Digital Futures Lab, on behalf of the Global South Network for Trustworthy AI (members include Global Centre on AI Governance, South Africa; ITS Rio, Brazil; CeRAI, India; International Innovation Corps, India; Derechos Digitales, Chile; GxD Hub, India (an initiative of LEAD at Krea University (IFMR); The Collective Intelligence Project; UCT AI Initiative, South Africa)

Building Safety Infrastructures for the Global South For many countries in the Global South, AI systems are already being deployed in high-stakes contexts, often without the ability to independently assess how they perform across different contexts. Existing evaluation and impact assessment approaches are largely focused on Western contexts, with limited contextual evidence from the Global South to inform upstream design and governance processes. The Global Dialogue can take concrete steps to address this gap by securing concrete commitments from Member States to build contextual evaluation capacity and localised safety infrastructure. This can include: - Supporting Localised Evaluations and Impact Assessments: Dedicated funding and technical support for multilingual evaluations, culturally grounded benchmarks, and the infrastructure required to run participatory, multistakeholder evaluations across the lifecycle of AI systems. This includes integrating gender-responsive and intersectional lenses to assess differential risks, exclusions, and harms, particularly for historically disadvantaged and underrepresented groups. - Ensuring Independence and multistakeholder collaboration: Establishing independent AI evaluations that involve a wide set of actors - including supporting local research institutions and civil society organisations to conduct ongoing, post-deployment product, user, and impact assessments. - Establishing Feedback Loops: Creating structured pathways for insights from these contextual evaluations to inform AI design/deployment as well as global governance processes. - Preventing Safety-washing: Setting minimum expectations for transparency in AI safety evaluation practices, including disclosure of methods, datasets, and known limitations, to prevent safety-washing. Lastly, the Dialogue also has an inherent mission to address the urgent need to define mechanisms for monitoring, oversight, supervision and certification of AI systems in a way that allows harm not to be escalated or replicated at the global level. Therefore, it should also work towards defining the criteria for analysing compliance of AI products and applications within their production chain and, thus, to support countries in making providers accountable for any violations.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

Our selection reflects the areas where we see the most urgent need for collective action, and where the Global South Network for Trustworthy AI is positioned to contribute. - Safe, secure and trustworthy AI is a priority because, in many Global South contexts, systems are already being deployed in high-stakes settings without sufficient evidence of safety or fitness for purpose and amidst uncertainty about the kinds of on-ground outcomes or impacts AI solutions are bringing about. Advancing safe and trustworthy AI, therefore, requires moving beyond high-level principles towards context-specific evaluation, impact assessments, and tailored risk mitigation strategies. - The social, economic, ethical, cultural, linguistic and technical implications of AI are central because safety cannot be understood in isolation from context. Current approaches often overlook how systems perform across languages and social environments, leading to uneven outcomes and unaddressed harms. - Ensuring the interoperability of governance approaches is critical given the fragmented nature of current efforts, while recognising the contextual, institutional and normative differences across jurisdictions. Many countries are grappling with similar challenges but lack access to shared methods and tools. There is a clear need for a structured exchange that allows approaches to be adapted across contexts, rather than developed in isolation. - Finally, transparency, accountability, and human oversight underpin all other priorities. Without access to information on how systems are built, evaluated, and deployed, it is not possible to generate evidence, enable scrutiny, or establish meaningful oversight. Together, these areas represent where the Dialogue can move from principle-setting to enabling practical, context-sensitive AI governance.

Two cross-cutting issues that are not explicitly captured in the listed themes are power asymmetries/market concentration in the AI ecosystem and the environmental implications of AI systems. First, questions of market concentration and political economy play a central role in shaping AI governance outcomes. The development and deployment of advanced AI systems are currently concentrated among a small number of actors with significant technical and financial resources. This influences not only access to infrastructure and models but also whose priorities, values, and risk frameworks are reflected in system design and governance processes. While elements of this are implied across themes related to inclusion, development, and governance, a more explicit focus is needed to address how these asymmetries affect the ability of different countries and communities to meaningfully participate in and shape AI trajectories. Second, the environmental implications of AI systems warrant more direct attention. The resource intensity of model development and deployment, including energy and water use, has uneven impacts across regions. These considerations are not fully captured within existing social or economic framings and require dedicated attention in governance discussions, particularly as AI adoption scales.

Across many countries in the Global South, AI is increasingly being deployed in critical sectors such as welfare delivery, agriculture, and healthcare. However, the technical and the institutional apparatus for governing and monitoring such systems remains largely underdeveloped: - At the model level: Existing safety guardrails and benchmarks are largely designed for high-resource languages, leading to weaker safety protections and more frequent model failures in local contexts. This disproportionately affects groups such as women and children, who are more likely to encounter biased or inappropriate outputs without effective avenues for redress. - At the institutional level: There is limited public consultation on AI deployment in public services, and few requirements for pre-deployment impact assessments, especially for underrepresented groups. Regulatory and judicial oversight mechanisms remain underdeveloped, while post-deployment monitoring is rare. Combined with limited technical capacity within public institutions, this creates conditions where AI systems are adopted and scaled without meaningful scrutiny. - There is also a general failure to observe existing obligations regarding access to information, transparency and data protection in the deployment of AI systems. Considering the aggregated opacity they bring to the public function, this implies a limitation on the possibilities for oversight and redress in case of violations of fundamental rights. There is growing recognition of the need to build localised AI safety infrastructure, including multilingual evaluation datasets, independent auditing capacity, and mechanisms for ongoing monitoring. Civil society organisations and public-interest researchers are beginning to play a critical role in shaping these efforts, particularly by grounding evaluation in real-world use and lived experiences. If supported, these developments could enable more accountable and context-sensitive AI deployment. Without them, current governance gaps risk entrenching existing inequalities in both access to and oversight of AI systems.

The Global Dialogue can advance international cooperation by creating structured pathways for knowledge-sharing across countries with very different levels of capacity when it comes to AI safety. While a few countries in the Global North have dedicated AI safety institutions (with some, such as the UK, also shifting focus from safety to security in recent institutional mandates) and technical expertise, most countries do not. The Dialogue can help bridge this divide by enabling the sharing of evaluation and impact assessment methods, governance approaches, and technical practices, particularly for assessing AI systems in different languages and social contexts. This is critical in a fragmented landscape where countries are moving at different speeds and often in isolation. By pooling insights across contexts, the Dialogue can support more plural, context-sensitive approaches to AI governance without forcing convergence on a single model.

The Global Dialogue can engage with emerging initiatives such as the Global South Network for Trustworthy AI, a recently launched, civil society-led platform focused on strengthening AI evaluation and governance across diverse deployment contexts. The Network brings together organisations from Asia, Africa, and Latin America to better understand how AI systems perform in real-world settings, particularly across different languages, sectors, and institutional environments. Its early work is focused on generating grounded evidence on deployment risks and gaps that are not captured by existing safety and evaluation approaches. This includes efforts to assess how models perform in multilingual settings, document harms that arise in practice, and explore how governments alongside research institutions, civil society, and affected communities can more effectively scrutinise AI systems before adopting them in public services. Additionally, it can also actively work with communities working towards building safe, inclusive, rights-based and sustainable AI applications, such as the AI4D Network of Labs. This initiative brings together a diverse set of research and innovation laboratories across Africa that collectively advance responsible, locally relevant AI research and solutions. The added value of the Global Dialogue would be to work with such initiatives to ensure that insights from deployment contexts inform global governance processes. This includes creating channels for evidence-sharing, connecting these efforts to standards-setting and regulatory discussions, and supporting the development of local capacity for independent evaluations. By engaging with actors of this kind at an early stage, the Dialogue can help shape a more inclusive and grounded evidence base for global AI governance as the field continues to evolve.

With regard to AI safety and evaluations, (i) governments can share regulatory approaches and implementation challenges, (ii) industry can provide transparency into system design, evaluation practices, and deployment constraints, (iii) civil society and public-interest researchers can surface evidence on real-world impacts, particularly from underrepresented contexts, and lastly, (iv) academic and technical communities can contribute methods for evaluation and risk assessment. To make these contributions meaningful, the AI Dialogue can be structured around focused thematic tracks with clear outputs. Each track should combine technical and policy perspectives and be anchored in specific use cases or sectors. The Dialogue should also include dedicated spaces for evidence-sharing, where various stakeholders can present concrete findings. A key design element for the Dialogue should be continuity. This could include working groups or task-oriented clusters that carry discussions forward between convenings, as well as mechanisms to synthesise inputs into actionable recommendations.

- Civil society and grassroots organisations, which are often the first to encounter and document harms arising from AI deployment, particularly in public services. Their proximity to affected communities positions them to identify risks that are not visible in technical or policy forums, yet their participation remains limited. - Local communities of researchers and developers working on multilingual and low-resource language AI remain underrepresented. Initiatives such as Masakhane and Deep Learning Indaba bring deep expertise on how models perform outside high-resource, English-dominant contexts. They are often best placed to surface issues related to bias, safety, and usability in systems that will be used by the global majority. How could they be included: - Organise Dialogue/specific regional sessions and related events in Global South countries to include grassroots organisations/independent researchers unable to travel to Global North countries - Live-stream deliberations at the Dialogue, with due consideration and safeguards to protect the privacy and freedom of speech of attendees - Support travel and participation in the Dialogue, including services for obtaining travel documents Additionally, we would also like to reiterate the suggestions recently submitted by civil society organisations to the Co-Chairs of the Global Dialogue [https://www.derechosdigitales.org/wp-content/uploads/2026/04/Letter-to-UN-AI-Global-Dialogue-Co-Chairs-Procedural-Recommendations-1.pdf], including the recommendations for (i) appointing a civil society representative liaison to support the Co-Chairs to facilitate meaningful multistakeholder engagement ahead of, during, and following the Global Dialogue, (ii) structuring sessions format that allow civil society and other non-governmental and governmental stakeholders to respond to each other in real time, (iii) convening support from government and non-government actors to offer financial and technological assistance for civil society stakeholders who are less able to acquire the necessary technology and connectivity on their own to engage with the Global Dialogue in a sustained manner.