Collaborating to Scale AI Adoption in the Global South

Contents

Executive Summary

This panel discussion explores the critical role of multistakeholder collaboration in developing and deploying AI solutions responsibly across the Global South. The panelists—representing academia, industry, nonprofits, government advisory roles, and startups—emphasize that sustainable AI adoption requires breaking down power dynamics, centering local languages and contexts, ensuring equitable data ownership, and educating diverse stakeholders rather than pursuing extraction-based approaches that concentrate benefits in the Global North.

Key Takeaways

Multistakeholder collaboration is not optional—it's foundational. From design through deployment, AI solutions require representation from governments, academia, nonprofits, private sector, end-user communities, and frontline workers. No single stakeholder group can ensure responsible, impactful AI.
Ownership and power dynamics matter more than access. The Global South's role in the AI value chain must shift from low-wage data collection to model ownership, governance, and decision-making authority. This requires explicit conversations about data ownership, licensing, and community control.
Localization is not a feature—it's a prerequisite for adoption. Language, offline capability, evaluation against local outcomes, and integration into existing workflows are non-negotiable. Forcing Global North–designed systems into Global South contexts without adaptation guarantees failure and mistrust.
Slow, participatory approaches yield better long-term outcomes than fast extraction. Extractive, rush-to-deploy strategies break trust and close doors. Investing in community relationships, equitable governance, and shared benefit models builds sustainable ecosystems.
Rhetoric must be managed alongside technology. Combating misleading narratives about job displacement, AGI timelines, and AI capabilities is as important as building systems. Education and transparent communication protect both public trust and informed policy-making.

Key Topics Covered

Multistakeholder collaboration frameworks for AI development and deployment
AI governance at global scale, particularly through UN advisory boards and policy mechanisms
Language and localization challenges in Africa and South Asia (2,000–3,000 African languages; low-resource language representation)
Data equity and ownership in the Global South; power dynamics in data collection workflows
Weather prediction, healthcare, agriculture, and education as domains for AI impact in underserved regions
Offline and low-resource model deployment (mobile-first approaches; compressed models on low-cost smartphones)
Evaluation methodologies that account for local context rather than one-size-fits-all metrics
Digital divide and literacy barriers to AI adoption
Responsible AI and trustworthiness in financial and B2B applications
Education and public understanding of AI to counter misinformation and fear
Workforce participation of Global South communities in AI value chains (data collection, annotation, evaluation)
Sovereign AI and model ownership for nation-states and communities

Key Points & Insights

UN-led global governance is essential but insufficient: The UN High-Level Advisory Board on AI produced "Governing AI for Humanity" (September 2024), emphasizing inclusive, all-inclusive AI rather than solutions designed for the wealthy West or China alone. However, the UN is under-resourced and must include China and diverse Global South voices at every policy table.
Language is a lever for equity, not just a technical problem: African languages (Swahili, Amharic, Yoruba, etc.) aren't missing from AI due to technical barriers alone; the social, political, and legal dimensions of language choice matter. Communities must control decisions about whether and how their languages appear in ML systems. Extractive approaches close doors and break trust.
Data ownership and community agency must be reframed: Currently, data is collected cheaply in four powerhouse countries (India, Kenya, Uganda, Philippines) primarily because they speak English. The model-building and value capture happen in the Global North. Shifting this requires rethinking data as an asset owned by communities, with equitable licensing and community governance—not one-size-fits-all approaches.
Context-specific evaluation beats global baselines: Impact evaluation cannot rely on a single global framework. Evaluation must be participatory, grounded in local causal relationships, and measure what matters to end users (usability, literacy access, voice interfaces, integration into workflows)—not just model performance metrics.
Offline-first and low-resource deployment is viable and necessary: Models compressed onto $8–10 USD smartphones can run offline, enabling frontline workers (nurses, farmers, teachers) to benefit from AI without internet connectivity. This requires deliberate architectural choices and multistakeholder input from deployment contexts.
Financial/B2B AI reveals the importance of trust and domain expertise: Fintech and financial operations AI (invoice processing, transaction analysis) require domain experts (accountants, CFAs, CFOs) to collaborate with model builders. Domain knowledge—not just engineering—is essential; raw model performance without user trust and regulatory compliance means no adoption or revenue.
Extractive practices undermine long-term impact: Rushing to deploy without community input, failing to share benefits, or ignoring local governance preferences damages trust and closes future collaboration opportunities. Slow, participatory approaches yield more sustainable ecosystems.
Education and rhetoric management are critical: The public is being "conned" by hype narratives (AGI by end of year; programmers will be obsolete). Educators must teach students to question AI critically, understand limitations, and think about downstream impacts collaboratively.
Open-weight models + local datasets enable localization without vendor lock-in: Strategies like Google's Gemma and Med-Gemma paired with curated local datasets (e.g., AfromedQA for medical QA across 16 African countries) allow Global South institutions to benchmark, evaluate, and fine-tune models for local relevance without depending on proprietary vendors.
Startups and SMEs need a voice in policy: Small organizations building solutions for end customers need seats at governance tables alongside big tech and governments. They navigate daily the tensions between impact, compliance, sustainability, and responsible AI—and their insights must shape policy.

Notable Quotes or Statements

Wendy Hall (UN AI Advisory Board): "The UN, although potentially flawed, is the only game in town when talking about global and inclusive governance of AI for the Global South and China."
Aisha Wal Bryant (Google Research Africa): "Your unique context is not a challenge or a barrier—it's a breakthrough or innovation that you can bring to the world."
Vukosi Marate (University of Pretoria, Laba AI co-founder): "If you only look at this as a technical problem, you will very quickly run out of railway track. You have to look at the social-technical dimension."
Safia Hussein (Karya): "Let's not become extractive. The people contributing to this should either benefit from the technologies or see data as their own asset."
Wendy Hall: "We're being conned. People think AI is going to take their job, and they're scared because they don't understand. The key is education, education, education."
Makran Tapasi (Vadrani AI): "If you really want to deploy solutions for the benefit of humanity, it cannot happen with ML scientists sitting in one place trying to do their thing."
Nyati Chaya (Hyperbots Inc.): "The only way to crack a value chain that becomes ethical is through multistakeholder collaboration because we have this unequal divide of models from the Global North and data from the Global Majority."

Speakers & Organizations Mentioned

Speaker	Role/Organization	Key Focus
Rahman Pandi	CEO, Center for Responsible AI; Moderator	Panel moderation; multistakeholder frameworks
Prof. Dame Wendy Hall	Regius Professor of Computer Science, University of Southampton; UN High-Level AI Advisory Board; UK Government AI Adviser	Global AI governance, policy, UN initiatives
Dr. Aisha Wal Bryant	Senior Staff Research Scientist, Head of Google Research Africa (Accra, Nairobi)	AI for African languages, weather prediction, food security, agricultural tech
Prof. Vukosi Marate	Professor of Computer Science, University of Pretoria; ABSA Chair of Data Science; Co-founder, Laba AI	African languages, NLP, machine learning for low-resource contexts
Safia Hussein	Chief Impact Officer, Co-founder, Karya	Data collection, annotation, evaluation; workforce equity; impact measurement
Dr. Makran Tapasi	Principal ML Scientist, Vadrani AI; Assistant Professor, CVIT, IIIT Hyderabad	AI for social good; deployments in education, health, agriculture; model compression
Dr. Nyati Chaya	Co-founder, Hyperbots Inc.; former Adobe Research	NLP, multimodal understanding, responsible AI in fintech; policy inclusion for startups

Institutions/Entities Referenced:

UN High-Level Advisory Board on AI
Google Research Africa
Vadrani AI (nonprofit applied AI institute)
Laba AI (AI startup for African languages)
Karya (impact-focused AI/digital services)
Hyperbots Inc. (fintech AI startup)
University of Southampton
University of Pretoria
IIIT Hyderabad
Masakhani (distributed research network for African languages)
UK Government AI Office
Bletchley Park AI Summit

Technical Concepts & Resources

Models & Tools Referenced

Gemma: Google's open-weight LLM
Med-Gemma: Medical-domain version of Gemma
Multilingual LLMs: Fine-tuned with smaller, purpose-sourced datasets for specific languages/contexts
AfromedQA: Pan-African medical Q&A dataset (25,000 QA pairs across 16+ countries, 32 disciplines)
Wahal NLP: Open-source speech dataset (11,000 hours ASR + text-to-speech) for Wolof language; hosted on Hugging Face
Global Weather Nowcasting: Satellite-data-driven precipitation estimation (5 km radius) for data-scarce regions

Methodologies & Approaches

Newborn Anthropometry via Vision: Computer vision + mobile ML to estimate infant weight from video, deployable offline on $8–10 smartphones
Equitable Licensing: Legal/governance frameworks giving communities control over language/data use (not one-size-fits-all)
Participatory Evaluation: Involving end-user communities and lived experience in impact assessment
Model Compression: Fitting trained neural networks onto low-cost, offline-capable mobile devices
Multilingual Fine-tuning: Adapting pre-trained models with smaller, locally-sourced datasets rather than training from scratch
Partnership-Led Data Collection: Communities own and control data collection processes (e.g., Makerere University + partners for African language speech data)

Domains of Application

Weather and climate: Nowcasting and precipitation forecasting in data-scarce regions
Healthcare: Medical language models, diagnostic support, newborn health assessment
Agriculture: Crop health, food security
Education: Literacy support, language-inclusive learning tools
Financial Operations: Invoice processing, transaction analysis, OCR + reasoning
Language Technology: ASR, text-to-speech, translation, NLP for low-resource languages

Key Concepts

Extractive vs. participatory approaches: Extraction (fast, short-term benefit to Global North) vs. participatory (slow, sustainable, community-owned)
Counterfactual evaluation: Difficulty isolating AI impact when AI is embedded in broader systems
Offline-first architecture: Designing for regions/contexts without reliable internet
Sovereign AI: Nation-state and community ownership of foundational models and governance
Digital divide: Encompasses not just internet access but literacy, language, understanding of AI, and policy literacy
Power dynamics in value chains: Concentration of model-building (Global North) vs. data collection (Global South, low-wage)

Research & Governance Initiatives Referenced

"Governing AI for Humanity" (UN High-Level Advisory Board, September 2024) — First major report on inclusive global AI governance
Global Scientific Board for AI (UN, recently announced) — Implementation of UN recommendations
Global Dialogue on AI Policy (AI for Good Conference, July, pending)
Masakhani Research Network — Distributed collaboration on African languages and NLP practices; ongoing evaluation research
Bletchley Park AI Summit (November 2022) — First global AI safety summit; launched AI Safety Institute

Note on Transcript Quality: The provided transcript contains significant repetition, audio artifacts, and incomplete sentences typical of automatic transcription. Where clarity was unclear, context from surrounding discussion was used to infer speaker intent.