Very good article. One question I’m curious for your take on: I could see a story where AI actually benefits LMICs if it commoditizes a lot of intellectual and managerial work, making physical work relatively more important.
Do you think that’s a possibility? Or do you think the gains from that shift would mostly accrue to higher income countries with strong industrial capacity and institutions?
Good question, but I think probably not, unfortunately —
- Knowledge work even if fully automated will mean big gains that flow to the frontier labs and their home countries, even if not to 'labor' / wages directly
- Manufacturing as a path to growth was already closing for structural reasons; it's more skill- and capital-intensive than it used to be.
- AI-enabled robotics are likely not too far behind, and the relative value of physical work will fall if robots undercut it; this may also drive re-shoring of manufacturing back to rich countries, which would compound the harm
I am planning a future post more focused on the likely impact to LMIC growth pathways!
Really great read - thanks for writing this! Articulates really clearly some different strains of thinking I’ve been mulling over for the last several months. A couple of reflections/musing aloud comments:
• I read recently that apparently open-source LLMs “catch up” on performance with the leading proprietary models with an average lag of about 3-4 months. If this is true, I would expect the advantage of owning the technology to be at least somewhat dented - people in LMICs having access to even a rough version of an LLM if it’s free/open source is still quite revolutionary
• On the low digitization rates example from BD: I think what makes current AI tools particularly powerful is their ability to use context. We’ve had OCR technology for years, but with the addition of context that LLMs bring, the same OCR ML algorithms can generate much more accurate results because it’s able to draw on context. This is just one example of course, but I feel digitization rates etc might themselves go up because of the greater ease and improved results that AI can deliver.
• Michael Kremer, and researchers at IFPRI & CGIAR have been working on applications of AI to agricultural workers - e.g, ag extension, but also forecasting better, etc. (some examples here: https://www.ifpri.org/project/generative-ai-for-agriculture-gaia/). Just to say that benefits are likely to accrue to non-service sectors as well.
I don’t think any of this detracts from your thesis overall, but on specific supporting arguments, there’s cause to be cautiously optimistic!
I really appreciate your thinking on this topic. Without disagreeing with your analysis, I also though there other channels for AI to improve things for poor countries. Some of the early research showed that AI was generally most useful to the highest performing staff/users and really accelerated their productivity - thereby increasing inequality. But it also brought a lot of poor performers up to median. And that seems like it could be helpful to countries and people at the bottom of the income/productivity range. I read some studies on this via ethan mollick.
I am hoping to write about this soon, but from reading the literature so far it seems to me like the dynamic is highly dependent on context. While, in these high-income country office settings, lower performers benefitted more, it seems like in other settings people with higher education levels or more context on a topic got better outputs from AI, leading to an increase in inequality of outputs. Pairing this with some of the early findings on cognitive offloading and the bifurcation we're starting to see among students in the US depending on whether they use AI as a tool or to do the thinking for them, I think AI helping higher-performers more than lower ones seems really plausible.
One possibility I keep wondering about is whether we may be underestimating AI's role as a teaching technology rather than simply a productivity technology.
The internet made vast amounts of information available, but access to information and successful learning turned out to be very different things. Millions of people can access courses, tutorials, and educational resources online, yet completion rates are often extremely low.
What makes AI interesting is that it introduces interaction into the learning process. Instead of simply delivering information, it can answer questions, adapt explanations, provide feedback, and potentially function more like a tutor than a library.
That obviously doesn't solve infrastructure, institutional, financing, or governance constraints, many of which you discuss persuasively here. But I wonder whether we are still early enough in the development of AI that some of the most important effects may emerge through new forms of human learning and capability formation rather than through direct automation alone.
In other words, could AI eventually become part of the mechanism through which skills and organizational capacity are developed rather than merely a tool that depends upon them?
I'd be curious how you think about that possibility.
I think this is a really good distinction in that interaction is an important element of learning and part of what makes AI different from things like MOOCs (though of course not all the way there - another thing that makes school different from online information is the legitimacy signal, clear credentialing, and commitment or motivational device of other people showing up).
But I don't know that the binding constraint on human capital formation is access to this sort of interactive teaching. For LMICs in particular, I think things like whether the returns to education are high enough to justify staying in school (e.g., there's a functional labor market ready to absorb graduates), or whether households can afford to forgo child labor matter a lot. AI tutoring is a complement to a system where those conditions are met — it makes a functioning education system better. It doesn't obviously fix one where the bottleneck is somewhere else entirely.
Deena, I think that's a useful distinction. Perhaps one of the questions is whether AI remains merely a complement to existing institutions, or whether over time it begins to substitute for some of the functions institutions currently perform—credentialing probably being the hardest one to replace
I largely agree with your observation that many of the most important constraints facing lower-income countries are not informational, but institutional, organizational, and infrastructural.
However, I keep thinking about a different possibility.
The printing press transformed how knowledge was distributed. Public education transformed how societies developed human capital. The internet transformed how information flows across the world.
What if AI ultimately becomes a new form of civilizational infrastructure rather than simply a productivity tool?
Could sufficiently capable AI eventually compensate for some of the functions that have traditionally depended on institutions, especially in places where those institutions are weak or underdeveloped?
In other words, might AI evolve from something that depends on institutions into something that becomes part of the institutional fabric itself?
I would be very interested to hear your thoughts on that possibility.
Thinking about it, the printing press and public education analogies are really apt in that they didn't just distribute existing knowledge but rather actively changed what was possible to know/do. If we think of AI as restructuring how capability is built (e.g., a community health worker operating above their license rather than just getting better and faster at their current role) that difference is important.
That said, the historical examples you cite required institutions to function properly (e.g., public education relied on the state, the printing press only had as much of an impact because of the already-literate populations, and eventually things like intellectual property law). So I don't know that they compensated for weak institutions rather than building on top of them.
Hi Deena -- thanks for putting this together, really thoughtful. You might find the Automation atlas below useful (Baier et. al, 2026). It touches on a lot of points about the specific tasks that are subject to automation generally and AI specifically. One point that I think about is that, even if overall adoption rates are low, the few firms and individuals that do make use of AI tools might see such large productivity gains that then affect society as a whole. Curious your thoughts, and thanks again for putting this out there!
Thanks for sharing this, will take a look at the Automation Atlas.
On your point, I think spillovers from concentrated adoption are real in principle, but the mechanism matters for these purposes. In richer countries, productivity gains at the frontier firm tend to diffuse, but those channels (e.g., competition, labor mobility) are weaker in most LMIC contexts. So I'd want to know more about what the spillover actually runs through before updating much on the concentrated-adopters story.
Additionally the greatest benefit of the ICT revolution to LMICs came from the hyper-globalisation in manufacturing in the 90s and gradual globalisation of services in the 2010s.
Why wouldn't this argument he applied to the ICT revolution in general? LMICs didn't manufacture network switches or telecommunication cables or mobile phones. All of this is still imported.
The largest welfare gains of technology comes from technology adoption and not production.
1- You're right that ownership of hardware wasn't what mattered for ICT, and the consumer surplus literature supports the idea that adoption gains can be large. But:
- AI adoption requires complements that are more demanding than mobile.
- The consumer surplus from + usage of AI is also highly skewed toward HICs and I expect that will continue for a while
- Even if LMICs 'catch up' the lag matters because some benefits are captured in the interim, it ossifies some advantages, etc.
2- It's true that consumer surplus from generative AI already dwarfs producer revenues in rich countries for now, so this is likely true at the moment, but AI companies' valuations seem to be pricing in this changing at some stage.
3- If ICTs caused benefits primarily through labor channels it's not totally clear why this would be replicated with AI. The pathways like service exports enabled by the internet are actually likely to be closed by AI.
What's your reaction to the argument "superintelligence and general-purpose robotics will make everything vastly cheaper, which will lift most developing countries out of poverty"? I.e., a disruption much greater than the industrial revolution, as outlined for instance in AI 2027?
Of course, it seems wildly implausible to most. But many of the people who are actually training the AI models believe it'll be this disruptive. I think we should take them seriously.
Some reasons for pessimism:
- Given the shift from labor to capital (in terms of relative GDP contribution), we should expect developed countries to benefit disproportionately, greatly increasing global inequality
- Many industries that currently generate income could be automated, which could reduce developing-world exports (IT outsourcing made obsolete by coding agents in the near term, agricultural robots reducing food prices, general-purpose robots eating into manufacturing wages)
- Some developing countries are very slow to adopt new technologies, which could cause them to fall further behind
Some reasons for optimism:
- If everything gets several times cheaper, that's equivalent to everyone getting several times wealthier, including most people in extreme poverty, which could mostly offset the above
- Global land prices could increase by a lot due to it being a fixed commodity; many poor people own agricultural land
- An unprecedented AI transformation would create lots of trillionaires and trillion-dollar foundations, and a small number of developed countries with enormous spikes in tax income, which could enable unprecedented levels of development cooperation; as a floor, a small number of trillionaires could scale GiveDirectly to nearly everyone
Overall, my guess is that truly transformative AI could indeed be the next leapfrog miracle, but am open to being convinced otherwise. Curious for your perspective.
Very good article. One question I’m curious for your take on: I could see a story where AI actually benefits LMICs if it commoditizes a lot of intellectual and managerial work, making physical work relatively more important.
Do you think that’s a possibility? Or do you think the gains from that shift would mostly accrue to higher income countries with strong industrial capacity and institutions?
Good question, but I think probably not, unfortunately —
- Knowledge work even if fully automated will mean big gains that flow to the frontier labs and their home countries, even if not to 'labor' / wages directly
- Manufacturing as a path to growth was already closing for structural reasons; it's more skill- and capital-intensive than it used to be.
- AI-enabled robotics are likely not too far behind, and the relative value of physical work will fall if robots undercut it; this may also drive re-shoring of manufacturing back to rich countries, which would compound the harm
I am planning a future post more focused on the likely impact to LMIC growth pathways!
Very informative article, thank you
Really great read - thanks for writing this! Articulates really clearly some different strains of thinking I’ve been mulling over for the last several months. A couple of reflections/musing aloud comments:
• I read recently that apparently open-source LLMs “catch up” on performance with the leading proprietary models with an average lag of about 3-4 months. If this is true, I would expect the advantage of owning the technology to be at least somewhat dented - people in LMICs having access to even a rough version of an LLM if it’s free/open source is still quite revolutionary
• On the low digitization rates example from BD: I think what makes current AI tools particularly powerful is their ability to use context. We’ve had OCR technology for years, but with the addition of context that LLMs bring, the same OCR ML algorithms can generate much more accurate results because it’s able to draw on context. This is just one example of course, but I feel digitization rates etc might themselves go up because of the greater ease and improved results that AI can deliver.
• Michael Kremer, and researchers at IFPRI & CGIAR have been working on applications of AI to agricultural workers - e.g, ag extension, but also forecasting better, etc. (some examples here: https://www.ifpri.org/project/generative-ai-for-agriculture-gaia/). Just to say that benefits are likely to accrue to non-service sectors as well.
I don’t think any of this detracts from your thesis overall, but on specific supporting arguments, there’s cause to be cautiously optimistic!
I really appreciate your thinking on this topic. Without disagreeing with your analysis, I also though there other channels for AI to improve things for poor countries. Some of the early research showed that AI was generally most useful to the highest performing staff/users and really accelerated their productivity - thereby increasing inequality. But it also brought a lot of poor performers up to median. And that seems like it could be helpful to countries and people at the bottom of the income/productivity range. I read some studies on this via ethan mollick.
Thanks, I think this is an interesting question.
I am hoping to write about this soon, but from reading the literature so far it seems to me like the dynamic is highly dependent on context. While, in these high-income country office settings, lower performers benefitted more, it seems like in other settings people with higher education levels or more context on a topic got better outputs from AI, leading to an increase in inequality of outputs. Pairing this with some of the early findings on cognitive offloading and the bifurcation we're starting to see among students in the US depending on whether they use AI as a tool or to do the thinking for them, I think AI helping higher-performers more than lower ones seems really plausible.
Deena, this was a fascinating article.
One possibility I keep wondering about is whether we may be underestimating AI's role as a teaching technology rather than simply a productivity technology.
The internet made vast amounts of information available, but access to information and successful learning turned out to be very different things. Millions of people can access courses, tutorials, and educational resources online, yet completion rates are often extremely low.
What makes AI interesting is that it introduces interaction into the learning process. Instead of simply delivering information, it can answer questions, adapt explanations, provide feedback, and potentially function more like a tutor than a library.
That obviously doesn't solve infrastructure, institutional, financing, or governance constraints, many of which you discuss persuasively here. But I wonder whether we are still early enough in the development of AI that some of the most important effects may emerge through new forms of human learning and capability formation rather than through direct automation alone.
In other words, could AI eventually become part of the mechanism through which skills and organizational capacity are developed rather than merely a tool that depends upon them?
I'd be curious how you think about that possibility.
I think this is a really good distinction in that interaction is an important element of learning and part of what makes AI different from things like MOOCs (though of course not all the way there - another thing that makes school different from online information is the legitimacy signal, clear credentialing, and commitment or motivational device of other people showing up).
But I don't know that the binding constraint on human capital formation is access to this sort of interactive teaching. For LMICs in particular, I think things like whether the returns to education are high enough to justify staying in school (e.g., there's a functional labor market ready to absorb graduates), or whether households can afford to forgo child labor matter a lot. AI tutoring is a complement to a system where those conditions are met — it makes a functioning education system better. It doesn't obviously fix one where the bottleneck is somewhere else entirely.
Deena, I think that's a useful distinction. Perhaps one of the questions is whether AI remains merely a complement to existing institutions, or whether over time it begins to substitute for some of the functions institutions currently perform—credentialing probably being the hardest one to replace
Excellent article.
I largely agree with your observation that many of the most important constraints facing lower-income countries are not informational, but institutional, organizational, and infrastructural.
However, I keep thinking about a different possibility.
The printing press transformed how knowledge was distributed. Public education transformed how societies developed human capital. The internet transformed how information flows across the world.
What if AI ultimately becomes a new form of civilizational infrastructure rather than simply a productivity tool?
Could sufficiently capable AI eventually compensate for some of the functions that have traditionally depended on institutions, especially in places where those institutions are weak or underdeveloped?
In other words, might AI evolve from something that depends on institutions into something that becomes part of the institutional fabric itself?
I would be very interested to hear your thoughts on that possibility.
Thanks, this is a really interesting point.
Thinking about it, the printing press and public education analogies are really apt in that they didn't just distribute existing knowledge but rather actively changed what was possible to know/do. If we think of AI as restructuring how capability is built (e.g., a community health worker operating above their license rather than just getting better and faster at their current role) that difference is important.
That said, the historical examples you cite required institutions to function properly (e.g., public education relied on the state, the printing press only had as much of an impact because of the already-literate populations, and eventually things like intellectual property law). So I don't know that they compensated for weak institutions rather than building on top of them.
Hi Deena -- thanks for putting this together, really thoughtful. You might find the Automation atlas below useful (Baier et. al, 2026). It touches on a lot of points about the specific tasks that are subject to automation generally and AI specifically. One point that I think about is that, even if overall adoption rates are low, the few firms and individuals that do make use of AI tools might see such large productivity gains that then affect society as a whole. Curious your thoughts, and thanks again for putting this out there!
https://automationatlas.org/
Thanks for sharing this, will take a look at the Automation Atlas.
On your point, I think spillovers from concentrated adoption are real in principle, but the mechanism matters for these purposes. In richer countries, productivity gains at the frontier firm tend to diffuse, but those channels (e.g., competition, labor mobility) are weaker in most LMIC contexts. So I'd want to know more about what the spillover actually runs through before updating much on the concentrated-adopters story.
Additionally the greatest benefit of the ICT revolution to LMICs came from the hyper-globalisation in manufacturing in the 90s and gradual globalisation of services in the 2010s.
Why wouldn't this argument he applied to the ICT revolution in general? LMICs didn't manufacture network switches or telecommunication cables or mobile phones. All of this is still imported.
The largest welfare gains of technology comes from technology adoption and not production.
1- You're right that ownership of hardware wasn't what mattered for ICT, and the consumer surplus literature supports the idea that adoption gains can be large. But:
- AI adoption requires complements that are more demanding than mobile.
- The consumer surplus from + usage of AI is also highly skewed toward HICs and I expect that will continue for a while
- Even if LMICs 'catch up' the lag matters because some benefits are captured in the interim, it ossifies some advantages, etc.
2- It's true that consumer surplus from generative AI already dwarfs producer revenues in rich countries for now, so this is likely true at the moment, but AI companies' valuations seem to be pricing in this changing at some stage.
3- If ICTs caused benefits primarily through labor channels it's not totally clear why this would be replicated with AI. The pathways like service exports enabled by the internet are actually likely to be closed by AI.
If AI closes the productivity gap between skilled and less skilled workers, it might incentivise more service exports from LMICs
I think this might be true on some time frames and for some assumptions about capabilities — planning on getting into this topic next post!
Great article! I had just been wondering about this topic myself.
Thank you!
What's your reaction to the argument "superintelligence and general-purpose robotics will make everything vastly cheaper, which will lift most developing countries out of poverty"? I.e., a disruption much greater than the industrial revolution, as outlined for instance in AI 2027?
Of course, it seems wildly implausible to most. But many of the people who are actually training the AI models believe it'll be this disruptive. I think we should take them seriously.
Some reasons for pessimism:
- Given the shift from labor to capital (in terms of relative GDP contribution), we should expect developed countries to benefit disproportionately, greatly increasing global inequality
- Many industries that currently generate income could be automated, which could reduce developing-world exports (IT outsourcing made obsolete by coding agents in the near term, agricultural robots reducing food prices, general-purpose robots eating into manufacturing wages)
- Some developing countries are very slow to adopt new technologies, which could cause them to fall further behind
Some reasons for optimism:
- If everything gets several times cheaper, that's equivalent to everyone getting several times wealthier, including most people in extreme poverty, which could mostly offset the above
- Global land prices could increase by a lot due to it being a fixed commodity; many poor people own agricultural land
- An unprecedented AI transformation would create lots of trillionaires and trillion-dollar foundations, and a small number of developed countries with enormous spikes in tax income, which could enable unprecedented levels of development cooperation; as a floor, a small number of trillionaires could scale GiveDirectly to nearly everyone
Overall, my guess is that truly transformative AI could indeed be the next leapfrog miracle, but am open to being convinced otherwise. Curious for your perspective.