Africa AI Safety
Prize Competition
Building Practical Safeguards for Beneficial AI Deployment
AI systems are being deployed across Africa at pace — but safety infrastructure has not kept up. This prize competition empowers researchers, practitioners, and innovators to build foundational, context-appropriate AI safety tools for African communities.
Africa AI Safety Webinar, 27.08.2026
Q&A Session, 14.04.2026
The 2026 winners & finalists
Fifteen projects reached the final round of the 2026 Africa AI Safety Prize Competition.
Winners
-
1st place ($5,000)
SafeAlert: Lightweight Pre-Procurement Safety Test Suite for AI Models in Nigerian Mobile Money and Fintech Systems.
Fraud losses across Africa's fintech sector are rising fast, and AI is increasingly used to fight it, including the use of chatbots and fraud detection systems. But these AI models are rarely evaluated for the contexts they actually operate in before deployment. As a PhD researcher in AI safety and Andrew, a cybersecurity researcher focused on Nigerian banking systems, this gap connects directly to both our work. SafeAlert is a lightweight, pre-procurement safety test suite that checks whether AI models resist generating fraud content and correctly distinguish real Nigerian bank alerts from scams. It's currently a working prototype, tested across six commercial models. To move forward, we're looking for access to real fraud datasets, partnerships with Nigerian fintechs willing to pilot it, and feedback from the AI safety community. Project on HuggingFace. Project on Github.
-
2nd place ($3,000)
Clinical Resource-Adapted Benchmark: Evaluating Large Language Models for Actionable Medical Safety in Nigerian Healthcare Settings.
AI systems are increasingly trusted in medicine, but they're built on Western clinical standards and rarely tested against what's actually deliverable in low-resource settings. I care about making healthcare technology genuinely equitable. CRAB benchmarks how safely free tiers/open-source LLMs (GPT-4o-mini, Claude, LLaMA, Gemini) guide care across Nigeria's tertiary, general hospitals, and primary health centres (PHCs), scoring accuracy, contextual adaptation, actionable safety, and cultural recognition. It's a working prototype revealing a 35.9% dangerous-response rate at the PHC level. GitHub link. Next: more clinician-collaborators to expand tier/specialty coverage, and a validation study running CRAB against a deployed (or discontinued) Nigerian clinical AI tool to test whether CRAB scores predict actual observed failure modes, which would turn CRAB into a pre-deployment certification instrument.
-
3rd place (tie; $1,000)
Veza-A Community AI Harm Reporting System for Gauteng.
Veza is a WhatsApp tool that helps people in Gauteng report AI scams, including deepfakes, voice clones, and fake investments, in English, isiZulu, or Sesotho. I built it because people around me, including friends and family, have been scammed but stayed silent for fear of being judged. Veza provides a safe way to report scams, share experiences, and receive clear next steps, while turning reports into an early-warning signal for emerging scams. It is a working, tested, open-source prototype. I am looking for a trusted local partner to help pilot it with real users and native speakers to improve translations. Prototype links: (1) Live bot: WhatsApp "join column-event" to +1 415 523 8886 (runs on free hosting for the prototype, so the first reply may take up to a minute-after that it is fast.); (2) Code and documentation<; (3) Demo video
-
3rd place (tie; $1,000)
MediSafe-GH: A Clinical Safety Screen for Medical AI Assistants in Ghanaian Languages.
Health AI is deploying across Ghana, yet models quietly fail in local languages exposing patients to clinical risk. Built by KNUST biomedical engineering researchers, G-MASS (Ghana Medical AI Safety Screen) is an open-source Python toolkit benchmarking health AI safety across Twi, Ghanaian English, and English using 300 customer-facing clinical probes. Current Stage: Pip package, Hugging Face dataset and gradio app demo. Needs: GPU compute for full 5-model evaluations, language expansion validators (Ewe, Ga, Fante, etc), and local health AI deployment partners.
Spotlight recognition
-
Kámárí and Kámárí-Safe Open: AI-Enabled Online Content Restriction for African Minors.
I grew up in Ikirun, Osun State, where a lot of kids my age were left unprotected online, and I watched what it did to them. Now I'm an AI engineer, so I decided to build something about it. Age checks on betting and adult sites can look accurate on average while passing roughly a third of minors as adults, and they fail worse on darker skin. Kámárí is an open benchmark that measures minor pass-through rate and fairness across skin tones, plus trained v0 models, a web and Android app, and a self-hostable API. To go further I need consented African face data with verified ages, compute, and platforms willing to pilot it. Web app link.
-
Evaluating AI Safety Methods in Yoruba: A Case Study for Local Language Safety in Nigeria.
Safety evaluations often rely on refusal rates to assess whether models reject harmful requests. In African languages, however, a non-refusal response may reflect either harmful compliance or difficulty generating coherent language. This matters to me because Yoruba is my language, yet we have limited evidence on whether safety evaluations accurately reflect model behaviour in Yoruba. My project addresses this by evaluating matched English–Yoruba prompts using a three-way classification of refusal, harmful compliance, and capability failure across established safety benchmarks. The pilot is complete, with code and results available in the YorubaSafety GitHub repository. I now need native Yoruba speakers to strengthen translation and validation, and support to expand the work to more models and languages.
Finalists
-
AgriSafe-MA: AI Safety Benchmark & Bilingual WhatsApp Pesticide-Advice Filter for Moroccan Smallholder Farmers.
-
AmiSafe: A Privacy-First, Multilingual Browser Extension/WhatsApp bot for Community-Led Monitoring and Reporting of AI-Generated Harms Across Africa.
-
An Africa-anchored AI safety benchmark for satellite-based carbon credit estimation pipelines.
-
CodeSwitch-Safety: A Tri-Probe Evaluation Framework for Code-Switched AI Harms.
-
Ethio Safety Audit: Addressing AI Toxicity in Amharic.
-
Safe Mama: An AI Safety Framework for Maternal Health Guidance Delivered to Nigerian Women via Mobile Platforms.
-
Safeguarding Kenyan Agriculture & CBC Education with AI Shield Kenya: Lightweight Multilingual SIR Harm-Propagation Toolkit.
-
SafeHer: Extending a Live TFGBV Reporting App with a Swahili/Sheng Abuse Detection Layer and WhatsApp Access for Kenyan Women.
-
SEEING WHAT THE SYSTEM MISSES: A Community-Based Playbook for AI Policing Harms (SENTINEL).
Supported by leaders across the African AI ecosystem
Reference information from the 2026 edition follows below
The safety infrastructure for Africa's AI hasn't kept pace
AI systems are increasingly shaping decisions, services, and livelihoods across Africa. While global AI safety efforts have grown, their context-appropriateness for African languages, sectors, and social dynamics remains limited.
Without safety tools designed for African contexts, the gap between AI's potential and its actual benefit to communities may widen — not close.
Community-Level Monitoring of AI Harms in Africa
AI-related harms are often first noticed in everyday life — by workers who lose income, women targeted by deepfakes, or communities affected by misinformation. In many African contexts, these experiences are shared informally and never recorded, meaning early warning signs are potentially missed. We are looking for community-facing ways to notice and make sense of AI harms as they happen, especially where formal reporting does not exist.
How can early signs of AI-related social harm in African communities be noticed and recorded in ways that are accessible, culturally appropriate, and safe, so emerging risks become visible rather than ignored?
Strong submissions will:
- Ensure privacy and security of those reporting the harms
- Allow people with no technical or legal knowledge to share what is happening to them
- Work through everyday communication (e.g. voice, storytelling, local languages, trusted intermediaries)
- Identify pathways for analyzing and effectively reporting the findings, potentially even feeding them into the AI development cycle
Example approaches (feel free to explore beyond these):
- A voice-based or story-based way for people to share experiences and a monitoring platform
- A simple method for noticing patterns across many similar stories
- A small pilot showing how harms that were previously invisible can be seen
Read the full challenge brief and context in the Concept Note.
Download Concept Note ↓Context-Appropriate AI Safety Evaluation for African Deployment
AI systems are often evaluated using benchmarks developed in high-resource settings. These benchmarks may not capture harms, misuse patterns, or safety concerns that are particularly relevant in African contexts. As AI systems are adapted and deployed locally, there is a need for targeted, lightweight evaluation tools that reflect real-world risks in African languages, sectors, and social environments.
How can AI safety benchmarks be designed or adapted to better capture risks and harms relevant to African deployment contexts?
Assignment — applicants should choose one:
- Identifying Benchmark Gaps — Which safety risks relevant to African contexts are insufficiently captured by widely used AI evaluation benchmarks? What is missing? Why does it matter? How could it be tested?
- Designing Context-Specific Test Cases — How can a specific African-context harm (e.g. election misinformation, AI-enabled fraud, deepfake abuse, local-language toxicity) be translated into a concrete benchmark task? What would failure look like? How could performance be measured?
- Local-Language Safety Evaluation — How do models perform on safety-related tasks in African languages compared to English, and what evaluation methods can capture those differences?
- Lightweight Safety Test Suites — What small, reusable prompt sets or evaluation methods can be used to test AI systems for context-relevant harms without requiring large datasets or compute?
A strong submission might deliver:
- A small benchmark task or prompt set targeting a specific harm
- A comparison of model performance across contexts or languages
- A documented gap in an existing benchmark, with a proposed fix
- A lightweight safety evaluation suite that others can reuse
Read the full challenge brief and context in the Concept Note.
Download Concept Note ↓All stages are fully remote
$10,000 for the best AI safety solutions for Africa
Meet the Judges
Final submissions will be evaluated by a panel of experts in AI safety, African AI deployment, and responsible innovation.
How submissions are evaluated
Directly reduces a specific, named AI-related risk or harm in a defined African context. Clear theory of change — the connection between output and safety problem is concrete and justified.
The solution is visibly shaped by the specific context it targets — language, community dynamics, infrastructure constraints. Evidence of local knowledge or community involvement is strong.
Deliverable is complete, coherent, and documented well enough for others to use independently. Methods are sound, limitations honestly acknowledged, and the approach introduces something meaningfully novel.
Clear pathway from output to real-world use — names intended users, institutions, or communities and explains concretely how they would adopt or build on it. Contributes to community empowerment.
Unintended consequences are actively considered. Trade-offs are named. The submission demonstrates awareness of how the tool could be misused, and proposes concrete mitigations.
Africa AI Safety Webinar
Registration closes 25 August 2026 (anywhere on Earth)
Frequently Asked Questions
Can we apply as a team?
Yes. If your team wins, the financial prize will be transferred to the lead author submitting the proposal.
Do I need official ethics approval?
Entrants should adhere to the necessary ethics requirements depending on their projects.
Can you disburse prizes to my country?
We can transfer prize payments to most countries. Please contact info@casa-ai.org before submitting your application to confirm your eligibility. Please note that currently, we are unable to transfer funds directly to certain countries, including Russia, Belarus, the Democratic Republic of Congo, Zimbabwe, and Lebanon.
I'm not affiliated with any institution. Can I still apply?
Yes. In the Affiliation field, simply write "Independent."
Can I submit more than one pitch?
Yes — up to one pitch per track, so a maximum of two submissions total.
Can I turn my proposal into a publication afterwards?
Yes. You are welcome to repurpose your submission into a publication after the competition.
Can I submit a previously developed idea?
Yes, as long as the idea has not already been operationalised by you or someone else.
I've already received funding for my idea. Can I still apply?
Yes, but you must disclose the funding amount and source in the Conflict of Interest section of the application form.
How will winners be selected?
In two stages: CASA shortlists up to 15 RFP pitches, then shortlisted candidates submit full proposals reviewed by an expert panel. Submissions will be evaluated through a double-blind review process. Judges will not have access to applicants’ personal identifying information. The identities of the judges will be disclosed in due course, but the allocation of judges to specific submissions will remain confidential.
Ready to build safer AI for Africa?
Submit your proposal and join a growing community of AI safety innovators.
Apply Now →