How AI could make markets worse

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

Greetings Free Lunch readers. Like everybody else, I’m trying to figure out how to think about the effects artificial intelligence will have on our societies, especially the economy.

It seems like most of what I read is about the technology’s ability to do (better and faster) tasks currently done by humans and therefore boost productivity. I see less about how AI may change the organisation and workings of markets. But productivity also depends on market functioning — not just on things being done better or faster or more cheaply, but the different things being done fitting together in an efficient way. So today’s column offers some thought about how AI could affect market functioning.

AI is, above all, an information technology: it reduces the cost and friction of acquiring, processing and producing — sometimes very impressively. So we should look to the economics of information to get insights about its effects. If AI leads to significant productivity increases, it’s by making information cheaper. But information economics is a weird field, with strange results. Here are some.

The classic case of a surprising insight from taking information imperfections seriously was George Akerlof’s “market for lemons” — a term for defective used cars, not the yellow fruit. Akerlof’s model showed how imperfect knowledge could prevent mutually beneficial trade. If a used car seller knows whether the car on offer is of high or low quality but the buyer does not, the buyer will only be willing to pay a heavily discounted price (to account for the risk of buying a lemon). But the owner of a good car (a “peach”) will find that price too low, and only owners of lemons will sell. Understanding this, buyers will adjust down their price further. The outcome: only lemons get sold, at a fair price, but the market for high-quality used cars breaks down.

Akerlof’s market for lemons showed how costly it can be when information is less than perfect. But it also illustrates a situation where AI should presumably improve efficiency. Armed with their favourite large language model, perhaps buyers could inspect your used car to establish its true condition to their satisfaction, and hence offer an acceptable price for a peach as well as for a lemon. No more would lack of knowledge prevent mutually beneficial exchanges.

But there are other cases of information economics where AI may not make things better and could, in fact, make things worse. Take, for example, a standard solution to when asymmetric information prevents mutually beneficial trades: signalling. In short, when uncertainty makes one side of a transaction unwilling to trade — Akerlof’s used car buyer, say, or an employer unsure of the quality of a potential hire — the other side can find “signals” that distinguish them from lower-quality alternatives.

In some cases this can be free (the seller of a “peach” can offer a guarantee against nasty post-sale surprises); in others it will require expending resources (future workers taking university degrees to prove their mettle). The key is that the signal’s cost is worth paying for the high-quality alternative but not the low-quality one (the seller of a lemon wouldn’t offer a guarantee; the less competent would find it harder to complete the course).

But the information-producing abilities that AI brings could disrupt the availability of such signals. Take education. The hand-wringing over how students are using LLMs at university is well known, and the main question is, of course, what it does to their learning. But we could ask, too, what it does to the signalling function of formal education. If LLMs in effect equalise every student’s measured performance in university, or, to go further, if they equalise their chances of getting into selective universities in the first place, then it’s hard to see degrees retaining their effectiveness as a signal of ability.

Would that make the economy more or less efficient? If it makes it costlier to match up the right jobs with the right workers, it presumably makes labour markets and the overall economy less efficient. Perhaps some of that would be compensated for by young people not spending years of their lives to invest in a signal rather than actual skills, if your view is that degrees are largely about signalling rather than learning. Or it could be that in the AI-infused world we find better signals to compensate for those that no longer carry the information they used to. But taken by itself, the destruction of a signal should be expected to amount to a productivity loss.

This is related to another area of interest to information economics: that of costly search. In a process of matching sides — take the job market or dating — it takes time and effort to assess how suitable a potential match is. So those looking for a job, a worker, a date or a mate will have to choose how much time and effort to put into assessing “candidates” and when to settle and stop searching. Now an important insight from search theory (a field that was awarded a Nobel Prize) is that one person’s search can affect the costs and benefits of the search process of others. In particular, if others search more, you may have to do the same to be able to count on an equally good expected outcome as before — a negative externality, in the jargon.

Cue AI-assisted job applications, recruitment processes and dating profiles. A likely outcome is that using AI makes rational sense for each individual, but requires more effort to be put in by everybody else — whether to sift through requests, filter out the slop, or to retain a chance of being noticed when others are swamping the field with applications. And, of course, the aforementioned problem of less valuable signals — if everyone can write a perfect job application or dating profile, for example — makes this worse.

We could think of many other examples — do send us yours ([email protected]). But they are likely to have in common that while AI makes information cheaper, “too much information” can come with its own efficiency cost. There is family resemblance to the contemporary approach to propaganda, of “flooding the zone” — it matters less to convince someone of something than to put out so many more or less plausible views that people give up trying to seek out the truth.

Perhaps AI will also help find solutions to these problems. But the key lesson as we assess what AI will bring is this: being able to execute tasks more productively does not necessarily make market interactions more productive overall.

Other readables

● Both on the battlefield and in the economy, things are turning against Russia and in favour of Ukraine. This is the right time to increase economic pressure on Moscow, I argue in my latest FT column.

● Some young Chinese find sanctuary in the country’s post-property boom ghost cities.

● Mind in the machine? The philosophers working for the AI creators.