What do you think is missing from the debate about AI and job loss?
The macroeconomists who are thinking about this question are focused on comparing two rates. Let’s say that eventually goods essentially get to be free, because AI makes more energy and makes more robots and goods become very cheap to make in terms of the actual cost of the resources used. Meanwhile, since the machines are making everything, we don’t need people anymore, so wages go down. So you’ve got two things happening at the same time: wages going down and goods getting cheaper. It’s possible that in that world, consumers are fine because they don’t need any income. But built into that is an assumption that people can access the goods at the real cost of making them, which is trending towards zero.
We need to be making sure that AI is not in the end captured by a small set of digital platforms that have market power, like we’ve seen with other kinds of digital services.
The point of the piece I wrote with Susan Athey is, what if the AI is owned by somebody who has market power and is a monopolist? Then people’s wages are trending towards zero, but the goods are still expensive. That’s a world that’s really terrible for workers. We don’t need a sophisticated model to understand that. It’s going to be bad.
That leads us to the moral of the story, which is we need to be paying close attention to competition in AI and making sure that AI is not in the end captured by a small set of digital platforms that have market power, like we’ve seen with other kinds of digital services.
A related problem is, does the party with market power in AI have different objectives than we do? Is the market power held by a hostile state—or just a hostile person, a tech bro who’s decided that he knows better than the rest of us how we should be living our lives and is going to execute on that vision without regard to what society wants?
How would competition in AI prevent that scenario?
If three or four of us have an AI tool that can grow carrots and there’s somebody who owns land that can grow carrots, they’re going to pick the least expensive tool that lets them grow the carrots. And competition is going to drive the cost of farming the carrots lower and lower and lower.
In that scenario where there’s multiple entities with an AI tool that can do farming, then landowners who want to be farmers all have the option of using those different tools and the creators of the AI tools are going to have to compete on the basis of quality and price and innovation to be the ones used to grow the carrots.
If one AI company was the only one out there, then somebody who owned land could choose to farm carrots with the old technology—tractors and people—or they could use these robots and data that are run by AI, but whoever’s selling them that is going to charge for it and is going to charge according to what the market will bear.
Now, that actually brings up an interesting issue. A monopolist in a general-purpose technology can raise prices not just for carrots, but computers and energy and clothes and trains and planes and everything because AI will be used in literally every good and service. That means that there’s what we call an income effect: everything gets expensive. At the same time, remember, the consumer is losing income because their labor is not needed for any of these industries.
Now, ironically, that makes it harder for the monopoly AI to charge a monopoly price. If you charge a high monopoly price and no consumer has a job, then no consumer can afford to buy your stuff. So in a hypothetical world where we suddenly move from today to monopolist on everything and consumers have no money, it doesn’t really do much good to be the monopolist. But of course in real life, we’re going to transition over many years and there’s going to be a lot of time for whoever has market power in AI to extract resources from the people who still have them.
You’re writing about the scenario in which AI has a transformative effect on the economy. Is there anything that is comparable from past human history that we can look at?
There are really big changes like the invention of electricity, the invention of the steam engine—similarly, these are general-purpose technologies that are applied across a huge swathe of the economy and really change the returns to factors of production. Once you’re doing manufacturing, the returns to land go down; once you’re doing certain kinds of services or manufacturing, you need some kinds of labor for them, but not other kinds of labor.
Some people think that AI will be different because it can invent new ideas. I don’t have a strong view on that. I don’t know if it’s right or not. But if so, it would make AI even more transformative than these past examples.
What can we do about this?
Engage in serious competition enforcement in the area of AI right now, and do it pretty vigorously so that we don’t end up in a bad spot in 5 or 10 years.
It’s hard, of course, because this is a quickly moving technology. It’s not clear who’s competing with whom and what the product market is, both because it requires some technological expertise to understand and because it keeps changing. But I think the enforcement is necessary nonetheless.
The judge just handed down a decision in the Meta Platforms case concluding that maybe at the time it engaged in its purchases of Instagram and WhatsApp, Facebook had market power in social media and was monopolizing. But today, 15 years later, according to Judge Boasberg, TikTok has come along and therefore Meta Platforms no longer is a monopolist, so the government isn’t allowed to prevail in its antitrust case.
That tells you that government enforcers can’t wait around because courts will say it’s too late. In the U.S. a typical antitrust case takes at least five years to litigate, on top of a couple of years for the government to develop it in the first place. You’ve got a seven- or eight-year lag to consider. That means a government has to bring the case immediately if it sees any kind of problem in order to have a hope of getting an answer from the court when the problem is still ongoing.
Are you talking about regulation or antitrust enforcement or both?
We don’t have regulation of digital platforms in the United States. We can only use the antitrust laws.
What do you think a remedy would be?
Well, you don’t know the remedy until you know the problem. Right now, we have a number of AI companies; we have a number of parties making chips; we have a number of parties making foundation models; we have a number of parties making applications. If many parties are competing for the same set of customers, then that market will typically be vibrant and innovative and competitive.
But we have pretty large parties that control distribution, like Google, and we have pretty large parties that control chips, like Nvidia. The European Commission has opened an investigation to figure out whether Google is using its distribution channel monopoly in search to try to monopolize consumer-facing AI. So that would be one potential problem.
Let’s imagine you investigate that and you find a competition problem. It should be fairly straightforward to say you can’t tie your search engine to AI—you have to make some kind of open interface that lets consumers choose among other AI services—and you can’t self-preference and pay end users or distribution channels to take your AI the way Google did in search.
There are a lot of parts of the technology stack behind AI with all different market structures. There are all sorts of imperfections in the way people can switch and competition can work and interoperability can work. So I would be shocked if there weren’t competition problems that arise over time.
You mentioned at the beginning that with a lot of our existing technologies, we have gone to very few players. Do you think that AI lends itself to being limited to one or two dominant companies?
We have concentrated markets in areas like search; social platforms like Instagram and Facebook; office productivity software, where Microsoft has a very strong position; perhaps something like e-commerce, where Amazon has a very strong position. Mobile operating systems are perhaps the best example, where we have Android and iOS and that’s all.
So why do those markets become concentrated? I wrote a long report about this in 2019 for the Stigler Center at the University of Chicago. And basically the story is you have network effects: the more users there are, the better the quality of the platform. And you have data effects: the more data you get, the more you can learn and make your product better, which makes it more attractive, which attracts more people, which attracts more data, which makes it better. When you have this kind of network effect, economies of scale with very low marginal costs and big fixed costs, and data effects, then you tend to get concentration.
In the AI world, some of the elements of that stack have these features, but I don’t know enough to tell you that all of them do. I don’t know whether, for example, foundation models are inherently going to get us to a monopoly or whether the market can sustain five or six or seven foundation models over time. The costs of the different parts of the stack keep changing, and the nature of the network effects could move over time also.
Concentration also depends on differentiation. If I want a foundation model that’s really good at biology and scientific problems, is it really one foundation model for all these different areas. Or is there going to be some return to taking a foundation model and making it really good at a subset of tasks?
But the more AI gets built into every area of our economic lives, the more potential harm from a monopolist in the stack.
“The Yale School of Management is the graduate business school of Yale University, a private research university in New Haven, Connecticut.”
Please visit the firm link to site