I was a development economist before I pivoted into AI safety, which has led people to mistakenly believe I have insight into how AI will affect developing countries. I do not. But I have takes! They are worth very little. Here are a few.
1. Service-led growth is dead
Most discussion I’ve seen on AI and development focuses almost exclusively on how AI will affect “service-led growth” – i.e. economic growth driven by exports of technologically-intensive services to rich countries, like call centers and back-office legal work. I think that kind of service industry is dead, for obvious reasons. But I also think that particular type of service industry was an overrated path to growth in the first place. People indexed on it pretty heavily because of the success of India and to a lesser extent the Philippines. But that doesn’t mean other countries could have followed the same path.
There is a different flavor of “service-led growth” that Dani Rodrik is trying to get the development community behind – growth through producer services that are targeted domestically (e.g. transportation, retail). By being grounded in the physical world, this strategy is more robust to AI than BPO exports. But it has a different problem: producer services are strong complements to manufacturing, so you can’t scale them without manufacturing. You can’t have a transportation industry without goods to transport; you can’t have a retail industry without goods to sell. So I just don’t see a world where domestically-targeted producer services can become a full development strategy.
2. AI-embodying capital will become the new engine of growth
Rafael Latham-Proenca has an excellent model for how developing countries can grow in a world transformed by AI. His post is long and criminally underread, but I’ll summarize the idea. After transformative AI, productive knowledge gets embedded in machines, so catching up mostly means getting the machines: AI-embodying capital like self-driving tractors, crop-spraying drones and warehouse robots. Thus, the key policy agenda that Rafa identifies is to reduce barriers to adopting AI-embodying capital. The most important barrier is financial: importing AI-embodying capital from rich countries costs money. So developing countries will have to either borrow money globally or raise money from their citizens to fund this investment. This parallels the “financial repression” that was essential to East Asian growth, where governments forced people to supply savings while rewarding them with a very low interest rate, so that businesses could invest that money and pay it off with economic growth.
I think this is the most optimistic and interesting growth strategy. It actually overcomes a really thorny problem with growth today. Governments can invest in people to boost labor – e.g. education to make people more skilled – but these investments are slow and take a long time to pay off, making it difficult for them to power a growth strategy. But capital can be accumulated quickly – it could even be as simple as shipping in machines from another country. Although it stings a bit that you don’t own the machines to begin with – and that means you have to pay for them – it is at least possible to mobilize them quickly.
3. Redistribution will become an existential policy question
Historically, development happened through labor-intensive industries (e.g. textiles). One benefit of this was that we didn’t need to do intensive redistribution, because labor was already earning most of the new income being generated. Redistribution came “for free”. But if capital becomes the new source of growth, governments will have to take pretty strong steps to redistribute that income before people can actually enjoy the benefits of growth. This is already a conversation in rich countries – the US debates about universal basic income or job programs or other ways to direct some of the surplus created by AI towards labor – but it will also become the most important conversation in developing countries too. I’m pretty confident that the US will redistribute adequately, but I’m not so confident about developing countries.
Relatedly, I’m worried about the intelligence curse: when AI makes human labor unnecessary for growth, governments lose their reason to invest in people. This works from the view that people’s bargaining power comes from the state needing them to do labor. For example, a 15th century autocrat couldn’t oppress the peasants too hard, because there would be nobody left to farm the fields. But if a time traveller gave him autonomous farms that didn’t need peasants, he might be a bit more willing to toss all the peasants in jail. A less extreme version is that there are all kinds of ways that governments redistribute towards people today (e.g. education, social services) that would not be necessary if all work was done by machines.
In rich countries, democratic norms are strongly embedded, enough that I’m not worried about governments becoming repressive or neglectful in response to AI. But in developing countries, and especially on the poorer end, I’m concerned that AI will enable a total impoverishment of the citizenry by governments that no longer need them.
4. In the short term, AI might trigger a debt crisis in LMICs
There’s a lot of demand for thinking about the long-term effect of AI in developing countries, but there’s a looming crisis on the horizon that people are really not awake to.
AI investment is likely to raise real interest rates. This will force developing countries to borrow at higher interest rates too, since they have to compete with lucrative AI investments. This means they will be forced to spend more and more money servicing government debts. Countries currently on the edge of being unable to service those debts may have to default on their debts, triggering a crisis.
This is a rough explanation, because I personally struggle to understand international macroeconomics. In this way, I am sadly typical. Most development economists don’t really understand international macroeconomics and sovereign debt, preferring to focus on non-monetary topics like trade or industrial policy, which gives them a blind spot for the financial channels through which global shocks hit poor countries first. The Venn diagram of (thinks about development), (thinks about sovereign debt), and (thinks about AI) is as close to empty as I can imagine. (I say that in the hope of being proven wrong, Cunningham’s Law-style.)
5. In the long term, most of AI’s effect will be through derived products
It’s worth remembering that AI will be valuable because it can help make stuff. It could make cures for diseases, or new energy sources, or new internet services. These derived products are not necessarily going to be identified with AI, but they will be most of the effect of AI.
For a short while, I contemplated starting a company that would use AI to accelerate agricultural R&D in developing countries. I was not equipped to do that, but the motivation was strong: agricultural R&D is currently concentrated in the US and focused on US crops and biomes, and research on the crops and conditions of poor countries is usually too expensive to be worth doing. If AI brings down the cost of breeding and testing new varieties, that research could become worthwhile for the first time. The resulting seed varieties could unlock agricultural progress and generate huge economic benefits in poor countries. This is the kind of derived product that will provide most of the value from AI. It sits alongside cures for diseases, new energy sources, new software – things that we could discover with AI, but are not themselves tied to AI.
These derived products may not always look like AI, and they may have unexpected effects. The whole purpose of AI is that it can invent general technologies, so we can’t say all that much about how general technologies will affect developing countries. But we do know that technologies tend to diffuse throughout the world. The internet, smartphones, computers, and electricity were all invented in rich countries and yet have found their way globally. So we should not be thinking only about things that we would label as “AI” when thinking about AI’s impacts.
6. AI will be jaggedly adopted; this will be very weird
In 2000, only 0.5% of Indians had access to the internet. That 0.5% had access to IRC chat rooms, internet culture, Those who did had access to chat rooms, online forums, foreign music and news, and a window into a global culture their neighbors couldn’t see.. Of course, this led to economic differences, but more importantly, it led to cultural and political differences. These 0.5% just lived fundamentally different lives from the rest of their compatriots.
This kind of within-country inequality – not just economic inequality, but inequality in access to the future – is likely to become widespread. Whether we see AI companions, or AI teachers, or other AI-induced cultural and political changes – these are likely to be extremely jagged within countries. Large chunks of country populations are going to live lives that are essentially similar to what they live today, while a small and growing contingent will live in sci-fi futures. Subsistence farmers will live in the same city as bounty hunters who track rogue AIs on the internet. This is going to be very weird!

