Design · Technology · AI for Good
← All work
Girl Effect · 2018 to now · South Africa, Kenya & India

AI Health
Chatbots

Building AI that young people trust with their hardest questions.

Two girls smiling at a red phone together
My role
Product & design, 0 to 1Built on 15 years of research and design with young people. Designed slowly, on purpose. AI & infrastructureThe move to GenAI and our LLM stack SafetyGuardrails, safeguarding and ethical AI guidelines TeamsNairobi, Johannesburg, Cape Town, Delhi, Mumbai and the UK
The brief

Give young people somewhere safe to ask what they can't ask a parent, a teacher or a nurse. And make sure a real person is there when an answer isn't enough.

Two girls reading an issue of Ni Nyampinga magazine together
The journey

From 10 questions a quarter 100,000 a day

2013 · Rwanda By hand

It started with Baza Shangazi, an Ask Auntie column. Young people wrote in, and real people wrote back.

2013 to 2015 · Rwanda & Malawi On air

The column grew into a text-in radio show. Questions arrived by SMS and were answered on air.

2016 to 2017 · Global Mobile site

The questions, and the answers, moved onto phones, ready any time of day. Free to access through our partnership with Facebook Free Basics.

2018 · South Africa The first chatbot

Built on years of real questions, so it spoke the way young people actually do.

Today · India, Kenya & South Africa GenAI, with humans

Guardrails throughout, trained people a tap away. Up to 100,000 questions answered daily.

A health worker in a headset responding to young people at her laptop

Three decisions
that shaped it

01Start with the questions, not the tech

Long before AI, we were answering young people by hand, through an Ask Auntie column and a text-in radio show in Rwanda. Those years of real questions became our foundation, and later the test sets we use to check the AI.

02Design the handoff first

I'm always going to bang this drum: automated answers alone are never enough. We built every chatbot around the moment a trained person steps in, and the route to real care after that.

03Meet them in their own words

Nobody texts in textbook English. We built for the way young people really write, from Sheng in Nairobi to Hinglish in Delhi, so no one has to translate a worry to be heard.

The human behind every chatbot: a trained health worker

Today, one of the largest AI health deployments across Africa

Benefiting 3 million lifetime users.

Big Sis on a phone screen
Meet Big Sis → South Africa · 2.5M users
Bol Behen on a phone screen
Meet Bol Behen → India · 350K users
WAZZII on a phone screen
Meet WAZZII → Kenya · 135K users

What changed

Independently evaluated
~2xBig Sis users were nearly twice as likely to seek mental health support as peers it didn't reach, the strongest result of any channel.
+21ptsrise in formal mental health service use where Big Sis was paired with outreach in the community.
225K+young people sought care through the chatbots in 18 months, and over 100,000 spoke to a trained counsellor.
+14ptsmore young women in Kenya saw contraception before marriage as normal where Tukisonga and WAZZII reached.

"With the chatbot you could ask things you wouldn't ask someone face-to-face."

Focus group participant, South Africa
How it works

Powerful AI, wrapped in safety

We use frontier AI models, wrapped in multiple layers of safety, to protect and support young people.

The risks of AI
Risks to the content
Hallucination
Bias
Cultural irrelevance
Harmful or inaccurate information
Risks to the human
Dehumanisation
Emotional dependency
Failure to spot a young person in crisis
Risks to the system
Exposed sensitive conversations
Excluding young people without smartphones, data or literacy
1. Community & expert governance
Youth advisors, local experts and safeguarding partners steer what we build.
2. Content & response guardrails
Vetted, local content. Locked tone and scope.
3. Real-time safety filters
Toxicity, off-topic and crisis detection.
4. Human in the loop
High-risk chats go to trained health workers.
5. Continuous evaluation
Safeguarded, reliable, accurate, resonant.
6. Frontier AI models
Several best-in-class models.
How we mitigate risk
Co-created with young people and communities. Continuous expert oversight.
Evidence-based information only, made for local contexts and languages.
Filters catch issues before they reach a young person.
Real people. Real support. The right care at the right time.
We test, measure and learn in every conversation.
End-to-end encryption. Minimal data. Built for trust.

AI can't replace human care. But with the right guardrails, it can help more young people get the support they need, when they need it.

What I've learned

A chatbot alone is not the strategy.The biggest gains came when chat was layered with face-to-face outreach. People meet regularly, and the chatbot carries them in between.
Private doesn't mean risk-free.We collect as little as we possibly can, and keep trained humans watching the edges.
Trust lives in small moments.A reply that sounds like a friend. A handoff that actually happens. Get those right and young people come back.
Recognition
Yale School of Public HealthTaught as a case study, 2025
Stanford Center for Digital HealthFeatured in its GenAI for health white paper, 2025
Fast CompanyWorld Changing Ideas winner for WAZZII, 2025
Go deeper
The full story at Girl Effect → Yale case study → Our white papers →
Next project Ni Nyampinga →