I just finished teaching my course on AI futures (Our AI Future; DSPC IA7175, formerly AI Institutions) for the second year at Columbia University’s School of International and Public Affairs (SIPA), for Master’s in Public Administration students.

Here’s the course description:

AI is rewriting the rules of society. This course invites you to understand and shape what comes next. We begin by turning the classroom into a living experiment on how AI could change education, then examine how abundant intelligence could reshape work, governance, and transportation. In a field often dominated by speculation, we will ground our discussions in evidence and theory. Together, we’ll explore what institutions are needed for a world transformed by intelligence.

The economic and social implications of AI will not be figured out solely by computer scientists in the labs. They will be figured out across society, by the entrepreneurs and tinkerers who see a bit ahead. I don’t have the answers, but my aim is to equip students to see further ahead, by helping them reason about what is coming. We read technical papers (scaling laws and on AI capabilities), emerging evidence on AI impacts, evidence from historical precedents (steam engines and electrification), and conceptual pieces.

The syllabus is here. (If you’re considering taking my course at SIPA, there are spoilers ahead.)

Principles

The course is designed on a few principles:

• AI allows, and may require, that we redesign institutions. Most discourse considers how AI may be grafted on to existing jobs, industries, laws, and government agencies. This class considers how these systems might be reshaped.

• We need both data and imagination. We combine guided speculation from theory and conceptual pieces, with evidence based on hard data. There’s a mismatch between the type of rigor prized in most academic work, which leads to carefully measured estimates that are out of date, and the forward looking guidance that is needed to understand our changing world. I used the class as a vehicle to think about this gap, and published some of my thinking as an essay in Nature.

• Education is a laboratory for the impacts of AI. I begin the course by acknowledging that that AI has expanded students’ capabilities, so what they are asked to deliver must also expand. Their first assignment is to propose a set of assignments for the course, to achieve learning and impact. In our second class session, we settle on a set of assignments. Deciding these assignments is always contentious: some want the traditional memo assignments to which they have become accustomed, many are wary of technical assignments, and many are wary of signing up for a class with uncertain deliverables. It’s chaos. That chaos is by design. At the end of the debate, I remind students that we will be considering a series of domains that have a status quo that may be disrupted by AI. In each of those domains, people will find their world shaken—we must respect how it feels.

• AI progress raises standards. Submissions are graded relative to what AI could produce. Last year, students chose to write memos on the readings. Halfway through the semester, I confronted the class with the fact that AI capabilities had increased. We graded two example memos together (incidentally generated by AI), and realized that student memos no longer exceeded in quality those that could be written by AI. That is an economic problem: an employer may not pay them $100/hr to produce insight that a computer could do for $0.01. The main change this year was to produce demos of minimum viable products.

Reflections

Students found a few moments particularly provocative:

• Is the potato a general purpose technology? Nathan Nunn and Nancy Qian’s work finds that its introduction from the Americas accounted for one quarter of population growth in the Old World between 1700 and 1900. The potato is a technology that reshaped society. Does it matter whether a technology is a GPT?

• Should self-driving cars respect local values? Nigerian respondents to the Moral Machine are more willing to spare higher status individuals, Chinese respondents are less willing to spare the young, and Americans are more willing than others to sacrifice human lives to spare animals. Who should decide these tradeoffs?

• Today’s AI policies will look quaint. Early airplanes did not have good navigation systems, so the U.S. built a network of giant concrete arrows that pilots could follow by looking down. Innovations to navigation made this stopgap obsolete. Our adaptations to AI will also go through multiple phases, and problems that appear severe today may find solutions we have yet to imagine. What are the stopgaps, and what are the supporting innovations?

• Visualizing the scale of modern AI systems. Visually relating the simple quantitative models students have used in statistics classes to the structure of LLMs is striking.

• We were grateful for a guest visit by Séb Krier of Google DeepMind, with a rich Q&A about AGI timelines and the role of policy.

I have now taught the class over two consecutive years. A few things have changed:

• Students are no longer skeptics. Last year, about half of my students were optimistic about AI, and half were skeptics. This year, nearly all of my students believe AI will transform society. They want to navigate that transformation.

• Students are capable of much more. Last year, two students built compelling demos for their final projects. This year, all students—even those with zero experience coding—built creative and interesting software projects. It’s a new world.

• Most economic and social changes are still yet to come. I kept much of the same material on economic principles, as it remained relevant. We are likely to be surprised by changes that have yet to come.

What students built

My students built creative demos! Here are a few:

• Policy conflict radar: anticipate political issues before they arise, based on how AI personas react to current events, by Sewon Sunwoo.

• Pathwise: assess whether your career path is exposed to AI, and get suggestions on how to pivot, by Karina Lages de Altavila.

• Resolve: write up the details of a dispute, then learn your rights and how to navigate the legal system, by Hani Iqbal and Ali Serdar

• SafeScroll: decide what content you see on the web using nuanced controls, for example blurring show spoilers, by Navya Sinha.

It has been an honor to think about our future with my students. I’m excited to see what they build next.

This piece was originally posted on Daniel Björkegren’s Blog.

Many low- and middle-income countries are preparing for AI to make information abundant. However, the frontier has shifted. Rich economies are preparing for AI to make some forms of intelligence abundant. Will advanced artificial intelligence allow LMICs to leapfrog human knowledge work—or leave them behind?

Information versus intelligence

Basic AI systems, like large language models circa 2024, are good at retrieving existing knowledge. In LMICs, they are being used for health, agriculture, business (Otis et al. 2025), and education (Björkegren et al. 2025). Many countries are pursuing frugal strategies that deliver these services with small models over mobile phones. For communities that the web barely reached, even basic AI offers an improvement in access to information.

However, frontier AI models are increasingly able to not just retrieve information, but apply forms of intelligence: planning, using tools, and completing complex tasks that require multiple steps. Coding assistants like Claude Code or OpenAI’s Codex are already shaking up how software is developed, how data is analyzed, and how computer scientists are educated. AI capabilities are improving and are expected to affect other knowledge sectors: the past few months have seen the release of tools that can perform general computer tasks and create digital designs. Since advanced models require much more computational capacity, firms are building vast data centers and the energy generation to power them.

Why returns to intelligence differ

The economic implications of this transformation can be characterized by the marginal returns to intelligence (Amodei 2024): how much can we improve economic outcomes as we better generate ideas, process data, and apply knowledge? Intelligence allows us to solve scientific problems, design better products, better anticipate demand, and ensure the right quantities are stocked in the right places. Low-income countries will benefit from innovations developed in rich ones. But within many LMICs, the complements to advanced AI are scarcer, including data centers, reliable electricity, and digital records, as well as experienced knowledge workers. Data centers can be located in countries that already have good infrastructure (‘the cloud’) and accessed remotely. But LMICs are less digitally legible: AI will be less able to understand and act in markets, firms, homes, clinics, and schools that do not record data in structured forms. Overall, we would expect LMICs to be at a disadvantage in integrating advanced AI (Korinek and Stiglitz 2021).

A crucial distinction is that LMICs have much smaller knowledge sectors. LMICs employ fewer than 10% of workers in skilled knowledge work, like managers, technicians, and professionals, relative to 41% in high income countries (Silva 2026). Current AI tools require substantial human guidance. So, firms in rich economies are pursuing a grafting strategy: existing knowledge workers are being asked to integrate AI into their roles, starting from producing slides and emails, and scaling to more sophisticated tasks. In countries with smaller knowledge sectors, there are fewer workers and processes to graft AI onto. Thus a key question is whether advanced AI will mainly empower existing workers, or automate knowledge work completely. In wealthy countries, advocates concerned about jobs suggest that AI systems be designed to augment rather than automate (Acemoglu, Autor, Johnson 2026). But in low-income countries, the more urgent question may be how to provide knowledge services when few knowledge workers are available. Fully automating knowledge work could in fact augment less educated workers, who could ask AI to complete macro tasks like developing marketing strategies, rather than micro tasks like reformatting spreadsheets. However, even automated systems will likely require oversight from entrepreneurs and scientists with deep expertise, which may be sufficiently available only in wealthier countries like Brazil and India.

If AI allows LMICs to grow automated knowledge sectors, would the returns be high or low? One indicator is in wages paid to human workers. The wage returns to college education are slightly higher in lower income countries (Psacharopoulos and Patrinos 2018 and 2025), but educated people often earn higher wages abroad, and some domestic knowledge workers are working on rich countries’ knowledge problems in call centers and business process outsourcing. Lower income economies may not currently be structured to fully tap the decision making entailed in knowledge work (Engbom et al. 2025). If we tasked millions of data scientists with helping smallholder farms, the returns are unlikely to be large: agriculture is constrained elsewhere.

However, if the price of some forms of intelligence declines by orders of magnitude, it may become worth applying intelligence to problems that were never worth assigning a human to. Small manufacturers might generate nuanced designs that would have required a team of industrial engineers, and implement advertising campaigns that would have required large creative teams. Many regions have struggled to agglomerate sufficient human talent; since automated intelligence can be accessed anywhere, it could make businesses more mobile. These opportunities could more fundamentally change economic structure.

Small knowledge sectors may also pose less resistance to the adoption of AI. Wealthy country workers are beginning to politically resist AI. But residents of low- and middle-income countries are less concerned that AI will displace jobs, and are more optimistic about AI, according to interviews conducted by Anthropic (Huang et al. 2026). LMICs also have less developed laws and ecosystems of institutions designed to manage human knowledge work. As a result, some uses of AI may become blocked in wealthier countries but allowed in low-income countries: for example, a proposed chatbot law in New York state would restrict AI systems from providing advice typically provided by licensed humans. Ultimately, the constraints in LMICs represent not simply deficits, but a different optimization landscape.

Building for abundant intelligence

The economies that LMICs trade with, compete with, and depend on are being reorganized around abundant intelligence. The question is not whether LMICs will be affected—they will be—but how they will engage. The analysis above suggests that AI that augments knowledge workers is likely to disproportionately benefit rich countries. It is less clear what alternative paths exist, but there are actions that could offer opportunities, particularly for middle-income countries and more advanced regions within low-income countries.

The most capable AI systems currently require large-scale frontier models and large amounts of compute. Governments, firms, and NGOs will need to work with the frontier labs to ensure that the most advanced models speak local languages and understand local contexts. Ensuring that there are multiple suppliers for both models and data centers can reduce prices and risks of lock-in and geopolitical disruption (Athey and Scott Morton 2025).

Governments will also need to push to make economic activity digitally legible, from markets to clinics to schools.

It is also important to ensure that AI can be productively used. That may require training humans to be more productive users of AI, both in applying the tool and having the deeper world knowledge needed to direct it. Firms can also invest in developing AI tools that are complementary to the industrial structure of LMICs, including tools for small scale entrepreneurs who have less education, and for agriculture, like weather forecasting.

The diversity of institutional conditions in low- and middle-income countries may be a comparative advantage. Wealthy countries have evolved similar institutions around human knowledge work; tweaks may lead to local optima. In contrast, systems in low-income countries can differ greatly. Tailoring to different constraints can generate opportunities: for example, Kenyan entrepreneurs coping with unreliable network connections developed techniques to create on-device AI models that are seeing demand around the world (Fastagger). Or, also in Kenya, 90% of people resolve disputes outside the formal justice system (Kenya 2020), and just two doctors serve every 10,000 people, compared to 37 in the United States (WHO, 2022). Firms and NGOs may find creative new solutions, such as offering more efficient ways to settle disputes outside of court, or dynamic medical advice. Governments can take advantage of opportunities to design new regulation for AI, rather than retrofit regulation designed for humans. A lack of established institutions around human knowledge work could also allow harm: what happens when medical AI makes mistakes and there are limited mechanisms to address malpractice? It will take care to develop appropriate new institutions.

We must begin to consider what economies with scarce resources might look like amidst abundant intelligence. The questions are challenging. If LMICs can’t graft AI onto knowledge work, what are they building instead? Will they wait for intelligence to reshape the global economy? Or will they develop new—and possibly different—economic structures?

This piece has benefited from feedback from Jeff Berens, Han Sheng Chia, Johannes Haushofer, and Alex Imas. It was originally posted on Daniel Björkegren’s Blog.

Frontier AI labs have begun publishing remarkably detailed reports on how their products are used. Reports from Anthropic, OpenAI, and Microsoft have documented not only overall usage, but also insights on how AI might interact with labor markets, by linking requests to job-related tasks, and differentiating between requests for help and for automating tasks. These statistics help policymakers grapple with one of the most important questions facing countries that have many knowledge workers: how AI will affect employment.

However, employment is not the most important question for billions of the lowest income people around the world, who for the most part are not knowledge workers and predominantly live in low- and middle-income countries. Interviews conducted by Anthropic found that low- and middle-income countries’ residents were much less concerned that AI would displace jobs, and were much more optimistic about AI. For the world’s poorest residents, AI may improve access to advice on health, business, education, and countless other needs.

The labs have made strides in reaching these communities, from developing language benchmarks to country-specific reports to flood forecasting for billions. But poor countries still use AI far less than rich countries. While this gap appeared to be shrinking for OpenAI’s ChatGPT from 2024-25, it was expanding in Microsoft’s more recent data. Understanding impacts on the poorest would both help target these companies’ efforts and help societies that are wary of downsides of AI assess the degree to which its benefits will accrue beyond Silicon Valley and the Fortune 500.

Because the world’s poorest have different needs, labs should release tailored statistics to understand them. I describe here simple, feasible analyses that would help the AI industry, governments, and social sector organizations better serve the poor. I also predict what we might find.

1. Are the world’s poorest using AI?

Anthropic, Microsoft, and OpenAI document that in low- and middle-income countries, a smaller share of the population uses AI than in high-income settings. However, some patterns appear puzzling. On average, users of AI in lower-adoption countries are more likely to use Anthropic’s Claude for programming, and countries that are less educated on average are submitting requests that AI more successfully completes. These counterintuitive patterns likely result because early adopters in low-income countries are wealthier people.

Wealthy people in low-income countries live like residents of rich countries: they may be knowledge workers and speak international languages like English, Chinese, or French. However, the poorest people in low-income countries have lives that are very different from anyone in wealthy countries: they may have income reliant on risky agricultural harvests with no safety net, have limited access to clean water, see medical providers who give incorrect information and counterfeit drugs, and send their children to packed classrooms where they learn little.

Because of the large difference between the wealthy and poor, country-level averages do not tell us whether AI is providing opportunities for the poor. There are early signs that some of the poor are using AI: my team analyzed usage of a custom AI chatbot, finding that a group of teachers in Sierra Leone used AI more often than web search, mostly to look up facts. The labs have data on many more users, and thus a correspondingly larger opportunity to illuminate the extent to which AI is reaching people in need.

There is not a single definition of poverty, but there are several indicators in the data the labs collect that could both identify the poor and assess barriers that limit their access to AI.

Devices

Lower-income people more commonly access the internet from smartphones, not computers, and typically use cheaper handsets. Many of the poorest have only basic phones that cannot access the internet. Low-literacy people may prefer interacting with voice rather than text.

What could be computed: What proportion of users have inexpensive versus expensive handsets? How does their usage compare? What proportion of consumer use is via smartphone apps or voice/SMS (possibly by third parties using the labs’ APIs)?

What we might find: In low-income countries, a large share of access is from wealthier people, using computers or higher priced smartphones. People with lower priced handsets have lower usage. Some custom apps tailored to local needs may be seeing substantial uptake, such as health or education apps by governments or NGOs. One large question is whether the poorest are better served by third-party solutions; if so, that may suggest that partnerships may be more important than improving the consumer app experience.

Language

Many of the world’s poorest residents primarily speak local languages like Tamil, Kannada, or Yoruba, even in countries where the wealthy speak major global languages. In India, if 95 percent of Claude requests are in English, that would indicate a very different kind of diffusion than if 40 percent of requests are in local languages.

What could be computed: What proportion of queries are written in each language, in each country? How does that compare to the breakdown of first languages in that country? Additionally, what are the failure or abandonment rates by language?

What we might find: A substantial fraction of AI usage is likely to be in major global languages, even in countries where the population predominantly speaks local languages. That may indicate a combination of lack of access and lower model capabilities in local languages. However, we are likely to find exceptions: broad take up in some languages, which can illuminate what is working and where additional language data need be collected. These measures may be stronger evidence about how AI performs in diverse languages than benchmarks, because they reflect whether models meet real world needs.

Remoteness

Cities contain a mix of rich and poor people, but in many countries rural areas are poorer. Additionally, people in rural areas may have less access to the internet and be less connected.

What could be computed: What proportion of use is rural versus urban, by country.

What we might find: There is substantial use in cities around the world. Some countries likely have substantial rural use, but for some it is almost entirely absent. That could indicate access barriers, that the models are less useful for the problems faced by rural residents, or lack of awareness.

2. What do the poor use AI for?

After the first step, we are likely to find a group of the poorest who are in fact successfully using AI. There is limited market research on these populations, and many do not often use web search, so their needs are absent from web query data. AI usage can provide a new window into the needs of the poor, analogous to Google Trends. This can help AI labs and a variety of organizations better serve these populations. We saw early examples of this among teachers in Sierra Leone, who submitted requests for not only facts and lesson plans, but also on handling reports for insurance claims and navigating interpersonal situations with students and supervisors. Another study found that one of the top uses of ChatGPT among gig workers in India and Brazil was for health queries.

What could be computed: An easy start would be to take the standard categorization of requests already reported by the labs (such as writing, technical help, or mapping to industries) and report them specifically for the subset of users in marginalized groups (defined by having cheaper devices, speaking local languages, or using from remote areas). However, these taxonomies are built around knowledge work, and may systematically undercount the ways poor people find the technology useful. Thus, it would be helpful to develop new taxonomies to understand poverty-specific needs, including particular uses within agriculture, health, navigating government bureaucracy, and business advice.

What we might find: The poor are likely to use AI differently from the wealthy: almost no software development, some use for navigating bureaucracy and social problems, more for help with homework, and less for writing assistance. Anthropic has already reported that Claude users in lower-income countries are more likely to request help with coursework. A further breakdown will help us understand if AI is being used only in the wealthiest schools or broadly, and help school systems ensure it is used in ways that support learning.

The governments and social sectors of low-income countries are all deciding how to respond to the spread and potential of AI. At the same time, AI labs are being asked whether their products will enrich a small few—or widely benefit humanity. These simple analyses are feasible. They would transform our understanding of whether AI is reaching people who may benefit most.

This piece is cross-posted at the Center for Global Development and Daniel Björkegren’s Blog. It has benefited from feedback from Jeff Berens, Han Sheng Chia, and Markus Goldstein.

While the internet has changed many things, the past year’s developments in AI suggests that accessing knowledge may have simply been a prelude. Once this knowledge has been ingested into machine learning models, it can be automatically synthesized.

What impact might AI have on developing economies, and on the world’s poor?

If AI is as important a breakthrough as mobile phones for these countries, then it could increase efficiencies across existing industries (Jensen 2007), alter political mobilization (Manacorda and Tesei 2020), and improve welfare by a decent share of GDP (Björkegren 2022). But if, as some are considering, it is as important as electricity or the printing press, then it could completely alter the structure of industry, politics, and production.

Developing countries have already seen widespread adoption of some early applications of machine learning. In addition to general applications that are common across societies (e.g., search engines, algorithmic ranking on social media), there are specific applications like digital credit scoring for people who are unbanked. Other applications are emerging, like digitally targeting aid during crises, or analyzing satellite images for policy. But those traditional ML systems are much more constrained. What we are facing is a new animal.

What can AI do and how might it affect these economies?

These technologies are evolving quickly, and their adoption across societies is likely to lead to complex dynamics. But here are some initial observations of what could be on the horizon.

Cheap, tailored expertise

Many of the poor have little access to experts, on topics ranging from physical and mental health, agriculture, and entrepreneurship. Hiring experts can be transformative: Bloom et al (2013) found that management consultants substantially improved production at textile factories in India. Many envisioned that the internet would level access to various forms of expertise, allowing entrepreneurs in Delhi and farmers in Western Kenya equal access to the world’s knowledge. But much of the poor’s knowledge is not written, and much of the world’s knowledge is not written for the poor. As anyone who owns a dusty textbook knows, raw information is not enough. Much of written knowledge is in English, some uses technical jargon, and has metaphors and references that make sense to people in Los Angeles but not in Lagos. AI has the potential to better synthesize local and global knowledge, and thus to unlock insights. The newest generation of AI chatbots can not only translate between languages, but also change reading levels, change jargon, and rewrite to use local customs and metaphors. These chatbots also allow you to have a conversation about a topic, allowing you to ask for clarification for specific parts, or specify that your needs differ from what the system had assumed. While these systems sometimes make mistakes, their quality, and ability to translate is improving quickly. And they will get better as local knowledge is increasingly digitized. They may allow almost zero cost access to the tailored expertise that would otherwise require hiring experts that would be prohibitively costly for the poor. This tailored expertise might improve business processes across developing economies for motivated people. Some startups are already developing targeted advisors for specific tasks, like choosing between schools (ConsiliumBots). Similar advisors could also help deliver medical advice to rural populations.

AI may alter pathways of development

AI has the potential to alter which industries lead to growth, and comparative advantage.

For one, there is a set of ‘high value’ industries that were seen as pathways to widespread good jobs which may no longer be, if automation replaces some outsourcing (Korinek and Stiglitz 2021). For example, software development is one job that is increasingly being transformed by AI: AI can write code to meet a specification, correct bugs, and describe code in everyday language. These impacts are likely to spread to other desk jobs, and may alter the returns to some jobs enabled by the internet (Hjort and Poulson 2019). Tasks for which a human can assess correctness of the end result are likely to be more affected.

At the same time, AI can provide some advantages to developing economies. AI models tend to have worse knowledge about lower income consumers, which will initially inhibit applications targeted towards the needs of the poor. However, these models have good knowledge about wealthy foreign consumers. This could help lower income countries to develop products for export. An entrepreneur in Lagos can query an AI to learn about the needs of consumers in Los Angeles: cheap market research. Additionally, strong grasp of ‘correct’ English may not be as important for engaging with the global labor market. This may diminish some of the returns to learning English, relative to other major languages. These systems can polish text to make it professional, which evens the advantage across clear thinkers who speak different languages.

A tutor for every child?

AI’s ability to reshape knowledge can also improve the quality of education. While there has been a dramatic increase in school attendance in developing countries over the past generation, quality has been a struggle, and many classrooms rely on rote learning. Having a good teacher can have a profound impact on learning and lifetime earnings (Chetty et al. 2011), but good teachers are hard to find. In developing countries, equipping relatively untrained teachers with standardized lesson plans and school procedures can lead to dramatic gains (one study saw students gain 3-3.5 years of learning relative to two years in status quo schools, in a randomized control trial; though that implementer has had ethical breaches).

AI tutors have the potential to improve on standardization by personalizing learning. Many education technologies have underdelivered, but well implemented ones have already produced large gains in learning, such as one in India (Muralidharan, Singh, and Ganimian 2019). Given the right setting, personalized AI tutors could result in far more learning, could be engaging, and iteratively learn the pedagogy that leads to the highest gain. They have the potential to give every child access to a tutor with good pedagogy. That could revolutionize the next generation. It’s unlikely that these modes can be simply set into existing education systems, so one large question is how to develop the physical and institutional environment that will best empower learning. Current systems can act as good tutors, and may be sufficient for advanced and motivated students. But younger ages will need engagement layers and supervision. Coming AI systems will generate video and interactive experiences, which could engage students (in ways both educational and not). Tutoring systems can also help adults: because they can tailor knowledge, they make it easier for a person trained in one field to learn a new field: AI chatbots can tailor the explanations to the mental model they already have.

Simultaneously, the needs of education systems may shift as economic activity shifts. Such changes have occurred in the past: as cognitive skills became more remunerative than memorization, parents invested more in children’s critical thinking (Hermo, Päällysaho, Seim, and Shapiro 2022). This is one reason that IQ has been rising over the past generations (the Flynn effect). These new innovations may also shift the returns to different skills. It’s unclear what skills will be most needed. For example, with current language models, figuring out tricks in how to phrase your request (“prompt engineering”) can help you get much more useful information, for example asking how a particular smart person would answer a question (Cowen and Tabarrok 2023). But this is likely to be more a symptom of these models not yet meeting our needs; as models improve, this skill may become less necessary. However, these systems are likely to produce some output that is incorrect, a defect which is likely to endure in some form. So, an ability to critique output and think critically will likely continue to be valuable.

While many of the first integrations of AI are into systems that knowledge workers use (e.g., create slides from revenue numbers), in the medium term this could change organization structure. These systems may enable a few decisionmakers to directly surface the information needed to make the decision. This may make it possible to produce more output of higher quality in smaller ventures. In developing countries, relying on AI rather than humans may help get around some of the contracting frictions that inhibit the growth of midsize ventures.

A challenge: AI knows less about poor consumers

Many of the poor in developing countries do not regularly access the internet, and are not well represented in the data used to train models. In particular, in Africa very few people have full computers, and smartphone adoption is still low among the poor (Milusheva, Björkegren, Viotti 2021). This has two implications. First, it will be difficult to access AI tools until smartphones are adopted. (Though, if AI tools are sufficiently valued, they could drive adoption.) Second, these tools will be less tailored to the needs of the poor.

Initially, AI tools will tend to make lower quality decisions for the poor. That will arise from two factors. First, the poor produce much less public content. Fewer smartphones mean fewer photos, and less writing about the topics important to the poor. This problem is more severe for minority languages. Additionally, much of the digital content that low income consumers create is posted to closed networks (like Facebook or WhatsApp groups) rather than public, annotated archives (like GitHub, Stack Overflow, or Wikipedia), which are the most useful for training machine learning models. Second, in the process of training an AI system, users provide feedback about whether a response is acceptable. So far, most of that training is being done by or for wealthy Western consumers, and as a result they will produce answers tailored to their preferences. People in other societies have different preferences (Falk et al. 2018), so some of these recommendations will fall short. For example, AI models tailored to the U.S. may recommend greeting consumers by first name in emails or encouraging customers to leave tips, techniques that may backfire in other settings. Yet, even untailored AI may be valuable for some tasks. And these weaknesses can improve. In particular, the presence of AI will increase the social returns to digitizing the knowledge of the poor.

What next

Over the past 20 years, developing societies have built out mobile phone networks and internet connections. This infrastructure already allows the fluid movement of information. It may increasingly also provide forms of cognition. This transition is likely to be disruptive—and could have a substantial impact on the world’s poor.

This post has benefited from feedback from a few people, who I thank without implicating: Jeff Berens, Mike Björkegren, Tyler Cowen, Avi Goldfarb, David McKenzie, Chris Neilson, and Paul Novosad. It was originally posted on Daniel Björkegren’s Blog. The cover image was generated by AI (Midjourney, with the prompt, ‘a manager of a high tech factory in Lagos looks over the factory floor holding a smartphone, in a vibrant cartoon style’).