Towards a Safer South: Launching the Global South AI Safety Research Network

Contents

Executive Summary

The Global South AI Safety Research Network was officially launched at the India AI Impact Summit to address the critical under-representation of Global South perspectives in global AI safety infrastructure. The network brings together research institutions, civil society organizations, and governments from Asia, Africa, and Latin America to conduct real-world AI safety evaluations, build localized trust mechanisms, and elevate Global South voices in international AI governance—particularly in contexts where AI systems are deployed rapidly in high-stakes sectors (healthcare, education, judiciary, government) amid low institutional capacity and deep societal inequities.

Key Takeaways

The Global South bears disproportionate risk from AI but has minimal voice in defining safety. This power imbalance must be corrected through institutionalized inclusion of Global South perspectives in international governance, standard-setting, and evaluation frameworks.
Real-world safety evaluation is a prerequisite, not an afterthought. Lab-based testing cannot anticipate contextual harms. Sustained post-deployment monitoring, incident reporting, and community-led evaluation must become standard practice—requiring infrastructure investment and institutional change.
Language and culture are not cosmetic add-ons to AI safety—they are foundational. Meaningful safety in multilingual, multicultural contexts requires teams with lived experience in those contexts, intentional design choices, and evaluation methodologies that capture sociolinguistic and cultural dimensions of harm.
Procurement and policy are underutilized levers for market shaping. Governments in the Global South can drive responsible innovation through procurement standards and public policy rather than waiting for voluntary corporate safety initiatives. India's "third way" of AI governance offers a model.
This network succeeds only if it drives problem-solving collaboration, not just information-sharing. Cross-border collaborative research on shared safety challenges—not siloed national efforts—will demonstrate the network's necessity and catalyze sustained support.

Key Topics Covered

Global South representation gaps in AI safety governance and safety institutes
Contextual and localized safety evaluation beyond English-dominant benchmarks
Multilingual AI safety and the inadequacy of translation-only approaches
Real-world harm assessment across diverse linguistic, cultural, and infrastructural contexts
Gender-based harms and gender safety in digital spaces
Civil society's role in identifying risks invisible to lab-based testing
Regional capacity building and infrastructure investment needs
Incident reporting systems and feedback loops for safety issues
Procurement as a governance lever for responsible AI adoption
Coordination across multiple AI safety networks and initiatives globally
Post-deployment safety evaluation and emergent harms
Sovereign capability building in the Global South
UN mechanisms (Global Dialogue on AI, Scientific Panel on AI) and multilateral governance

Key Points & Insights

Under-representation is structural: The Global South remains severely under-represented in global AI safety institutions. Only Kenya (as noted by Ambassador Tiggo) is a member of the International Network of AI Safety Institutes; many Global South countries lack even their own safety or oversight institutes, creating a power imbalance in defining what constitutes "risk," "harm," and "safe performance."
Safety is contextual, not universal: Safety cannot be defined as a one-size-fits-all technical metric. It varies by language, cultural norms, religious beliefs, gender dynamics, and local infrastructure. Current benchmarks—predominantly English-language—miss critical societal risks specific to Global South deployment contexts.
Language is more than translation: Large language models with strong translation capabilities still fail to capture lived meaning and local context. Example: "waters have broken" in a maternal health context is a critical incident, but literal translation could obscure clinical urgency. Models must be trained on sociolinguistic understanding, not just vocabulary.
Emergent harms occur post-deployment: Safety issues often surface only after a system is live and in widespread use—not during pre-release lab testing. This requires sustained, ongoing evaluation and robust incident-reporting mechanisms that currently don't exist in many Global South contexts.
Civil society is uniquely positioned to surface real-world harms through grounded evidence. Their proximity to deployment contexts, cultural knowledge, and credibility with local communities makes them essential partners—not peripheral stakeholders—in safety evaluation.
Misuse takes multiple forms: Technology can enable both unintentional harm (e.g., content moderation failures due to underpaid workers and insufficient non-English oversight) and intentional harm (e.g., deepfakes for election disinformation, AI-powered surveillance). Design must anticipate both.
Infrastructure gaps compound safety risks: Absence of rule of law, weak civil society mechanisms, limited computational access, and delayed feedback loops mean harms persist longer and scale faster in Global South contexts. These structural gaps must be explicitly addressed in safety design and evaluation.
Benchmarking is a governance act: Benchmarks are not neutral. They encode power—determining which risks are measured, which harms are prioritized, and what "safe performance" means. Concentrating this power in a handful of institutions is problematic; decentralization is essential.
Safety must be costed into financial systems: Without regulatory mandate or financial penalty for unsafety, companies have insufficient incentive to invest in contextual evaluation. Safety must become a line-item cost structure, not a peripheral concern.
Coordination across networks is critical: Multiple initiatives (African network, Chinese network, UN network, this network) are launching simultaneously. Without coordination, fragmentation wastes resources; with it, there's potential for significant force multiplication.

Notable Quotes or Statements

Orvisha Nea (Founder, Digital Futures Lab): "Independent civil society organizations are uniquely positioned to address this gap. Their proximity to real world deployment context enables them to surface risks that are invisible to lab-based evaluations or testing."
Ambassador Philip Tiggo (Kenya): "The global south has always been excluded from this conversation... that model that is not inclusive to a global majority that in most cases bears the brand and the impacts of AI is not acceptable."
Abishek Singh (India AI Mission): "The objective is not to stifle innovation. The objective is to ensure that more and more users benefit from the usage of AI... in a responsible manner... in a safe manner... in a trustworthy manner."
Dr. Rachel Sabande (Gates Foundation): "If you translate [clinical language] from the local language to English... that will literally mean 'I have thrown away water.' So if the model is not trained to understand that context then you will miss that flag."
Quentin Chao Lambert (UN Office for Digital & Emerging Technologies): "Safety can be thought of as a relational concept not an absolute concept... when we talk about safety, we always think about who is safe and who is protected from which threats."
Jennai Choi (Masakani African Language Hub): "There are over 2,000 languages documented on the African continent. Masakani is only working on 50 of those African languages... what you then find is when people are deploying technologies, people just don't speak [one variant of] Kiswahili across East Africa."
Amir Banatam (Cognizant Chief Responsible AI Officer): "There is no penalty of not being safe. So as long as there is no constraint to put safety as a cost structure which strong mandate companies will not pay attention... to enough attention."
Dr. Balaram Raindran (IIT Madras): "There are just too many [safety] initiatives getting launched... We have to figure out a way how we coordinate operations among these initiatives... that would be a great force multiplier."

Speakers & Organizations Mentioned

Founders & Leadership:

Orvisha Nea, Founder & Director, Digital Futures Lab
Dr. Balaram Raindran, Head, Center for Responsible AI, IIT Madras

Government & International:

Abishek Singh, India AI Mission
Ambassador Philip Tiggo, Special Envoy on Technology, Republic of Kenya
Quentin Chao Lambert, Chief of Office & AI Lead, UN Office for Digital and Emerging Technologies

Research & Civil Society:

Natasha Krompont, VP and Chief Responsible AI Officer, Microsoft
Dr. Rachel Sabande, Senior Program Officer AI for Africa, Gates Foundation
Jennai Choi, Director, Masakani African Language Hub
Amir Banatam, Chief Responsible AI Officer, Cognizant

Founding Partner Organizations (Network Members):

Sir (Sarai) — IIT Madras
Global Center for AI Governance
IT Rio (Brazilian research organization)
International Innovation Corps
Collective Intelligence Project (CIP)
CARA
GXD Hub
Kaya Organization (Samishka project)

Referenced Initiatives & Institutions:

New Delhi Frontier AI Commitments
UN Scientific Panel on AI
UN Global Dialogue on AI Governance
Bletchley Park Summit (international AI safety declaration)
Seoul Declaration
Paris AI Safety Summit
UK AI Security Institute
Microsoft Research
Cognizant Labs (Bangalore and San Francisco)

Technical Concepts & Resources

AI Safety & Evaluation Frameworks:

Multilingual benchmarks (language-specific performance evaluation)
Contextual harm assessment (localized to cultural, linguistic, infrastructural contexts)
Real-world deployment evaluation (post-launch monitoring, not pre-release only)
Gender harm taxonomy (incident reporting database for gender-related harms)
Incident reporting systems (capturing emergent harms, safety issues, misuse patterns)
Red teaming (adversarial testing; noted as a capacity gap in Global South)
Sociolinguistic evaluation (capturing lived meaning, not just vocabulary/translation)

AI Models & Applications Referenced:

Large language models (LLMs) — multilingual performance, contextual appropriateness
Small language models (SLMs) — localized, low-cost alternatives better suited to Global South contexts
Generative AI tools (healthcare, government, education use cases)
Voice-based AI systems (particular challenges in cultural appropriateness, gender representation)
Sentiment analysis and content moderation (disproportionately affecting Global South languages)

Datasets & Benchmarking Initiatives:

Samishka project (CIP + Kaya + Microsoft Research) — contextually-aware evaluations
African language datasets (Masakani working on 50 African languages; 2,000+ documented)
Multilingual benchmark datasets (New Delhi Frontier AI Commitments initiative)
Incident databases (planned as part of network activities)

Policy & Governance Tools:

Procurement standards (leverage government purchasing to drive responsible innovation)
New Delhi Frontier AI Commitments (multilingual performance data sharing, usage data transparency)
Policy interventions informed by shared usage data
Full life-cycle accountability (from model development through environmental impact)

Methodological Frameworks:

Vulnerability-based approach (relational safety concept — who is safe from which threats)
Contextual evaluation methodology (embedding cultural, linguistic, infrastructural context)
Community-led evaluation (inclusion of civil society, marginalized groups in design & testing)
System-level evaluation (models as part of larger applications, infrastructure, data ecosystems)

Referenced Research & Reports:

Yeshua Benji's scientific panel report (frontier AI risks)
Secretary General's High Level Advisory Body on AI (UN conceptual framework)

Document Generated: This summary synthesizes the complete launch event transcript from the Global South AI Safety Research Network, delivering technical precision while preserving the lived examples and human context that made the event's arguments compelling.