these activities as investment, and indeed, in the past usually they were not. Nonetheless, these days they count as investment into the quality of AI models because they are generating useful data, even though they are not recorded in national investment accounts in the standard manner. Investment is one part of GDP, and furthermore, this point implies that the contribution of the tech sector to GDP growth is higher than we had thought.
As Nav Vaidhyanathan wrote on Substack, “Billions of people perform trillions of hours of productive activity—generating the behavioral data that trains AI systems, calibrates advertising markets and drives the market capitalizations of the world’s most valuable companies—and this activity is classified as idle leisure because no fiat wage is exchanged. The System of National Accounts, in its current form, cannot see it.”1
Vaidhyanathan roughly estimates the magnitude of this effect at $1.3 trillion a year globally. Whether or not that is the correct number, we all know that people spend plenty of time writing content for the internet. And Vaidhyanathan’s number does not count the time people spend writing books, or for online periodicals. Those activities now also count as investment, because so many books and online articles are “fed” to the AIs as useful data.
Another study, by Maryam Farboodi and Laura Veldkamp, examines the role of apps. Many apps are given away to consumers for free, in large part because the app-issuing enterprise will receive data from those consumers. The researchers note correctly that the enjoyment and practical gain from the app should count as a service. In essence, the consumers offer data as a kind of barter trade for the services of the app. Typically, that exchange is unpriced and not recorded for the purposes of national economic accounts. Yet it is a useful form of economic activity.2
Farboodi and Veldkamp estimate that the value of all this data, circa 2018, is about 6% of GDP. To be clear, and as the temporal range of the data indicates, most of this earlier app-related data probably has not been fed to the AIs. Still, this study is useful for showing that the value of data is very high, that our current economic statistics do not measure this value well and that American economic activity is more robust, and more dependent on the tech sector, than we had thought.
As the AI sector grows, these more general points are likely to prove more relevant yet. For instance, we can imagine that consumers will be given free services, such as the use of a Roomba “cleaning robot” for their houses. In return, the company supplying the robot will receive data, useful for improving the robots for the future. In this scenario, the consumers receive something for free; this does not go into GDP, yet ought to, as it is a useful service to those who receive it. In a similar fashion, imagine receiving something akin to free Zoom calls, but again on the grounds that the conversation will be used as data. The possibilities are numerous, and I expect further developments along these lines. The net result, again, will be that GDP undervalues the human welfare gains from AI.
In sum, there are a few possible mechanisms through which strong AI will boost our economy in ways we do not measure well. This is overall a reason for optimism. We might wish to consider improving our economic statistics, but in the meantime, we should realize that unmeasured economic gains from AI are a significant force indeed.
Notes
1 See Nav Vaidhyanathan, “Dark GDP: The Economics of Entangled Time,” April 12, 2026, https://navvaidhyanathanatvysdomai.substack.com/p/dark-gdp-the-economics-of-entangled. Do note that most developed economies are working on incorporating data into the national accounts in a more accurate manner. These projects, however, for the most part are not complete and do not yet show up in published GDP and investment statistics.
2 See Maryam Farboodi and Laura Veldkamp, “A Model of the Data Economy,” Review of Economic Studies, July 6, 2026, https://www.restud.com/a-model-of-the-data-economy.