Bridging AI divides: capacity-building, access and digital foundations

Robert Opp defines the AI divide in two linked ways: unequal access to compute, skills and capacity, and unequal ability of institutions to adopt AI safely. His point is that countries are already receiving AI through markets and procurement, so the challenge is not only getting access but ensuring national uptake is inclusive and governed well.

He explicitly says there are likely many AI divides, but highlights a divide around access, meaning affordable and available access to compute power, capacity and skills . He then adds a divide around adoption, explaining that AI is already entering countries through procurement systems and vendors faster than institutions can keep up, so the task is to ensure AI is safely and inclusively adopted at country level .

Opp argues that AI has moved beyond being just a technical topic and has become an implementation and systems problem. That means countries need strong digital foundations and institutional capacity, not only interest in the technology itself.

He states that in UNDP's work across more than 170 countries, AI is evolving from a technology conversation into an implementation and systems challenge . He says this shift underscores the importance of capacity building and good digital foundations .

Robert Opp presents capacity building as essential because institutions are struggling to keep up with the speed of AI diffusion. His position is that countries need people, skills and institutional readiness if they are to manage adoption responsibly.

He notes that AI is entering countries through procurement and vendors faster than institutions can keep pace . He then says UNDP sees AI as an implementation challenge that underscores the importance of capacity building and strong digital foundations .

Opp argues that a key part of the AI divide is institutional lag. Because AI deployment is moving faster than governance and administrative systems, national adoption has to be managed in a way that is both safe and inclusive.

He says AI is entering countries through procurement systems and vendors and is outstripping the ability of institutions to keep pace . He therefore defines the divide around adoption as ensuring AI is safely and inclusively adopted at country level .

Opp frames the event as a structured dialogue designed to bring out multiple aspects of the AI divide. He uses the sequencing of formats to indicate that the session is meant to move from framing to discussion and then to wider intervention.

He introduces a fireside chat with distinguished guests as the next segment of the session . Later he explains that the afternoon will continue with two moderated panels separated by moderated discussions with the co-chairs .

Loretta Hieber Girardet argues that access alone is too narrow a conception of inclusion. In her view, real bridging of AI divides means countries and stakeholders must be able to shape AI in ways that fit their own needs and realities.

She says bridging AI divides is more than simply expanding access to AI . She explains that it is about strengthening conditions that allow countries and stakeholders not only to use AI, but also to develop, adapt, evaluate and govern it in ways that respond to their own needs and realities .

She stresses that AI cannot substitute for the basic ingredients of functioning public systems. Responsible use depends on digital infrastructure, trusted institutions, data systems and local capability being in place first.

From a disaster risk reduction perspective, she says AI creates opportunities but also reinforces the need for sustained capacity strengthening and technology transfer, especially in developing countries . She adds that AI cannot substitute for strong institutions, data systems, local expertise, community engagement or accountable decision-making, and therefore countries need digital infrastructure, institutional capacity, technical skills, governance frameworks and locally led solutions to use AI responsibly and safely .

Hieber Girardet sees long-term capacity strengthening, not one-off deployment, as necessary for countries to benefit from AI. She particularly emphasises developing countries, where technology transfer and institutional support remain critical.

She states that the opportunities from AI also reinforce the need for sustained capacity strengthening and technology transfer, particularly in developing countries . She later says the priority is ensuring countries have infrastructure, institutional capacity, technical skills and governance frameworks, alongside locally led solutions .

She argues that countries should not become dependent on AI tools that they cannot understand, interrogate or govern. For her, AI should be adapted locally and remain accountable to national and community realities.

She says stakeholders should be able to develop, adapt, evaluate and govern AI in ways that respond to their own needs and realities . She also warns that countries should not become dependent on technologies that cannot be interrogated, adapted or governed .

Hieber Girardet stresses that AI should not displace institutional accountability. Instead, governments need governance frameworks and decision-making structures that allow AI to be used responsibly, safely and effectively.

She says AI cannot substitute for accountable decision-making, nor should countries depend on technologies they cannot govern . She then states that the priority is not simply deploying tools, but ensuring countries have governance frameworks and locally led solutions needed to use AI responsibly, safely and effectively .

She offers disaster risk reduction as a concrete area where AI can deliver public value. At the same time, she insists these benefits depend on strong institutions and capacities, especially in developing countries.

She says AI is already beginning to transform disaster risk reduction through risk assessment, early warning systems and risk-informed investments with direct implications for public safety and protection of critical infrastructure and services . She then cautions that this opportunity still requires sustained capacity strengthening, technology transfer, strong institutions, data systems and accountable decision-making .

Samba Diouf argues that AI offers major opportunities, but those benefits will be lost if access remains unequal and the people developing AI are disconnected from those consuming it. He therefore frames AI governance as a responsibility to ensure it becomes a source of shared prosperity rather than deeper inequality.

He says AI offers many opportunities, but they can only be fully realised if there is unified access . He points to broad disparities in connectivity and says that people who develop AI do not align with the consumer . He then states that the collective responsibility is to make AI a driver of shared prosperity rather than a source of deeper inequality, especially by helping developing countries build their own capability and adapt AI to national priorities .

Diouf argues that countries need to build their own capabilities rather than just import AI. He links this to national development strategy, using Senegal's digital skills, AI, infrastructure and innovation pillars as an example.

He says countries, particularly developing countries, must build their own capability and adapt AI to national priorities . He cites Senegal's New Deal technology vision, built around digital skills, AI, digital infrastructure, public platforms and innovation .

He sees cooperation as necessary to make AI benefits universal rather than concentrated. His emphasis is on practical solutions, stronger partnerships and recommendations that lead to a more inclusive global AI system.

He says collective effort must be strengthened so that the benefits are for all . He adds that he hopes the session will identify practical solutions for stronger partnership and generate ambitious recommendations for a more inclusive, equitable and development-oriented global AI .

In his concluding synthesis, Diouf reframes digital sovereignty less as owning the whole technology stack and more as building human capability to understand and use technology. He also argues that smaller countries cannot build sovereign AI alone and should instead cooperate regionally, especially on data governance.

He says digital sovereignty, in his understanding, is about upskilling because countries may not have the technology but can have people able to understand and execute it . He also says smaller countries cannot afford the scale of investment or demographic advantage of larger countries, so they should collaborate regionally and build shared strategies, including for data governance .

Diouf argues that for small states, individual national AI sovereignty is often unrealistic because of limited resources and scale. Regional cooperation therefore becomes the practical route to building shared AI capabilities.

He says countries like his do not have the large populations or investment capacity of big countries . On that basis, he argues that small countries should work together and build a regional strategy, particularly around data governance, rather than trying to build AI entirely alone .

Jovan Kurbalija frames AI primarily as a knowledge issue rather than a data issue. Because AI is built from accumulated human knowledge and shapes current knowledge production, he argues that AI divides cut across education, cognitive life and societies everywhere.

He says AI is about knowledge, not merely data, and that it is created from the knowledge of all people who lived before us, ranging from Shakespeare and Dostoevsky to African oral traditions . He adds that the people in the room are also generating knowledge now, and links this to concerns that affect people everywhere, from education and plagiarism to cognitive and social divides .

Kurbalija uses humour to comment on the event format and constant movement between podiums. The remark serves as light process management and helps maintain energy during a long session.

He jokes that there has been a lot of exercise and calls it 'healthy diplomacy' and 'gym diplomacy', thanking the organisers for the innovation in diplomacy that makes everyone move around .

In his closing reflection, Kurbalija emphasises that the central issue is not only technology but knowledge and wisdom across civilisations. He argues that AI must preserve and develop local, indigenous and historical knowledge in inclusive ways.

He refers to the Ark of the Covenant as an early codification of human wisdom and says the discussion should retain a historical sense of knowledge across civilisations in Asia, Africa, Europe and Latin America . He adds that local knowledge and the knowledge of indigenous communities are critical, and that advanced technology should be used to preserve past knowledge and develop future knowledge inclusively .

Pedro Manuel Moreno argues that the clearest way to understand the AI divide is by looking at where investment is going. He says AI infrastructure spending is vast but geographically narrow, with most investment and research concentrated in a few countries and firms, leaving poorer countries structurally behind.

He says that from an investment perspective, to understand the AI divide one should 'follow the money' . He cites figures that the world will spend close to $800 billion on AI infrastructure in the year, with cumulative spending approaching $8 trillion by the end of the decade . He adds that memory chips are sold out through 2027, companies are installing gas turbines at data centres, and strategic sectors now take half of all new investment worldwide, up from 20 per cent five years ago . He then says the geography of this boom is narrow, with two economies hosting the majority of AI and semiconductor investments, 100 companies accounting for 40 per cent of global corporate R&D, and least developed and lower middle-income countries capturing only 10 per cent of strategic sector investments .

He argues that because the global capital boom is concentrating in a handful of places, poorer countries cannot catch up without deliberate support. The barriers are cumulative: infrastructure, power, bandwidth, skills and governance all have to align for investment to come.

He explains that a data centre needs reliable power, water for cooling, high-bandwidth connectivity, technical skills and predictable data governance all at once, and if one of these is missing, the investment goes elsewhere . He also notes that least developed and lower middle-income countries have captured only 10 per cent of strategic sector investments, half their share in other industries .

AHM Bazlur Rahman argues that in places like rural and climate-vulnerable Bangladesh, the AI divide goes far beyond connectivity or compute. He says the deeper gap is institutional, linguistic, financial and political, especially because developing countries are still treated mainly as users rather than co-creators.

He asks what happens when AI reaches a community before skills, trusted information, safeguards and institutional capacity, arguing that the digital divide then becomes an AI divide . He says that for millions in Bangladesh, the divide is in language, data, skills, reliable information, institutional capacity and power, and that developing countries are still being treated mainly as users and consumers of technology . He concludes that the real gap is institutional, linguistic, financial and ultimately a power gap .

He argues that effective AI inclusion must be rooted in local ownership and local language, not external project cycles. In his view, community participation and cultural relevance are essential if AI governance is to avoid reproducing existing inequalities.

He proposes three priorities: build locally anchored AI capacity ecosystems, invest in linguistic and cultural inclusion, and create sustainable and accessible financing . He says funding should support local language AI, AI literacy, synthetic content verification, prevention of technology-facilitated gender-based violence, public interest AI and meaningful participation in AI governance . He also says access without capacity creates dependency, capacity without local ownership does not survive project cycles, and AI governance without linguistic, cultural and community participation will reproduce the divide .

Sarah Maria argues that countries do not enter the AI era on equal footing. She says the AI divide is rooted in structural differences in infrastructure, data, financing, technical expertise and institutional readiness, not merely in ambition.

She says that for developing countries such as the Philippines, the key question is whether they have the digital foundations, institutional capacity and human capital needed to participate meaningfully in the AI economy . She defines the AI divide as differences in connectivity, computing capacity, quality data, technical expertise, research systems, financing and institutional readiness, and says countries with stronger digital infrastructure are better positioned to benefit .

She argues that inclusion in AI must be broader than basic access and must deliberately reach groups often excluded from digital transformation. Her point is that these communities must be able not only to use AI but also to question, improve and govern it.

She says efforts must extend beyond major urban centres to women, youth, persons with disabilities, indigenous peoples, workers in transition, small enterprises, rural communities and geographically isolated areas . She adds that inclusion cannot be limited to access but must include the ability to use, question, improve and govern technology .

Ambassador Nahida Sobhan argues that AI divides are fundamentally development divides. She says AI should accelerate development and empower people across lines of nationality, gender, age and income rather than deepen existing inequality.

She opens by saying that bridging the AI divide is ultimately about bridging the development divide . She adds that if AI is to be a force for sustainable development, it must be accessible, inclusive and governed in a manner that upholds human rights and puts people at the centre . She further says the AI revolution should accelerate development, not deepen existing divides, and must empower all people regardless of nationality, ethnicity, gender, age or income .

She argues that affordable and lightweight open-source models can make AI more inclusive for developing countries. Such models should be trained on diverse languages, cultures and datasets from the Global South so that AI does not reproduce bias and exclusion.

She notes that most AI models are trained on data, languages and contexts from the Global North, often producing biased outcomes . She proposes investing in lightweight, affordable, resource-efficient, open-source AI models trained in diverse languages, data, culture and contexts of the Global South to democratise technology and prevent digital exclusion .

Sobhan argues that developing countries cannot close AI gaps through aspiration alone because frontier AI requires substantial compute and investment. She therefore supports a global fund and concessional mechanisms targeted at least developed countries.

She says frontier AI requires enormous computing power and investment while AI talent remains concentrated in a handful of technologies and locations . She then proposes democratising access to AI by operationalising a proposed global fund on AI and establishing concessional pathways for LDCs to access AI compute and frontier technologies .

Shaza Fatima Khawaja argues that the real divide is not simple access to AI tools but structural inequality in who shapes the technological order. She says this asymmetry concerns control over compute, models, data, standards and governance, with serious implications for global stability.

She says the divide is not in access to tools but in the ability to shape the rules, the standards and the foundational infrastructure of the technology . She describes the world as increasingly 'technopolar', where power concentrates around those controlling frontier compute, models and data, and warns that without consequential decisions most countries will be locked into the role of rule-takers .

She argues that openness and interoperability are essential for fair participation in AI. For emerging economies, open models, open data and open standards are the tools that can convert them from passive recipients into co-designers of AI systems.

She calls for emerging economies to be at the centre of designing open foundational capabilities, including models, datasets, compute access frameworks, evaluation benchmarks and safety regimes . She says interoperability and strategic openness are non-negotiable, and that open standards, open data and open models are fair strategic equalisers that let nations leapfrog while preserving sovereignty and choice .

Khawaja uses Pakistan to illustrate how an emerging economy can begin shaping AI in practice. She presents Pakistan as building public-sector capability and contributing open-source AI solutions tied to development needs.

She says Pakistan has set its own national AI direction and invested in building capabilities across the public sector . She adds that Pakistan is ready to contribute an open-source powered stack and is working on use cases in climate resilience, agriculture, inclusive public service delivery and multilingual data .

Ambassador Tsegab Kebebew Daka argues that one of the greatest risks of AI is unequal access to its benefits. He frames bridging the divide as both a fairness issue and a development imperative so that AI contributes to sustainable development rather than inequality.

He says the scientific panel's report recognises that one of the greatest global AI risks is not only misuse of AI but also unequal access to its benefits . He concludes that bridging the divide is not only a matter of fairness but essential if AI is to serve all humanity and support sustainable development .

He argues that Ethiopia is responding to the AI divide through concrete national investments in skills and institutions. The strategy is to develop talent pipelines and research capability so that the country can participate in AI on its own terms.

He says Ethiopia's Digital Ethiopia 2030 strategy includes the establishment of the Ethiopian Artificial Intelligence Institute, the launch of the 5 million Ethiopian Coders Initiative and plans for an Artificial Intelligence University . He presents these as part of building an inclusive AI ecosystem .

He argues that AI inclusion requires genuine multistakeholder and international cooperation. No single actor can provide infrastructure, skills, knowledge transfer and responsible innovation by itself.

He says no country can bridge the AI divide alone and therefore international cooperation and a genuine multi-stakeholder approach are very important . He lists governments, academia, the private sector, international organisations, civil society and the technical community as all having complementary roles in expanding access, strengthening capacity, promoting knowledge transfer and supporting responsible innovation .

He presents Ethiopia as using AI in concrete development sectors while building the institutional and digital foundations behind it. This links AI policy to public service delivery and broader digital public infrastructure.

He says Ethiopia is expanding its FIDA digital national ID system as a key component of digital public infrastructure . He also says the country is promoting AI applications in healthcare, agriculture, education and public service delivery while supporting AI solutions that reflect local language and context .

Kanono Ramashamole argues that developing countries need a set of basic enabling conditions before they can move from AI consumption to co-creation. These include affordable connectivity, reliable electricity, digital skills and better-organised public data.

He says that affordable connectivity is a basic condition, but connectivity alone is not enough, noting that Lesotho has 100 per cent connectivity while only 50 per cent of people have internet access . He adds that reliable power is critical because some people still walk long distances to charge phones . He identifies digital skills, compute and especially data as key barriers, saying public-sector data is often fragmented or unstructured .

He argues that AI capacity is multidimensional and cannot be reduced to narrow technical training in AI tools. Countries need a wider skills base spanning cyber security, data management and infrastructure operation.

He says AI does not depend on a single skill and that countries need specialists in cybersecurity, people who can manage or build data centres, and broader skills related to data .

Ramashamole argues that open source and digital public infrastructure are practical tools for stimulating local AI capability. His point is that if countries create demand and interoperable systems, compute and innovation will follow.

He says that if the conditions on data and skills are met, demand will grow and compute will become available . He adds that what has worked in Lesotho's environment is investment in digital public infrastructure and open source . As an example, he mentions participating in Open Source Week in New York, where developers learned and networked with peers from around the world .

Gilles Thonet argues that the invisible architecture of standards only works if very visible physical infrastructure exists. Reliable power, secure data centres and connectivity are therefore foundational to any AI system.

He says standards are the 'invisible infrastructure' that help products operate safely and across borders . He then stresses that this invisible infrastructure depends on physical infrastructure such as energy-efficient data centres, grid security and energy supply . He sums this up by saying that without reliable power, there is no AI .

Thonet argues that international standards are a key bridge between high-level principles and concrete technical implementation. They allow countries of any size to shape trustworthy AI by participating in consensus-based rule-making.

He describes the AI governance landscape as five normative layers: international law, soft law, national law, international standards and the technological frontier shaped largely by the private sector . He says standards connect these layers by translating principles into practical tools, giving governments a basis for regulation and industry a common language for safer and more interoperable products . He also notes that standards are developed by consensus, giving every country an equal seat at the table regardless of its economic size .

He argues that participation in standards work is not only about compliance but is itself a form of capacity building. Countries can become co-creators of AI by contributing their expertise and learning through the process of international standardisation.

He says standards are one of the most practical tools for developing countries to become not only AI users but AI co-creators . He adds that participation itself builds capacity by giving national experts knowledge they can use and adapt . He also notes that more than a third of participating countries in the IEC-ISO AI standardisation committee come from developing economies, which he presents as an important milestone .

Sid Ali Zerrouki argues that meaningful AI governance is impossible without digital foundations. If countries lack connectivity, trusted governance, cyber security and compute access, they cannot shape AI themselves and instead become dependent on others.

He says there can be no meaningful AI governance without strong digital foundations . He lists connectivity, trusted data governance, cyber security, digital skills, cloud capability, research and development and access to computing capacity as necessary conditions . He concludes that if these are missing, countries are being asked to govern a technology they cannot shape, which amounts not to inclusion but dependency .

He argues that the real challenge is not just applying AI tools but building sovereign capacity. This means being able to align AI to national priorities, assess risk, protect citizens and participate in shaping standards and governance frameworks.

He says the challenge for developing countries is not merely to use AI applications but to build sovereignty capacity . He adds that adapting national priorities, evaluating risks, protecting citizens and having standards and governance frameworks that shape the future are priorities .

Mondli Gungubele argues that inclusive AI requires action on three fronts at once: infrastructure, people and access to core AI resources. He says without these, developing countries will struggle to participate meaningfully in the AI ecosystem.

He says the first priority is digital infrastructure, including connectivity, broadband, electricity, computing resources, cloud capacity and digital public infrastructure . He identifies human capacity development as the second priority, including AI users, developers, researchers, entrepreneurs and policymakers . He then names access to data, computing resources and innovation opportunities as the third priority, warning that without them AI innovation could become concentrated in a small number of countries and corporations .

Gungubele argues that domestic effort alone is insufficient because developing countries face structural constraints in access to data, compute and financing. He therefore supports stronger international cooperation and investment arrangements to widen participation.

He says many developing nations face constraints in accessing datasets, computational infrastructure and financing needed to build local AI solutions . He then states that South Africa supports greater international cooperation, partnership, knowledge sharing and investment mechanisms that enable broader participation in the AI economy .

He uses South Africa as a national example of an inclusive AI strategy. The focus is on building enablers so that AI supports development, transformation and social progress rather than reproducing exclusion.

He says South Africa believes AI must be a force for inclusive development, economic transformation and social progress . He then outlines South Africa's priorities in digital infrastructure, AI skills across sectors and access to data and computing resources as part of its approach .

Niamh Smyth argues that governments must ensure broad access to digital infrastructure and services if AI is to support rather than deepen social divides. Her point is that citizens, communities and small businesses all need the skills and access to participate meaningfully.

She says that in Ireland the government is working to ensure all citizens, communities and small businesses have the digital skills needed to thrive, along with equitable access to technology, essential public services and high-speed broadband . She adds that Ireland's national digital and AI strategy targets socio-economic, enterprise and geographic divides through infrastructure investments, business supports and digital literacy initiatives .

Smyth argues that open-source and open-weight models can widen access to AI capabilities and support local development. She links openness to local innovation, language diversity, sovereignty and the ability of countries with weaker infrastructure to have more influence over AI.

She says countries with less developed AI infrastructure often have less influence over AI governance even when impacts on their populations are significant . She then states that open-sourced and open-weight AI models can play an important role in widening access to AI capabilities, supporting local innovation, language diversity, digital sovereignty and access to open-source AI .

She argues that AI governance must be risk-based, human-centred and shared across stakeholders. Because AI affects society broadly, no government, company, university or civil society actor can govern it effectively alone.

She says governing frameworks must evolve and that Ireland focuses on inclusive, multi-stakeholder governance models that are risk-based and human-centric and take a human rights-based approach . She adds that no single entity, whether government, tech company, academia or civil society, can alone navigate the complex societal and ethical impacts of AI .

She presents Ireland as using whole-of-government strategy to spread the benefits of digital technology and AI. The emphasis is on enterprise competitiveness, citizen empowerment and equitable service access.

She says Ireland's national digital and AI strategy targets socio-economic, enterprise and geographic divides through infrastructure investments, business supports and digital literacy . She explains that it aims to keep SMEs competitive, empower citizens to access and engage with technological change, and guarantee equitable access to essential public services .

Ambassador Subedi argues that a country such as Nepal needs more than aspiration to operationalise AI. It needs open and interoperable standards, an innovation ecosystem and support mechanisms that address its specific structural constraints as a mountainous and graduating LDC.

He says Nepal faces challenges in connectivity, infrastructure, affordability, digital literacy and access to reliable data and computing resources, especially in rural and remote areas . He then says Nepal's national AI policy treats AI as a human-centric and ethical lever for development and that what is needed most is local expertise, stronger institutions, applications aligned to national priorities, and an open, interoperable, standard and thriving innovation ecosystem .

Carlos Mendoza Alvarado argues that for countries like Guatemala, AI readiness begins with basic digital foundations rather than advanced models. He emphasises that data is a strategic public asset and must be supported by interoperability and affordability to improve state planning and decision-making.

He says that for countries like Guatemala, the starting point is connectivity, data, digital infrastructure, institutional capacities and human talent . He notes that Guatemala's national strategy recognises data as a strategic asset with three pillars: regulatory framework, interoperability and affordability . He also says that from the Planning Secretariat's perspective, data is not simply a technical input but a strategic asset for better public decision-making and better orientation of state action .

Ali bin Amer Al Shidhani argues that bridging the AI divide rests on three pillars: talent, infrastructure and cooperation. He also presents physical AI infrastructure, including trusted and green digital hubs, as a central part of national readiness.

He says the AI divide exists in model development, data, skills, computing infrastructure and power . He identifies talent as the first foundation and notes national initiatives in AI product management, data science, chip design and cloud services . He identifies infrastructure as the second foundation, explaining that Oman is developing a 'digital triangle' of interconnected multi-gigawatt digital hubs to host green AI and high-performance computing clusters . He names cooperation as the third foundation and says no country can cross the divide alone .

He offers Oman as an example of active national implementation, linking AI strategy to infrastructure development. The aim is to create a safe and trusted environment for next-generation AI systems.

He says Oman is advancing this through its National AI Program . He explains that Oman is developing a national digital triangle to host green AI and high-performance computing clusters, drawing on its political neutrality, submarine cables and green energy diversity as a 'safe space' for trusted intelligence .

Ambassador Jessica Hunter argues that trusted AI depends on underlying digital capability, not only on AI-specific tools. Her emphasis is that communities need connectivity, systems and foundational capacities first.

She says Australia views capacity building as essential to ensuring AI is not only accessible but also adopted in a trusted manner that respects human rights and delivers benefits equitably . She adds that AI capacity building must be based on understanding challenges across the AI lifecycle, including governance, skills and digital infrastructure . In the Pacific, she says Australia's work focuses on strengthening foundational digital capabilities and is underpinned by investments in connectivity .

Hunter argues that there is no one-size-fits-all approach to AI capacity building. She presents Australia's regional partnerships as examples of tailored cooperation that combines infrastructure, governance, skills and standards alignment.

She says Australia supports AI capacity building programmes across the Pacific and Southeast Asia through a focus on skills, governance and infrastructure . She cites the Australia-Pacific Digital Economy Program, partnerships with governments on public sector capacity, and work in Southeast Asia with governments, industry and academia on workforce training, AI governance and innovation .

Ahmed Hefnawi argues that openness by itself does not equal inclusion. Open systems only close divides when countries also have local inference capacity, skills and governance arrangements that allow them to use open tools safely and effectively.

He says open systems can be genuine equalisers only under two conditions: openness must be paired with governance that preserves trust, and it must be matched by domestic capacity to use what is open . He adds that an open model without local inference, structure or skill does not close a divide but simply moves the barrier from cost to capability .

He argues that capacity building should be judged by whether it creates a pathway from training to real participation in the economy. Training without deployment or income opportunities is not genuine capacity.

He says the most binding barrier is often the absence of a pathway from access to genuine participation, and that training that does not lead anywhere is not capacity but only a credential . As evidence from Egypt, he says the Digital Egypt Pioneers initiative was deliberately paired with the Industry Development Agency's GIGS programme and the Freelancer.EG platform, backed by freelancer taxation legislation, so that reskilling would connect to real income rather than end with a certificate .

Hefnawi argues that openness and safety should be designed together rather than sequentially. He supports open AI, but only if misuse safeguards and national governance rules are built in from the start.

He says international support for open AI should always come with support for the conditions of its use, and that safeguards against misuse should be designed alongside openness from the outset rather than retrofitted after harms appear . He gives Egypt's example of open-sourcing Karnak, a national Arabic large language model, while governing downstream uses such as legal assistance, language tutoring and healthcare tools under national guidelines for trustworthy and responsible AI .

He uses Egypt's practice to argue that open-source AI does not have to mean ungoverned AI. A country can make models available while still steering responsible downstream use through policy and guidelines.

He states that in Egypt, Karnak, a national Arabic large-language model, was open-sourced . He says its downstream uses in legal assistance, language tutoring and healthcare tools are governed under national guidelines for trustworthy and responsible AI, demonstrating that openness and safeguards can go together .

He presents Egypt as trying to convert AI skills training into practical labour-market outcomes. The emphasis is on linking training, policy and platforms so that participants earn income and participate economically.

He says Egypt's Digital Egypt Pioneers initiative is linked to the Industry Development Agency's GIGS programme and Freelancer.EG platform . He explains that this linkage, supported by formal freelancer taxation legislation, was designed so reskilling would connect to real income rather than stop at certification .

Khaled El-Enany argues that the international AI conversation must now move from high-level ethics principles to practical implementation. UNESCO's contribution, as he presents it, is large-scale capacity building across public administration, justice and education systems.

He says UNESCO's 194 member states adopted the first worldwide recommendation on ethics of AI in 2021, and that the task now is to turn from principles to implementation . He says UNESCO has organised training programmes and tools for more than 50,000 civil servants and judicial actors in 192 countries . He also says UNESCO has developed AI competency frameworks for teachers and students used in more than 60 countries and aims to train more than 1 million teachers by 2028 with the Varkey Foundation .

El-Enany argues that AI inclusion must be universal but also attentive to groups and regions that are frequently excluded. He explicitly centres women, youth, Africa and small island developing states in UNESCO's approach.

He says inclusion and equality are very important for UNESCO and that everyone should benefit from AI . He specifically highlights women, youth, Africa and small island developing states . He also refers to plans for a new normative instrument on youth in the era of AI and a guide for parents on teenagers, social media and AI .

He argues that a normative framework for ethical AI already exists at global level, and the current task is implementing it. For UNESCO, ethical, inclusive and safe AI depends on bridging gaps in infrastructure, skills and legal frameworks.

He says UNESCO's objective is to make AI ethical, inclusive and safe . He notes that UNESCO member states adopted in 2021 the first worldwide recommendation on the ethics of AI . He then says the challenge is to bridge multiple gaps, including connectivity, infrastructure, skills and legal frameworks, in order to implement those principles .

El-Enany argues that AI opportunities and risks outpace any single government or institution. He therefore treats international cooperation, especially within the UN family and across sectors, as indispensable.

He says the opportunities and challenges of AI outpace any institution or single government . He then says international cooperation is essential and calls it the missing link, adding that the UN family has a major responsibility to work across agencies, with governments, civil society, the private sector and scientists .

Félix Ulloa argues that education is the foundation of El Salvador's AI strategy. The country is using major public investment, devices and AI-enabled educational platforms to build the human capital needed for the digital era.

He says that under President Bukele's government, El Salvador is investing around 5 per cent of GDP in education, making it the largest item in the national budget . He explains that the country signed an agreement with Google Cloud to use Google Classroom for all students in the public system and distributed more than 1.2 million laptops and tablets to students across the country . He adds that through an agreement connected to X AI, each public-school student has an AI tutor for personalised learning .

Ulloa argues that AI deployment must rest on a clear national legal and institutional framework. El Salvador has therefore created an AI agency and a set of interlocking laws built around responsible innovation, transparency and human oversight.

He says El Salvador established a National AI Agency and built a normative ecosystem based on four laws: the AI and Technology Promotion Act, the Data Protection Act, the Cyber Security Act and the Act Promoting Robotics Technologies . He says these rest on founding principles including responsible innovation, competitiveness, transparency, human oversight and non-exclusion .

He presents El Salvador as a country already using AI across several sectors, not merely planning for it. The examples are intended to show that AI can be applied practically under a coherent legal and institutional structure.

In education, he describes the nationwide use of Google Classroom and AI tutors for public school students . In health, he explains the Dr. SV telemedicine programme, which gives citizens 24-hour access to doctors through phones or laptops and recorded 1.6 million consultations in its first six months . He adds that AI is also being used across ministries including agriculture, justice and the public prosecutor's office, and notes El Salvador's role in digital assets and Bitcoin-related regulation .

Ulloa argues that readiness assessment is useful not simply for praise but for identifying gaps that require action. In El Salvador's case, the RAM process highlighted weaknesses in rural connectivity and women's participation.

He explains that RAM stands for Readiness Assessment Methodology and is used by UNESCO and partners to assess national readiness . He says El Salvador received a strong score but that the more important part was the recommendations . He highlights identified weaknesses such as the rural-urban gap in the use of AI tools and the need for greater women's empowerment, especially among rural women . He describes practical responses including local white-space routers and the rollout of Starlink to help close the connectivity gap .

Shikoh Gitau argues that AI capacity building should not be reduced to narrow user training on popular tools. She wants a broader approach focused on creating AI locally, with investments in skills, jobs, standards, open systems, language and cultural context.

She says capacity building in AI should not just be about how to use AI and ChatGPT . She argues that it needs a coordinated approach to building and creating AI 'for us, by us', with investments in skills, jobs, standards, open systems and sustainable financing . She also stresses inclusion of Africa's thousands of languages and says that beyond language, culture and context must be included .

Gitau argues that creating AI that works for diverse communities requires not only skills and standards but also financing and South-South collaboration. She links this directly to the need for linguistically and culturally grounded AI systems.

She says capacity building should include sustainable financing and explicitly mentions South-South cooperation . She ties this to the need for all languages, cultures and contexts to be included in AI systems .

Gitau introduces the first panel as a response to a major governance question: how to distribute AI's benefits globally and locally. She deliberately emphasises communities as well as countries.

She says the session tackles one of the most consequential questions in global AI governance: how to ensure AI's transformative potential reaches all countries and communities . She underscores communities in particular, noting Africa's communal-led economy .

At the end of the first panel, Gitau distils the discussion into a set of recurring themes. Her synthesis shows that the panel converged on structural enablers, local relevance and the social foundations of trust.

She summarises the panel by saying 'follow the money', and that connectivity, power, data and skills are important . She adds that context decides outcomes and that purpose-driven, locally aware AI matters . She also highlights standardisation, equitable access, trust-building over time and starting with grassroots actors .

Diera Gala Paksi argues that AI capacity building must be ecosystemic and human-centred. Infrastructure matters, but durable readiness depends on preparing the people and institutions around AI, including educators, policymakers and communities.

She says AI capacity building is not simply about increasing AI literacy but about building an ecosystem that enables government, educators, industry, academia and communities . She then lists four essential conditions from ASEAN's perspective, beginning with human readiness, which should focus not only on infrastructure but also on educators, students, policymakers and communities .

She argues that local partners are indispensable because they understand local culture and implementation realities better than regional bodies alone. Sustainable programmes therefore need local organisations and government ownership from the start.

She says local organisations understand cultural context better than regional organisations . She notes that ASEAN is working with 23 civil society organisations across the region . She also says capacity building becomes sustainable only when governments are involved from the beginning, which is why ASEAN convenes policy roundtables linking policymakers and local implementing partners .

Paksi argues that sustainable AI readiness needs both state ownership and broad partnerships. She presents regional and multistakeholder cooperation as what enables large-scale and lasting impact.

She says one essential condition is government ownership, because capacity building becomes sustainable only when governments are involved from the outset . She also recommends regional multi-stakeholder partnerships combining government, academia, industry and civil society to deliver locally relevant AI capacity building . She illustrates scale by saying the ASEAN AI Ready programme has empowered millions of people across the region .

She uses the ASEAN AI Ready programme as a practical example of broad-based AI capacity building. The programme is intended to reach multiple segments of society and combine awareness, training, policy discussion and research.

She says she serves as project manager for the AI Ready ASEAN programme, one of the largest AI initiatives across ASEAN . She explains that it aims to empower millions of individuals including youth, parents and educators through awareness raising, platforms, training of trainers, policy roundtables and research . She later says the initiative has already empowered 8 million individuals across ASEAN .

Alessandra Sala argues that workplace AI training should not be limited to efficiency gains or labour replacement. Instead, it should help workers rethink their roles, use AI to scale value creation and shape a more empowering future of work.

She describes leading AI transformation at Shutterstock and says one of its pillars is literacy among staff, while her work with Women in AI focuses on bringing more women into core AI skills so they can be at the table . She recounts telling employees not to use AI simply to save time or risk replacement, but to use the right tools to reinvent workflows . She gives a concrete example of helping a salesperson move from contacting 50 customers a day to 500 with personalised offers and customer analysis, showing AI as a way to shift from reactive to proactive work rather than merely cut time .

Sala argues that one of the most achievable and necessary actions is investment in education and literacy, especially for groups often excluded from technology transitions. She stresses that this education should be about creating value and redesigning processes, not merely extracting answers from AI systems.

In the rapid closing round, she calls for investment in upskilling and literacy for women and children and says they need to be brought along . She adds that this should not be 'vanilla education' aimed only at getting answers from AI systems, but education that helps people reinvent processes and create value .

Ambassador Vladimir Cuc argues that capacity constraints are widespread, including in Europe, and that the response must focus on institution-building rather than declarations alone. He sees skills, stronger public institutions and digital foundations as central to making AI more transformative and less disruptive.

He says governance gaps, resource gaps and capacity gaps create disparities and unequal access, and that no country can deal with this alone . He notes that Moldova's UNESCO Readiness Assessment found an ecosystem that is developing but lacks capacity in the public sector, private sector and academia . He therefore says capacity building and investment should be at the core of AI dialogue and cooperation so that AI becomes less disruptive and more transformative .

He argues that AI governance must be anchored in human rights, democracy and rule of law, but also recognise the practical role of private firms. Public-private partnership is therefore necessary because companies are already central actors in AI development.

He says Moldova advocated for the Council of Europe's framework convention on democracy, human rights, rule of law and AI, and adheres to UNESCO's recommendation on the ethics of AI . He states that Moldova is aligning its legislation with the EU AI Act and will create an agency . He concludes that one of the key takeaways should be investment in public-private partnerships because private companies are currently shaping AI while governments regulate it .

The Smart City representative argues that training-focused approaches to digital and AI divides are incomplete because they ignore what happens afterwards. Real capacity requires labour-market demand, public procurement pathways and access to underlying assets such as compute and data.

The speaker says Smart City has trained over 400,000 Filipinos across more than 60 local government units in the Philippines . Based on this experience, the speaker identifies three gaps not captured by 'pipeline thinking': absorption into labour markets and procurement systems, a seat at the table in governance and standard-setting, and access to compute, data and institutional negotiating power . The speaker concludes by calling for capacity building to be tied to jobs, procurement and enterprise pathways rather than only counting completions .

Diego Belevan argues that many countries do not lack values or principles but the means to operationalise them. He therefore calls for cooperation that is horizontal and mutually beneficial, centred on access to infrastructure, tools, information and technical support rather than one-way transfer.

He says that without real capacity, agreed principles cannot materialise and that the problem for many countries is not the absence of shared values but the impossibility of operationalising them . He argues that cooperation is most valuable when it complements the normative layer with concrete access to infrastructure, tools, information and technical assistance . He adds that the logic should be reciprocal benefit and horizontal cooperation, not unilateral or directional transfer of predefined models .

Emma Theofelus argues that technical capacity in the Global South requires deliberate investment in young people and academic collaboration. She sees twinning arrangements between universities as a practical route to scaling training.

She says one of the immediate priorities is high technical skills and academia, especially twinning agreements between universities in the Global North and Global South to train as many young people as possible across the Global South .

She argues that AI governance must remain human-centred and interoperable, while also holding large technology companies accountable. She specifically links accountability to harms such as technology-facilitated gender-based violence.

She says Namibia is preparing its national AI governance framework around human-centred governance, targeted public-sector innovation, public-private and global collaboration, equitable infrastructure and interoperable rules and shared standards . She then stresses the need for greater accountability geared towards megatech companies, especially where technologically facilitated gender-based violence is concerned .

Theofelus argues that inclusion requires access to lighter and more usable AI models as well as open standards. She links this openness to both broadening participation and improving accountability.

She says one of the issues that must be addressed now is access, which includes lean models, open standards and greater accountability towards major tech companies .

Girmaw Abebe Tadesse argues that technology alone does not determine whether AI helps people; the decisive factor is context. For AI to deliver fair outcomes, systems must be built in ways that are purpose-driven, locally aware and shaped by the people and cultures they are meant to serve.

He says their report contains the line that for AI to be successful, context is important, and indeed that 'context decides outcome' . He argues that positive impact depends not only on technology but on who accesses AI, how it is adopted into institutions and how it diffuses across sectors . He invokes Wangari Maathai to warn that even doing the right thing without context can take one off track . He adds that having 7,000 languages does not mean technology will serve them equally, so AI must be purpose-driven and locally aware . Finally, he says understanding context requires bringing in governments, private actors, youth and grassroots communities and putting people at the centre of design .

Arutyun Avetisyan argues that AI should eventually allow all people to speak and be understood in their native languages, including speakers of small languages. At the same time, common standards must still leave room for different national cultures and values.

He says he hopes AI will eventually make it possible for everyone to speak their native language and still understand each other, and that even smaller languages should be covered by AI . Later he says that while countries may work towards common standards, each country has its own culture and must solve some issues itself, taking account of cultural specificities and richness .

He argues that trust and security have been underappreciated in AI governance and need to be addressed through shared standards and practical safeguards. However, those common standards must still respect cultural diversity across countries.

He says scientists have been saying for several years that trust is more important than many other issues, and he highlights security challenges including deepfakes, elections and social impacts . He says countries cannot solve this alone and proposes that open models and code should be equally accessible to students and universities worldwide . He also calls for institutes and standardised work together on security, noting that simple internet tools are not enough without adequate security and knowledge . He further suggests that AI systems will need special labels and equitable access regardless of company or country size .

Avetisyan argues that the complexity of trustworthy AI means states need shared resources and cooperation. Open access to models and code, coupled with common work on security, is part of making co-creation possible across countries.

He says it has been demonstrated that countries cannot work through isolated campaigns and need to work together . He proposes understanding at UN level whether open models and open code should be equally accessible to any student and any university in the world . He then argues that one internet tool is only one solution and that a whole range of security and deployment issues requires collective and standardised work .

Huanzhang Fu argues that law enforcement agencies need more than access to AI tools. They need AI literacy, governance frameworks and confidence-building so that use of AI respects the rule of law and public trust.

He says AI literacy and readiness among law enforcement agencies vary considerably across countries, making the AI divide a capacity-building imperative . He states that effective AI adoption begins with knowledge, trust and responsible governance and that technology alone is not enough . He adds that agencies need the skills to understand AI's opportunities and limitations and governance frameworks to deploy it responsibly while upholding the rule of law and public trust .

He uses Interpol's work to show how responsible AI adoption can be operationalised in a sector-specific way. The emphasis is on practical guidance, training and international cooperation for police forces.

He says Interpol, with UN partners, UNICRI and the UK, published a revised toolkit on responsible AI innovation in law enforcement in 2024 . He adds that Interpol and UNICRI are implementing an EU-funded AI project and launched the TRAIL e-learning programme, which is freely accessible to law enforcement in all member countries . He also notes webinars and in-person training delivered to officers in several regional groupings to translate principles into operational practice .

Valts Ernštreits argues that most of the world's linguistic and cultural communities remain structurally excluded from AI. He says these communities need representation, dedicated methodology and control over their own data if AI is to support rather than marginalise them.

He says the report shows that of 7,000 languages, only 1,000 have the foundations for meaningful inclusion in AI systems, leaving 6,000 communities without them . He adds that many of these communities are indigenous or speak severely endangered languages . He argues that AI for these communities is methodologically distinct because their data ecosystems and usage patterns differ from mainstream ones . He then proposes three responses: communities must be at the table, there must be dedicated international and national research and resources, and data and technologies must be governed so that AI is fit for purpose, non-extractive and supportive of communities' rights and ways of life .

Urvashi Aneja argues that a neglected but crucial divide is the oversight gap. Countries in the Global South need their own evaluation, procurement, reporting and monitoring capacity so that AI systems reflect local values rather than simply importing frameworks developed elsewhere.

She says the gap that is not often discussed is the oversight gap, and that it is widening as AI systems become more complex . She says this area is routinely underinvested in by governments, funders and development agencies in the Global South . She identifies needs at the state level such as evaluation, procurement standards, incident reporting and continuous monitoring . She argues that domestic evaluation capacity and infrastructure are the way countries can ensure AI systems promote local values and priorities, and warns that without civil society and independent media, countries end up borrowing frameworks from the Global North instead of building their own .

Aneja argues that the AI discussion often overlooks the infrastructure of oversight that makes accountability possible. She sees underinvestment in public institutions, civil society and media as a major reason why AI can expand without adequate checks.

She says oversight is underinvested in by governments, funders and development agencies . She lists procurement standards, incident reporting, continuous monitoring and evaluation as examples of state-level gaps . She also says civil society and independent media are the checks and balances that make oversight real, and where they cannot flourish there is no public literacy or accountability ecosystem .

She argues that open, locally usable evaluation infrastructure is one of the most practical levers available to build agency in AI governance. This would allow non-technical actors and communities to influence how AI systems are assessed and aligned with their goals.

In the final round, she calls for support to open and local evaluation infrastructure that is easy for non-technical people to use . She says this is how local communities can participate in shaping the values and goals of AI systems and that it offers a way to manage inevitable dependency on more equitable terms .

Diego Rodrigo Beleván Tamayo argues that openness can help reduce AI divides, but only if countries have the institutional capability to use what is made available. Otherwise, the gap simply shifts from access to practical use.

He says open source, open models and open AI could significantly contribute to bridging divides provided their adoption is accompanied by developing institutional capacity . He warns that otherwise the gap is merely transferred from access to a gap for use .

Katharina Frey argues that a practical way to bridge AI divides is to connect demand for AI resources with existing global supply. She presents networks like ICANN as mechanisms that can link researchers and institutions to compute, datasets and expertise for locally relevant work.

She says that what is needed is access to powerful compute, qualitative datasets and domain as well as AI experts, and notes that ICANN's members are academic partners . She explains that ICANN connects demand and offer . As an example, she says Data Science Africa was linked to Swiss supercomputing centres so researchers could work on AI weather forecasting systems for farming communities when they lacked enough compute . She also describes a newer programme connecting UN organisations in Geneva with researchers from ICANN's network across Africa, Europe and Asia because of strong demand for co-created AI solutions .

David Tshere argues that broad calls for partnership are insufficient for countries that lack the infrastructure needed to benefit from AI. He wants the dialogue to produce a specific, actionable global funding mechanism for developing countries.

He says he had already spoken earlier about how developing countries are left behind and hears a recurring message that they should partner with others and 'figure it out' . He then argues that the dialogue should come up with an idea for a global AI fund for developing countries because infrastructure remains the major issue in Africa and without it countries cannot harness AI's benefits .

Eugenio Vargas Garcia argues that digital sovereignty should not mean isolation or complete self-sufficiency. Instead, it should mean gradually reducing dependency, diversifying partners and increasing the political agency of developing countries within the global digital ecosystem.

He says digital sovereignty is important because it allows countries to make sovereign choices about investing in, adopting or playing a proactive role in AI development . He stresses that it should not be seen as isolationism or the attempt to produce everything locally . Rather, it is a continuous process of reducing structural dependencies and expanding political agency . He says this requires diversified partnerships and suppliers, selective national alternatives, international cooperation and stronger participation by developing countries in international discussions .

Garcia argues that the international community already has many ideas on AI capacity building and now needs implementation and coordination. He highlights the new global network of centres and suggests linking UN dialogue more directly to other summit processes to avoid duplication.

He says capacity building is not new for the UN and that existing resolutions and examples already indicate what should be done, so the problem is implementation . He points to the newly established global network of centres for exchange and cooperation on AI capacity building as a useful voluntary mechanism where institutions and labs can pool resources and courses . He also suggests connecting the Bletchley process and its summits with the UN global dialogue so the Geneva and New York tracks are not treated as standalone efforts .