BLOG POST

Cutting Through the Noise: One Way to Make AI Valuable for the World’s Poorest

As the AI revolution booms in the Global North, many in the global development sector are asking how this technology will impact the world’s poorest. Optimists assert that with generative AI, every child will get a personalized tutor. Others caution that the same underlying technology can deliver tailored disinformation and sow discord.

As a General-Purpose Technology (GPT), generative AI has the potential for both. GPTs have, as their name suggests, broad applicability, and include examples such as electricity and the steam engine. They are defined as having wide use, large innovation spillovers, and capacity for continuous improvement. They are subject to the motivations of its users and can be used for ill, the public interest, and everything in between. However, as powerful as GPTs like AI are, they do not teleologically result in advancements for the world’s poorest. Several key barriers must be overcome first. This blog discusses one of these barriers—lack of internet access— and puts forth one way that extremely poor households can benefit from AI without the internet.

Cautionary results from rural electrification

If you are reading this and thinking, “But surely AI is bound to permeate every technological interface and transform every sector, and that has to in some way benefit the poorest”, consider this study from Kenya on another general purpose technology—electricity. 

The randomized impact evaluation found that connected households saw no improvements in economic well being or children’s learning levels at the 16- and even 32-month mark, challenging the notion that electrification alone can benefit low-income households. Even with electricity, families were so financially constrained that they could, for example, only buy one light bulb, and not productive appliances to expand their businesses. Children could study more at night but potentially studied less during the day. If you had asked me to bet before reading the study on whether connecting a rural Kenyan family to electricity would boost their wellbeing, I would have put real money in favor of it. The theoretical benefits of electricity seem so intuitive.

This does not mean that subsidized electrical connections should not be provided, that this result will generalize across contexts, or that benefits won’t eventually manifest. The point is that general purpose technologies, and their expensive infrastructure networks, are still dependent on microeconomic constraints at the level of households and firms. Individual limitations and decision making still matters. With this in mind, what do we have to do to bring the benefits of AI to the world’s poorest?

What generative AI can be useful for

In a previous blog post, I argued that the international development sector should not try to apply AI to every intervention. Instead, the sector should invest where we already have promising causal evidence on what works and supercharge those interventions. Take the field of basic education (getting kids to read and write) for example. We know that an approach called Teaching at the Right Level (TaRL), which is a classroom-based intervention that targets instruction by learning level rather than by age, has been shown across 20 years of studies to be incredibly cost-effective. During Covid-19, when students were not allowed to attend school, Botswana adapted TaRL-based lessons and delivered them over the phone. A randomized controlled trial showed impressive gains in learning.

The intervention was simple—authors of the study write in Nature, “In one treatment arm, SMS messages provided a few basic numeracy ‘problems of the week.’ A second treatment arm supplemented these weekly SMS messages with a live 15–20 min phone call walk-through of the problems. Each student in this arm received 3 hours of direct instruction spread over 8 weeks.” The authors go on to write that the combination of SMS messages and live calls were found to be “among the most cost-effective interventions to improve learning.”

As cost-effective as the intervention is, there are limits on the number of teachers that a Ministry of Education or NGO can mobilize to provide individualized tutoring. Generative AI technologies could remedy this. The underlying technology to deliver a 20-minute interactive phone call that personalizes tutoring and testing for children ages seven and up at massive scale already exists, but has yet to be translated into a tested product for low income countries.

Why is this example significant?

There is growing angst in the “AI for Good” sector that the digital divide will only be exacerbated with AI. The fear is that those in the Global North with access to AI will become exponentially more productive, leaving the world’s poorest who do not have internet, let alone AI tools, that much further behind. This fretting is not unwarranted. While more than 90 percent of people in high-income countries used the internet in 2022, only one-in-four in low-income countries have access. Executives from Google write, “After decades of private and public investments in infrastructure, internet access remains persistently challenging with large shares (e.g. Africa 60 percent, Asia Pacific 36 percent) of the population having either unreliable, unaffordable or no access at all to the internet.” Globally, 2.7 billion people remain offline.

While we may be getting closer to global coverage through companies like Starlink, it is important to remember the study on electricity cited above. The study found that the poorest are incredibly price constrained. Lowering the price of being physically connected to the grid (full price of $398) by 57 percent increased demand by less than 25 percentage points. Monthly electricity expenditure was between $1.80-$2.17– just enough to power basic lighting or a fan each day of the month.

I suspect that even after all the internet infrastructure has been laid and internet satellites have been launched, finding even a subsidized price for internet connection and productive use that can be borne by the more than 600 million people living below $2.15 will be challenging.

This is why the study on Teaching at the Right Level delivered over the phone and 2G networks was particularly exciting. Paying for 2G talk time via an old school feature phone is something households living under $2.15 a day are already doing. In the Kenyan electrification study cited above, 5 percent of control households had electricity connection, while 85 percent had a cellphone. Cellphones and their use are affordable. Feature phones in Kenya can cost as low as $8, and Safaricom charges less than $1 for three hours of talk time. Leveraging what households already have and can afford just makes sense.

A company called Viamo is already starting to apply generative AI to 2G voice networks. Their product “Ask Viamo Anything” is being tested on feature phones without internet access. It has the potential to transform their previous offering, known as Interactive Voice Response (IVR) that had services like “Press 1 to learn more about what to expect when pregnant, Press 2 for when to visit your nearest health facility…” into a more intuitive experience for callers. They could enable conversational, targeted coaching to new parents, much like an AI tutor can offer targeted instruction on basic literacy at scale. Sending voice messages to mothers prior to the advent of AI already had promising results. The mMitra Program in India sent pregnant women and new mothers voice messages on nutrition, prenatal care and breastfeeding practices, and demonstrated impacts on maternal knowledge and some behavioral practices.

What needs to be done

While I am advocating for AI technologies to be run over existing technology that the extreme poor can already afford and use, there are still challenges to overcome. Translating generative AI to be applicable to voice, especially in languages and dialects that have no written form can be challenging. That said, OpenAI recently publicized VoiceEngine, which is able to use text input and a single 15-second audio sample to generate natural sounding speech. In partnership with Dimagi, VoiceEngine could even read voice messages in Sheng, a creole popular in Kenya that has relatively less standardized written conventions. 

I suspect that making generative AI available over 2G phone networks and conversational in multiple local dialects will form a foundational technology layer applicable to several sectors. While there has been exciting early work I also suspect that without strong philanthropic backing, this application of AI technologies to 2G systems will be underinvested in relative to internet-based applications which are easier to monetize and have a more robust developer ecosystem. This does not mean we should not be investing in bringing internet access to the world’s poorest. I am simply arguing that we should move faster, deliver what we can through networks that already exist, and not wait for the $100B investment that is required for Africa to reach universal broadband coverage.

Concurrent to funding these innovations to deliver AI over voice networks, philanthropy should fund rigorous impact evaluations of their effectiveness. In the example of tutoring delivered over the phone, we may learn that part of what drove students to engage with remote tutoring was the accountability between the teacher and student. Generative AI may be able to deliver targeted instruction, but that may be less effective without the sense of accountability that comes from human interaction. Rolling out new technology will expose unknown unknowns, and those working to bring them to market should operate like a startup, developing minimum viable products and testing rapidly.

As groundbreaking as AI will be for the world, those working on “Technology for Development” have gone down similar roads. The promise of a new technology is often quickly followed by vague proclamations about the need to provide access to the world’s poorest. Where possible, we should deliver AI services through technologies already adopted and financially sustained by extremely poor households. 

Disclaimer

CGD blog posts reflect the views of the authors, drawing on prior research and experience in their areas of expertise. CGD is a nonpartisan, independent organization and does not take institutional positions.


Image credit for social media/web: Ken Banks / Flickr CC BY 2.0 DEED