The Cell Phone, AI, and the Future

What does technology mean for global equality?

The cellphone has had an enormous impact on life, even in the poorest corners of the world. Just about half of people in sub-Saharan Africa own a cell phone now, with enormous consequences for productivity and welfare. Just about ten percent of Africa’s GDP (I shall henceforth use Africa to refer to the sub-Saharan regions) is due to cellphones – half from the direct utility to the consumer of being able to talk with friends and family, and other half from knowing more information. A fisherman can now know what the price of fish is in different villages, or a laborer knows they will be able to find a room and a job if they move to the big city.

Information is valuable in its own right. George Stigler’s two papers, one from 1961, the other from 1962, on information are what really started our modern understanding of information as a good. People do not possess information effortlessly, and must expend energy and money to find it. Thus, there is a dispersion of prices away from what the market clearing price would be, and some degree of static inefficiency.¹

We can see this empirically. One of the cooler papers I know, Steinwender 2014, looks at the opening of the transatlantic telegraph line in 1866 from America to Britain. The big use case of the line was giving the prices of commodities in Britain to American traders, in particular cotton. She found that variability in prices plummeted after the line was introduced, with large and positive effects on welfare. Simply knowing whether to ship or not ship was equivalent to removing a 6% tariff on cotton, or an 8% increase in the total value of American cotton exports, and this was in spite of cotton being a storable good. (As an aside, I thought her method of calculating this was really cool. In order to estimate deadweight loss, you need to know the slope of the supply and demand curves. These are simultaneously determined, and you can’t simply observe the data. What she uses to identify the curves is that once a ship leaves America to go to Liverpool, it won’t turn back, no matter what the price is in the destination port. From this, she can then estimate how valuable the information is). Jensen (2007) measures the effect of cellphones directly, by looking at the rollout of cellphones in Kerala, India. Kerala has a substantial fishing industry, and is even more sensitive than cotton to gluts and deficits. By knowing where their products were most in demand, fishermen were able to cut down on price dispersion and on waste.

In Africa, Hjort and Poulsen (2019) looked at the effect of high speed internet on employment. The identification here is similar to Steinwender. Internet access comes from the opening of undersea cables, which happens discontinuously. They find that having access to the internet substantially increases employment, without causing displacement, and that incomes rise. The internet and mobile phones have been an especially big deal for small businesses – in 2010, Aker and Mbiti report that firms had almost universally adopted cell phones in Kenya, while individual use lagged behind. Another study from Aker on cell phones, this time in Niger, found that it reduced dispersion in grain prices by at least 10%. Niger had basically no landline telephones, only having 2 for every 1,000 people, so the mobile phone was a true revolution in information availability.

But why could they adopt the cellphone, but not landlines? The cellphone has been adopted because of how little it requires from the state. To get people the telephone requires a competent, or at least non-confiscatory, state. One must run wires to every house, and also have home electricity. A lot of inventions are like this. The automobile means little if you cannot build roads. Electrification is pointless if you cannot build large power plants and install wire across the country.

By contrast, cell towers require nowhere near the same level of infrastructure investment. A foreign company need only slap up a few towers and put a fence around them. The cell towers themselves are not particularly valuable to steal, and they do not need to be connected to a grid, being runnable off of standalone power sources. (During the initial rollout, they were mainly powered by diesel generators, but with the falling costs of solar they’re now almost entirely powered by solar panels and batteries).

And we’ve seen this sort of thing before, where innovations are particularly valuable only in a particular institutional context, although it does seem rarer for it to benefit the less-developed parts of the world. The hand-cranked transistor radio is a curio in the developed world, but a wonder in unelectrified parts of the world, for example. (Please don’t look too closely at the effect of radio access on the Rwandan genocide). The effect of institutions on growth is therefore not constant over time. Even if what a good institution is is constant (a rather big assumption) we cannot expect them to have the same differential impact on growth rates in different technological contexts. It isn’t possible to take a point estimate from studies estimating the effect of institutions on growth, and expect it to stay the same in a different context.

This matters looking backwards too. It seems likely that, even if institutions were a necessary factor in the industrial revolution, they were nowhere near sufficient. You couldn’t have had an industrial revolution during the time of the Romans because technology – and, perhaps, the people – had not attained the necessary level for sustained economic growth. It’s really humbling to read Joel Mokyr’s book, “The Lever of Riches”, a book which on account of its detail does not abide summarization. Living standards – if narrowly defined to encompass only the barest necessities – may well have remained stagnant, but technology did not. Europe on the dawn of the 18th century was as alien from when it was at the time of Christ as we are from them.

So, will AI tend to have a bigger impact in developing countries or developed countries? There are reasons to think it might go either way. Nevertheless, I think it will be much more important in the developed world than in the developing world.

AI is really two different things, large language models and application-specific machine learning. AI in the sense of large language models (generative AI) becomes more useful the more non-physical work is rewarded. They are complementary to skill. Yes, they are going to produce cheap but mediocre work. Within countries, they are likely going to benefit the middling more than the very advanced. But between countries, a lot of what it does doesn’t matter in the poorest countries in the world. AI might make it easier to deal with regulatory compliance, which will reduce the demand for lawyers – but will that matter in a country which lacks the state capacity to regulate at all? AI might allow less doctors and nurses to make more accurate diagnoses – but what does a more accurate diagnosis matter, if you lacked access to treatment anyway? What is the point of becoming a much more productive computer programmer if there is no return to computer programming at all in your country? The returns to skill are dependent on the skills that other people have, and a technology can increase equality within a country while decreasing it between countries.

And I do want to emphasize that the evidence which we have points toward generative AI compressing skill in the developed world. Think about software engineering – much of what a software engineer has been done a thousand times over. It is boilerplate, and so AI is incredibly good at it. Really it’s just splicing together a corpus of stack overflow posts and splicing them together; but it does it so fast and well. Experiments confirm this. Cui, Demirer, Jaffe, Musolff, Peng, and Salz (2024) conduct three randomized controlled trials at major companies, in a continuation of prior work which by many of the same authors on freelancers. They found that using CoPilot substantially increased productivity, with heterogenous effects by skill level – exactly as predicted, less skilled engineers saw more of a benefit than the most skilled.

In writing boilerplate in an experimental context, Noy and Zhang (2023) find that it substantially increases productivity and reduces inequality in production. ChatGPT benefits those who find writing hard or unenjoyable the most. (Hard luck for your dear author here!) And in customer service, Brynjolfsson, Li, and Raymond find the same compression in wages accompanying an improvement in average productivity. AI serves to disseminate the standard way. And Merali (2024) found that low-skilled translators saw much greater gains from access to AI than high-skilled translators. It is possible that some of this is because more skilled workers are prideful, and do not adapt to new technology uses — there is a study of radiologists which shows that the doctors simply didn’t trust AI enough. This is belied by the next section, though.

Conversely, I believe that AI in the sense of application specific deep-learning – things like self-driving cars and predicting how proteins will fold – will benefit both the top of developed countries, and developed countries generally, far more than the poor. Take improvements in self-driving cars. I think this is a big deal, and actually quite underrated by the public – it will not only greatly reduce traffic deaths, it will also allow trucks to stack up on top of each other to avoid drag, combining the advantages of the railroad with the advantage of trains – but all of this is dependent upon good roads, good infrastructure, and the ability to afford quite expensive vehicles.

Firms are much smaller in the developing world. Investing into technology entails a fixed cost, and if the firm isn’t big enough, doing so is unprofitable. This is why international trade can improve productive efficiency in a world of increasing returns. See Melitz and Treifler for an exploration of how bigger markets allow us to have more varieties and more specific production processes. There’s a really cool paper out just a couple months ago on slack in economic development which has a related thesis. Machinery is discontinuous – you cannot purchase half of a tractor, or half a lathe. If your business isn’t very large, there are periods where you are going to be simply unable to use the inputs you bought. As firms get bigger, they’re able to use their machinery all of the time, which in turns allows them to use more optimal production processes. Investments can be made which would not otherwise be. Simple survey evidence backs this up – large firms were much more likely to implement some sort of automation in the year sampled. (Yes, yes, small firms might in systematically different industries. Nobody is going to seriously argue that bigger firms aren’t more likely to undertake capital expenditures). In short, any capital intensive investments will be done first in the developed world.

Recent work on the impact of AI on science shows that it is most beneficial to those at the top, and not very beneficial to those at the bottom. Aidan Toner-Rodgers has a working paper out on the adoption of AI in an anonymized materials science company. The AI could identify promising target molecules, and the scientist would evaluate whether the claims were accurate. Using AI greatly increased the rate of success, but it did so by increasing the top performers by a lot, and changing the worst performers not at all. This fits with the framework of Agrawal, Gans, and Goldfarb (2018), who see the recent advances in AI as improving prediction, but not judgement. They define “predictions” as anything where similar situations have occurred before, and AI is able to make predictions from large datasets, and “judgement” when a novel situation arises. (It is, of course, wrong to think that AI is totally unable to make judgements in novel situations, and will never be able to. Humans are not fundamentally unreproducible by machines. One should start out by analyzing what would happen if things continued progressing at much the same rate and way as they have before, before seeing what would happen if you were to change around some of the assumptions). At the very cutting edge, where people have to create new ideas, AI is going to benefit the best more, if it benefits them at all. It is in the humdrum – the legal filings, rudimentary statistical analysis, drawing up a powerpoint – that it will compress productivity.

Artificial intelligence is not the only technological change coming, and not all of them will benefit the developed world more than the developing world. The use case of blockchain technologies is primarily in avoiding extortionary governments, and to the extent that that is more common in developing countries, it will benefit the poor of the world more. Medical advances are likely going to disproportionately benefit the developing world too, as it seems that the low-hanging fruit is primarily in vaccines for malaria and HIV. Artificial intelligence is the biggest of them, though, and I expect the first order effects to widen inequality, even as it makes everyone better off. The second order effects – of what sort of innovation will occur in the new world – is so uncertain that rendering any prediction here would be an exercise in false advertising.

1. This is too much a detour for even me to insert in the main text, so read this after. This dispersion in prices is why inflation is harmful under conventional macro models. These use something called Calvo pricing, where in each unit of time, randomly chosen firm change their prices. (That is, they’re visited by the Calvo fairy). Since all changes in price level are not going to be fully reflected by firms changing their prices, the optimal monetary policy is one where the central bank targets an inflation rate of 0. Unfortunately, the empirical evidence has not been kind to this assumption. Nakamura, Steinsson, Sun, and Villar (2018) test whether the size of price changes has increased during a time of greater changes in the price level. If firms are limited in the number of times they change their price, then during periods of higher inflation firms must have bigger price changes. They do not find this. When dispersion gets worse, firms simply change their prices more often. Optimal policy is likely stabilizing nominal wages, rather than targeting inflation, which you arrive at if you say that changing prices incurs a fixed cost which is invariant to size. See Caratelli and Halperin (2023) for a characterization of this.

I’m not so sure that this is optimal, though, since it seems like firms should be able to economize on changing prices if they have expectations of the future course of prices. Very little of so-called “menu costs” is the literal changing of menus, or other markers of price. Prices are way too sticky for it. (See Eife 2009, which uses changes in currency when the Euro came out to estimate menu costs. The literal signage would all have to reprinted when countries shifted to the Euro, but people wouldn’t have to really put any thought into what the price should be. The literal cost of menus was insufficient to explain empirical observations). Suppose that inflation is 1% every month (terribly high, I know), and that everyone knows this. At the beginning of the year, the firm decides that every month prices will be raised by 1%, and if something weird happens that necessitates some deviation, then they’ll reconsider. Fixed menu costs would say that each and every month incurs the same cost, but that just clearly isn’t true! Menu costs have to be endogenous to the regularity of monetary policy. Unfortunately, I do not have the ability to critically evaluate what happens when we change this assumption, but if any of my enterprising and every so talented readers would like to take it up with me, I’d really quite like to work on a paper with you. :)