Why AI is different for Poor Countries

Thanks to web-based applications and cloud computing, you can use the same AI tools in the United States and West Africa. But even if these countries use the same models for the same purposes, AI will have a different impact on the two economies. The different impacts will have very different political effects.

Thanks to Anthropic’s Economic Futures report, we can get some insight into how AI use differs across countries. In the U.S., based off 220,000 observations, here are the top five uses of AI (reworded according to my interpretation):

1. Cheating in school

2. Coding

3. Cheating in school

4. Comparison shopping

5. Business consulting

In Benin, a small country in West Africa I’m currently visiting, the top five uses across 536 observations are:

1. Coding

2. Business consulting

3. Cheating in school

4. Coding

5. Business consulting

Since the sample sizes are different by over two orders of magnitude, we can’t assign too much significance to the differences. But some interesting patterns arise.

Americans like using AI for cheating in school. This has less to do with virtuous students in Benin as it does with their limited access to the internet.

Benin also does not use AI for comparison shopping. That’s not a huge priority in a country where GDP per capita is about $2500.

But the common tasks across both countries is coding and business consulting. It’s interesting that distinct economies and cultures are converging on similar uses of AI. But it would be hasty to conclude that since their uses are similar, the impact will also be similar. Even when uses are the same, the type of innovation is very different.

Clayton Christensen made his career by thinking about innovation. He unfortunately died an early death in 2020, meaning he missed the explosion of AI growth that has come in the last five years. Undoubtedly, if he was alive today, we’d be peppering him for insights on AI.

I thought of Christensen’s work when reflecting about how different AI is for poor countries and rich countries. His framework groups innovations into three categories: efficiency innovations, sustaining innovations, and market-creating innovations. In this framework, AI doesn’t fit into one spot. It fits into two, depending on which country you are in.

Efficiency innovations are the ones that give us the same thing but cheaper. Think of shipping containers, a relatively simple but radical efficiency innovation. Before containers, cargo was not standardized and therefore loading a ship was difficult. Dockworkers carried cargo on and off, but each ship was different. Moving cargo took a lot of time and labor, and since each ship required unique care, logistics were a mess. Delays occurred regularly and no one knew whether they would get a Turboman in time for Christmas.

But this changed with the container. Cargo came in a standardized, stackable container. Docks installed cranes to easily load and unload them, and they could transfer them immediately to a train or a truck. Since every ship was the same, processing times dropped dramatically and logistics became more predictable.

This efficiency innovation gave us international trade, but cheaper. The cost was a dramatic drop in the most expensive part of shipping: human labor. Dockworkers, so common they became a trope informant in detective stories, weren’t needed once the job could be effectively automated.

This is where most people put AI. It’s threatening coding jobs by making software development cheaper. It’s threatening business consulting jobs by making business analytics cheaper. It’s even threatening plagiarizing jobs by making cheating cheaper. It’s the efficiency innovation aspect that has people worried about the future of work.

But this is the discussion in rich countries. It’s a different thing in poor countries.

Christensen loved the market-creating innovation. These innovations don’t focus on delivering the same good or service cheaper. They target people who aren’t consuming it at all and make it accessible. In targeting nonconsumption, they usually provide a product that is inferior to the market leader. But the product isn’t supposed to compete with the market leader. It’s competing with nonconsumption.

A go-to example is Ford’s Model T. We often think about Ford’s innovations in assembly lines, which were efficiency innovations. But we don’t talk about how the Model T was an inferior car. It didn’t ride as nice as a Cadillac, and it certainly didn’t look as nice inside. But Ford wasn’t trying to convert Cadillac drivers. He wanted the person who didn’t have a car to start driving, and to do that he had to offer a “just good enough” version at a price they could afford.

Christensen’s claim to fame is pointing out the irony of this kind of innovation. He called it the Innovator’s Dilemma. Since the innovation targets nonconsumption, the market leaders ignore it. The new guy is attacking the least-profitable part of the market. The incumbent says good riddance; it doesn’t like being in that market anyway. So the incumbent ignores them, and the new guy starts making money. Then as the innovation improves over time, the new guy moves up market, taking the next least-profitable sector. He works up the ladder until suddenly he’s the reigning champion.

In rich countries, AI as an efficiency innovation is targeting people who already use the service. But in poor countries, there’s a lot of nonconsumption. For example, the United States has an 80% college education enrollment rate. Across Sub-Saharan Africa, it’s less than 10%. That means Africa has access to significantly less college-educated intelligence to fill jobs. AI rapidly expands access to this intelligence at a low cost. Yes, it sometimes hallucinates or doesn’t do exactly what you want, but even if it’s only good 90% of the time (which is lower than I would estimate), that still may be better than having no access to intelligence.

So AI is a different innovation in poor countries. It has the potential to create markets and target nonconsumption. But I think this has another implication for the spread of AI.

Because AI is an efficiency innovation in rich countries, it’s facing resistance. The same thing happened with shipping containers. Dockworker unions knew they would lose jobs, so they dug their heels in. They slowed the adoption of containers and captured some of the gains. With AI, we’re seeing similar discussions. The workers aren’t organized in unions, so maybe it will ultimately be ineffective. Nate Silver is probably right that we’re underestimating the political effects of AI adoption.

But in poor countries, there could be less resistance. AI is competing with nonconsumption. It’s not displacing workers, because those workers don’t live in developing countries. So there isn’t a class of people who will push against it. The political cost of widespread AI adoption could be much lower in poor countries.

Of course, I’m considering only one political cost of AI adoption, so I could be underestimating which barriers are most significant. Tech in the US has deep pockets and growing influence, so it may be able to overcome political barriers. On the other hand, many governments in poor countries are already compromised by rent seekers. If they feel threatened by the potential for AI to unseat them, they could impede adoption. But I think the general thrust is right: there will be less popular resistance to AI in poor countries.

I’ve become convinced that there’s a compelling case for targeting AI adoption in poor countries. First, it fits Christensen’s market-creating innovation, expanding access to services that were previously unattainable. Second, there is less likely to be a labor market disruption that creates political backlash.

But to advance AI adoption in poor countries, we need to be more intentional. Here’s an example. Mark Zuckerberg told Dwarkesh that Meta’s models spoke good French, but the personality felt like an American who learned French. The French users could tell they weren’t talking to something that reflected their values. How does this translate to users in developing countries?

Would an African entrepreneur receive advice consistent with the laws and regulations in his country, or will the AI assume he’s in a country like the US and France and give improper advice?

If I ask it advice on bribes that I need to pay to make sure my business deals make it through whatever loops have been erected, will it understand that that’s how business works in these countries or will it refuse to work if there are bribes?

Are models going to be available at prices that poor countries can afford? What does that business model look like? Cross-subsidization from rich countries to poor? An older vintage model that is less costly to run but still useful?

For economic and political reasons, AI has a different opportunity in poor countries. But we need business plans and policies that are focused on the potential there.