“The robots work at night”: How AI helps the world’s poorest, and where it falls short

GiveDirectly is built on the core belief that people living in poverty know what they need better than anyone else. That belief is part of what led our team to test an unrestricted AI chatbot – with funding from Imara Fund, a foundation co-founded by Reed Hastings and Constance Jones – alongside cash transfers in Rwanda. This represents a different approach from the restricted AI tools many other organizations are offering.

The two of us had seen how much can hinge not only on money, but on access to the right information at the right time – Sharon from years of working directly with Rwandan entrepreneurs, and Carolina from designing programs for vulnerable communities around the world. 

We know that cash gives people room to make their own choices. AI can be one way to help them make more informed ones, especially in places where reliable coaching, advice, and services are harder to reach.

AI is filling gaps that cash alone cannot

Alongside our usual one-time ~$1,000 transfers, we offered recipients in Rwanda a ChatGPT-powered chatbot on WhatsApp, developed with Turn.io, so they could ask questions in a tool many already know how to use.

Our team had three main uses in mind: helping people plan how to use their cash, think through business ideas, and get answers about the GiveDirectly program. And people do often use it that way. They ask how eligibility works, how to make spending decisions, and how to manage the practical details of new investments– from how to get fruit to a more distant market to caring for livestock.

But the more revealing pattern was how quickly recipients moved beyond those program-specific questions. They used the chatbot for the ordinary, messy decisions people bring to AI everywhere:

• How do I manage conflict with my wife?
• What should I do if my child is sick?
• Where can I find markets outside my community?
• Who can I trust in my neighborhood?

Recipients using the chatbot in this way isn’t particularly surprising or unprecedented. They use it the way people use AI everywhere: to think through money, work, family, health, and trust. But in many of the places we work, people often have fewer places to turn for a second opinion, practical advice, or timely information. The community health facility, business coach, or legal aid office that might otherwise have the answer may be far away, expensive, or entirely unavailable in their community. The chatbot suddenly put tons of it at their fingertips.

“The robots work at night”

Traditional training and support programs are often delivered during the day, at fixed times, and in group settings. But many people do not have the luxury of structuring their lives around when support systems are open, especially in places like rural Rwanda, where many people spend their days tending farmland or working a market stall.

A chatbot, on the other hand, is available whenever they are. It’s there after the children are asleep, during a quiet moment at work, or at the end of a long day when there is finally time to think. As one recipient put it in a focus group discussion, “The robots work at night.”

In many cases, that timing is part of the value. The chatbot creates space for people to reflect privately, experiment with ideas, and ask questions they may not otherwise have an opportunity to ask.  

People trust the chatbot because it is not human

People everywhere bring sensitive questions to chatbots, often because they want privacy, patience, or a lower-stakes place to think out loud. Recent research from Anthropic found that millions of people already turn to Claude for personal guidance on health, careers, relationships, and finances – often because they cannot easily access or afford professional advice.

In close-knit communities in rural Rwanda, that choice can be especially delicate. People may be reserved about personal matters, and many are cautious about sharing information that could expose them to scams, pressure, or criticism. In that context, the chatbot’s non-human nature can be part of what makes it feel safer. It offers privacy, patience, and freedom from social judgment.

Because of that, many people treat the chatbot like a sounding board. They greet it, confide in it, and return to ongoing conversations over time. 

The next frontier for AI inclusion may be language, voice, and localized context

Many of the barriers to AI uptake in lower-income countries are structural. Patchy internet, expensive data, and low-quality phones all shape how people use these tools, and the broader infrastructure work to address them has been underway in Rwanda for years. Chatbots are also a new interface for most recipients, so there's a learning curve.

Other barriers come from gaps in the data that the models have been trained on:

• Language support is still limited. Language determines whether many users can meaningfully engage at all, and tools perform inconsistently in African languages like Kinyarwanda. A handful of small vendors are starting to close this gap. GiveDirectly is among the first users of Proto, which is doing this work in Rwanda.
• Voice functionality is underdeveloped. For many lower-literacy users, voice notes are their primary interface, but voice features are often unreliable, slow, or poorly adapted to these contexts.
• AI knows the cities better than the villages. Models mostly know about capital cities like Kigali; the more rural the setting, the less the AI knows about local markets, services, and conditions. That asymmetry shapes how useful these tools actually are for the people we serve.

Closing these gaps will take deliberate investment from AI developers and partnership with the organizations and communities who understand what 'useful' means in practice. We'll keep testing, learning, and sharing what we find – both to make our own programs better and to help the developers building these tools see what works in these contexts.

We're continuing to learn how AI can amplify the effects of cash

As we continue testing what makes cash more impactful, AI is becoming an increasingly important part of that conversation. This pilot suggests that AI has huge potential to expand what people living in extreme poverty can do with the resources they already have.

"I was so inspired by the conversation with Latifah. She shared how she had always wanted a smartphone, and when she got it she put it immediately to use: gathering prices to inform her farming business, sending videos and pictures to customers to sell goods, and asking questions about pregnancy and health. Much like cash, this chatbot is a tool that gives people the opportunity to decide what they need to improve their own lives. I was struck by the range of uses and unique questions each person explored using the chatbot - specific to their own life and needs."

Constance Jones, Founding CEO of the Imara Fund

Imara Fund is an innovative Africa-focused foundation, co-founded by Reed Hastings

We’re also learning from organizations focused on building responsible AI: Research from Anthropic found that Claude users in sub-Saharan Africa were among the most optimistic about AI globally, and more likely than respondents anywhere else to want to use it to start a business or access education. In our pilot, we see the same optimism among people at the margins of today’s AI ecosystem.

Most major AI labs say their mission is to extend the benefits to all of humanity. The world’s poorest communities are not an edge case; they’re central to that mission. We think they are one of the clearest tests of whether these tools actually work for the people who could benefit from them most.

We are still early in this work, and we do not think the best answers will come from one pilot or one organization. We’ve already launched a similar pilot in Malawi, and we still have much more to learn. These pilots are helping us assess uptake and usability. Next we run A/B tests - first of different prompts to see which responds most effectively to recipients, then of ‘cash’ versus ‘cash plus chatbot’ to evaluate economic indicators.

If you are building, funding, or studying AI for low-resource settings, we would welcome the chance to think it through together. Get in touch - carolina@givedirectly.org.