Leaders TalkX 3 - Artificial Intelligence : Policy, Ethics, and Global Cooperation

Chaichanok Chidchob argues that while financial and technological transfers are welcome, what developing countries most urgently need is access to knowledge, experience, and governance frameworks. This is because there is a vast disparity in understanding of AI across nations, and closing that gap is foundational to equitable participation. Sharing governance models enables emerging economies to catch up and engage meaningfully in building AI together.

Chidchob noted that a co-chair at a previous event highlighted how very few people in the room fully understood the software, hardware, or technology behind AI, which underscored the enormous knowledge disparity in the industry . She explicitly stated that what the international community should prioritise is sharing - not just funds and technology, but knowledge, experience, and governance models and frameworks, so that everyone can reach a level playing field .

Chidchob highlights that Thailand has emerged as a potential AI hub for ASEAN due to its geographic location, strong connectivity infrastructure, and competitive energy prices. Prior to recent years, Thailand was largely unaware of this potential. She invites the international community to study Thailand's evolving AI governance experience as a real-world sandbox.

Chidchob noted that Thailand is currently being discussed as the focal point of the next potential AI hub of ASEAN , and that the country had previously been unaware of its potential stemming from its geolocation, strong connectivity infrastructure, and electricity prices . She formally offered Thailand's situation as a sandbox for AI governance and invited all participants to study from what Thailand is going through, framing it as a shared responsibility for emerging economies to catch up .

Mavetera argues that it is not sufficient for developing countries to simply adopt AI technologies created elsewhere; they must be participants in shaping AI's development trajectory. This means moving beyond passive consumption to producing AI solutions and contributing to the governance of the technology. Zimbabwe's approach reflects this ambition through its national strategy and innovation initiatives.

Mavetera explicitly stated that developing countries must not be merely consumers of AI but also producers, and that they need to be participants in shaping their own development . She also referenced Zimbabwe's vision of a human-centric, inclusive, and trusted AI ecosystem that aids sustainable development in an ethical, trustworthy, and secure manner .

Mavetera outlines the suite of policy instruments Zimbabwe has developed to prepare for AI adoption. These documents collectively address strategy, security, child protection, ethics, and accountability. Together they represent a comprehensive policy framework designed to embed AI responsibly into Zimbabwe's national development.

Zimbabwe launched its national AI strategy as a key step towards embedding AI as a country , and also launched a cybersecurity strategy and a child online protection policy as necessary preparatory measures . Additionally, Zimbabwe developed an AI charter covering ethics and considerations, as well as a mandatory algorithmic impact assessment .

Mavetera describes Zimbabwe's practical efforts to foster homegrown AI innovation through the creation of sandboxes for testing startups and a presidential challenge to build capacity among ICT entrepreneurs. These initiatives are designed to cultivate local talent and ensure Zimbabwe is future-ready. The emphasis on talent development reflects a belief that human capital is the most critical resource for AI readiness.

Zimbabwe created sandboxes to enhance and test startups, enabling a controlled environment for innovation . The country also launched a Presidential AI Innovations Challenge specifically to capacitate ICT startups, with a belief that talent is what needs to be developed to prepare for the future .

Zeledón reframes the challenge facing developing countries as an 'agency divide' rather than merely a digital divide. He argues that the critical question is not whether countries use AI, but whether they can shape the models, data architectures, standards, and markets that determine their development opportunities. True inclusion means having the capability to adapt AI to local needs, evaluate risks, protect rights, and retain public value.

Zeledón stated that the central question is no longer whether developing countries will use AI, since it is already happening, but whether they will help shape the models, data architectures, standards, and markets that will determine their development opportunities . He defined the agency divide as the distance between using AI and having the capacity to adapt it to local needs, evaluate risks, protect rights, develop solutions, and retain public value .

Zeledón presents Costa Rica's national AI strategy as a comprehensive framework that integrates ethical values with practical implementation priorities. The strategy aligns with internationally recognised standards from the OECD and UNESCO while addressing domestic needs such as talent development and digital infrastructure. It represents Costa Rica's commitment to moving from principles to measurable action.

Costa Rica's national AI strategy 2024-2027 combines human dignity, transparency, oversight, equity, security, and sustainability with priorities such as talent, productive adoption, digital infrastructure, research, public sector transformation, and alignment with the OECD AI principles and UNESCO's recommendations on the ethics of AI .

Zeledón highlights the Community Innovation Labs (LINCS) as a concrete mechanism for distributing AI capability across Costa Rican society, including beyond urban centres. The initiative has reached thousands of people, with a particular focus on women's digital empowerment. This approach reflects the principle that AI agency must be broadly distributed rather than concentrated in elite or metropolitan settings.

More than 7,500 people have been trained through the LINCS network, more than 6,000 women have received AI and digital skills training, and nearly 5,000 public officials have received AI-related capacity building . Costa Rica is extending training and experimentation spaces beyond metropolitan areas through this initiative .

Ozols challenges the assumption that countries must develop large frontier AI models to gain meaningful benefits from AI. He argues that high-quality, domain-specific data can yield results comparable to large-scale models, giving smaller countries a viable path to AI agency. Latvia's stroke detection platform serves as a concrete example of this principle in practice.

Ozols cited the Synapse Hospital Stroke Detection Platform, developed by Latvia's national hospital and a local company, which achieves a stroke detection rate of 93%, comparable to industry-standard tools of similar scale . He argued that quality and specificity of data outperforms quantity, suggesting many countries have this agency and need not be dependent on large-scale models .

Ozols describes two concrete steps Latvia has taken to move from AI consumer to contributor. The first is a civic initiative in which citizens voluntarily recorded their voices to train open speech models that are then distributed. The second is a National AI Sandbox that allows pre-testing of high-risk AI systems in partnership with developers and deployers, building regulatory competence through real-life experience.

Latvia established a civic initiative where citizens donated their voices to be used for training open speech models, which were then distributed . The National AI Center developed a National AI Sandbox that allows pre-testing of high-risk systems in partnership with deployers and developers, enabling regulators to gain real-life experience in controlled settings before setting guardrails .

Ozols highlights that the dominance of large language models is exacerbating linguistic inequality, with smaller and underrepresented languages receiving progressively less representation. This is not merely a technical issue but a governance failure, as current governance models are creating and widening these gaps. Targeted policy responses are needed to protect linguistic diversity in AI systems.

Ozols noted that the gap between dominant and underrepresented languages in large language models is widening, and that this gap is being created by the current governance model . He framed this as a particular concern for smaller languages, which are disproportionately affected by the concentration of risks in frontier-scale AI systems .

Ozols reveals that when Latvia was building its EU open model, disinformation networks had silently contaminated the training data with distorted facts, which then propagated into downstream models and products. This illustrates the serious data integrity risks facing AI development, particularly for smaller countries. Resolving the issue required collaboration with other countries and researchers to clean the corrupted data.

Ozols described how, during the construction of Latvia's EU open model, millions of articles planted by disinformation networks had silently poisoned the training data, distorting facts that then impacted downstream models and services . Latvia had to partner with other countries and researchers to clear out the corrupted data .

Ozols advocates for the creation and sharing of open AI assets — including language models and speech datasets — among countries as a strategic response to the dominance of large proprietary models. This approach enables smaller nations to contribute to and benefit from AI development without being entirely dependent on a handful of powerful actors. It also serves as a mechanism for protecting linguistic and cultural diversity in AI systems.

Ozols identified building open assets and sharing them among member states and countries, such as open language models, as the first of three moves Latvia recommends for becoming a contributor rather than a user . The civic voice-recording initiative, where citizens donated their voices for open speech model training, was cited as a practical example of this approach .

Ozols argues that smaller nations can amplify their weight in global AI governance by strategically joining coalitions and multilateral bodies rather than acting alone. Participation in structures such as the EU AI Board and UN bodies provides leverage that individual countries cannot achieve independently. Sharing open assets and institutional expertise further strengthens this collective approach.

Ozols noted that as a compact country, Latvia can gain leverage by joining coalitions that multiply its weight, citing the EU AI Board and UN bodies as examples of such multipliers . He concluded that no country can address AI governance alone and that those with something to contribute must put it on the table so that evidence, standards, and oversight are shared and coordinated globally .

Shehail advises institutions to resist the temptation to deploy broad autonomous agentic AI systems and instead begin with clearly defined, limited workflows. Starting narrow allows organisations to learn, test, and improve without exposing themselves to unnecessary risk. This disciplined approach creates the conditions for responsible scaling over time.

Shehail stated that organisations getting real value from agentic AI are not beginning with open-ended autonomous agents but with defined workflows where success can be measured and failure can be contained . She noted that a bounded use case gives the institution space to learn, test, and improve without exposing the whole organisation to unnecessary risk .

Shehail emphasises that wherever a decision carries significant consequences — legal, financial, operational, or safety-related — a human must remain responsible for the final outcome. Agentic AI systems can assist and support decision-making processes, but they should not be permitted to replace human accountability. Each institution must define what constitutes a consequential decision in its own context.

Shehail stated that if a decision has legal, financial, operational, or safety implications, the system can support the process but a human should remain responsible for the final decision . She added that each institution needs to define clearly what a consequential decision means in its own context .

Shehail argues that observability and reversibility are not optional features to be added later but must be built into agentic AI systems from the very beginning. Every action taken by an agent should be logged and traceable so that institutions can understand what happened and why. Where possible, actions should also be reversible to allow for correction when errors occur.

Shehail recommended that every action should be logged, traceable, and, where possible, reversible . She gave examples such as an agent sending an email, updating a record, or triggering a process, noting that the institution should know what happened, why it happened, and how to correct it if needed .

Shehail warns against several common pitfalls in agentic AI adoption. Granting broad autonomy based on controlled demonstrations is dangerous because production environments are far more complex and unpredictable. Treating agentic AI primarily as a cost-cutting tool leads to underinvestment in oversight, and the security risks introduced by agents with system access or code execution capabilities are frequently underestimated.

Shehail cautioned that a demo is controlled while production is not, and that real environments include incomplete data, unusual cases, changing conditions, and real consequences . She warned against treating agentic AI as a headcount reduction strategy from day one, as this leads institutions to underinvest in oversight, redesign, training, and change management . She also highlighted that an agent that can browse email, access a system, or execute code introduces security risks that traditional controls may not fully address .

Shehail's overarching message is that agentic AI represents a genuine capability shift, but its benefits will only be realised by institutions that approach it as a governance and organisational challenge rather than purely a technical one. Expanding the autonomy of AI agents must be earned incrementally through demonstrated trustworthiness and evidence-based evaluation. Moving deliberately and measuring honestly are prerequisites for responsible scaling.

Shehail stated that the institutions that benefit from agentic AI will be those that treat it as a governance and change management challenge as much as a technology one . She concluded with the message to move deliberately, measure honestly, and expand autonomy only when trust has been earned through evidence .

Kleiner argues that the IGF must take an role in shaping how agentic AI develops in relation to the internet, treating it as an extension of existing internet architecture rather than a separate or disruptive force. Key issues such as standards, identity management, trust mechanisms, and communication parameters for agents need to be addressed within the multistakeholder community. Failing to do so risks allowing agentic AI to develop in ways that undermine the open internet.

Kleiner stated that the Internet Governance Forum should take a role in shaping the way we look at agentic AI for the future . He raised specific questions about what standards are available, how to treat agent identity in ways that build trust and avoid spam-like problems, and how to ensure discovery tools and communication parameters make agents efficient and valuable .

Kleiner advocates for the use of IGF Policy Labs as a dedicated forum where diverse stakeholders can proactively examine how AI — and agentic AI in particular — may affect internet governance. The goal is to anticipate problems before they manifest as fragmentation or exclusion rather than reacting after the fact. This requires genuine engagement from the full range of multistakeholder actors.

Kleiner expressed the EU's wish for the issue of agentic AI to come to centre stage in IGF discussions and not be a forgotten item, and hoped it would be addressed through the policy labs together . He noted that the EU recognises a lot of work remains to be done to engage with the multistakeholder model and use the new policy labs feature of the IGF to explore current issues .

Kleiner stresses that the open, unified internet is a shared global resource that must be actively protected as AI develops. He warns that agentic AI has the potential to become autonomous from the internet in ways that could fragment it, and that this fragmentation would destroy the common fabric of digital opportunity for all nations. Preventing this outcome is a collective responsibility of the international community.

Kleiner noted that there is an acceleration in the potential of agentic AI to become autonomous from the internet , and argued that there is a collective responsibility to avoid a situation where the one internet we have built becomes fragmented, as this would completely undermine the common fabric and common opportunities for the whole world .

Behrendt argues that the current pace of international cooperation on AI is insufficient given how rapidly the technology is advancing. Nations risk being overtaken by developments they have not collectively prepared for if they do not accelerate their collaborative efforts. This urgency applies equally to regulation, capability building, and joint model development.

Behrendt stated that nations need to find ways to cooperate in a very speedy way because they are a little bit too slow in their cooperation, and that they cannot afford to be slow in the future .

Behrendt emphasises that effective AI governance cannot be limited to addressing AI as it currently exists; leaders must engage in forward-looking scenario planning about what AI will look like in the near future. This kind of anticipatory thinking is too complex and consequential for any single nation to undertake in isolation. It requires collective effort and shared imagination across borders.

Behrendt referenced Joshua Benjo's call for a 'mind exercise', arguing that serious AI strategies must consider what AI will look like in two, three, four, or five years rather than only its current state . She stated that this exercise must be done together and in scenarios, and that no single nation can do it alone .

Behrendt argues that Germany and its partners must strengthen their own AI capabilities to reduce dependence on a small number of private companies, predominantly based in the United States. This does not mean pursuing complete independence, but rather building capacity and developing AI models in collaboration with partners who share democratic values and commitments to human rights. The initiative is framed as open to global partners, not exclusively European.

Behrendt noted that there is a strong dependency on private companies mainly situated in the USA, and that while complete independence is neither possible nor the goal, countries must boost their own capabilities and create their own models . She described the goal as creating AI models made in Europe or with trusted partners who share values of democracy and human rights, and invited others to join this initiative .

Sinha argues that AI's transformative potential can only be realised equitably if the foundational resources — compute, data, skills, and applications — are made broadly accessible rather than concentrated among a small number of countries and companies. This is framed as the first of three core priorities for India's AI policy. Without deliberate action to extend access, AI risks deepening existing inequalities.

Sinha stated that AI must not remain limited to a few countries, companies, or communities, and that compute, data, skills, and applications must be available to the Global South and to underserved populations .

Sinha outlines India's three-pillar approach to AI policy: ensuring broad access, building trust through transparency and accountability, and delivering real-world impact for citizens. The India AI Mission operationalises these priorities by investing in compute infrastructure, data quality, indigenous AI development, and startup support. This framework reflects India's ambition to be both a responsible and an contributor to global AI development.

Sinha identified three priorities for India's AI policy: access (AI must not remain limited to a few), trust (AI systems must be transparent, accountable, privacy-preserving, secure, and fair), and impact (AI must solve real problems for farmers, students, patients, women, small enterprises, and citizens) . The India AI Mission is described as building an ecosystem that demonstrates AI compute, improves data quality, supports indigenous AI capabilities, promotes startups, and advances safe and trusted AI .

Sinha highlights India's digital public infrastructure journey as a model for translating connectivity into concrete development outcomes at the grassroots level. The Samridhgram Vegetable Village project demonstrates how AI-enabled connectivity can deliver services across multiple sectors — healthcare, education, agriculture, e-governance, skilling, and livelihoods — directly to village communities. The project's recognition as a WSIS champion further validates this approach.

Sinha mentioned the Samridhgram Vegetable Village initiative, where connectivity has been converted into healthcare, education, agriculture, e-governance, skilling, and livelihood services at the village level . He noted that this project has been selected as a champion project for WSIS 2026 prizes .

Sinha argues that no country can govern AI alone and that effective global AI governance requires a comprehensive multilateral framework. This framework must encompass shared principles, interoperable standards, responsible data governance, capacity building, and AI safety testing. Crucially, it must also ensure that developing countries participate equitably rather than being marginalised in global decision-making.

Sinha stated that no country can govern AI alone and that shared principles, interoperable standards, responsible data governance, capacity building, AI safety testing, and equitable participation of developing countries are all needed . He also noted that India stands ready to work with ITU, the UN system, and global partners to build an AI future that is ethical, inclusive, trusted, and truly beneficial for humanity .

Sinha challenges the common assumption that ethical constraints on AI necessarily impede innovation. He argues instead that ethical guardrails generate the trust that is a prerequisite for innovation to scale effectively and sustainably. India's New Delhi Declaration on AI embodies this integrated approach, linking economic growth, social good, responsible innovation, and international cooperation.

Sinha stated that ethics and innovation are not competing objectives, and that ethical guardrails create trust, which in turn enables innovation at scale . He referenced India's New Delhi Declaration on AI Impact as reflecting this approach by emphasising AI for economic growth, social good, responsible innovation, and international cooperation .

Chami frames the central challenge of the session as one of collective governance: whether the international community will govern AI's transformation together or allow it to govern them. She emphasises that the speed of AI's impact on daily life, the information space, and the global economy makes timely action imperative. The session is presented as an opportunity to examine what action is needed now.

Chami noted that AI is no longer a distant horizon but is already transforming daily life, the information space, and the global economy at breathtaking speed . She posed the question of whether the international community will govern this transformation together or let it govern them , and framed the session as examining the nature of action needed before technology outpaces our ability to govern it .

Chami identifies the core normative challenge for global AI governance as ensuring that the technology's benefits are distributed equitably, with particular attention to developing economies. She frames this as a question of governance, ethics, and global inequality, not merely a technical matter. This framing sets the agenda for the panel discussion.

Chami asked what the international community's priority should be to ensure AI remains inclusive, trustworthy, and benefits all countries, particularly developing economies, noting that AI has enormous potential to accelerate sustainable development but also raises challenges in governance, ethics, and global inequality .

Chami poses a question that reframes the development challenge from one of mere access to one of genuine agency and capability. She asks how global AI governance can ensure that developing countries are not just recipients of AI but shapers of AI systems, governance of risks, and beneficiaries of public value. This reflects a more sophisticated understanding of digital inclusion.

Chami asked how global AI governance can move beyond access and ensure that developing countries build the agency, capacity, and partnerships needed to shape AI systems, govern risks, and capture public value .

Chami highlights Latvia as an instructive case study in the limits of digital advancement as a guarantee of AI agency. Even as an EU member with strong digital infrastructure, Latvia continues to struggle for linguistic representation in AI systems, demonstrating that the gap between consuming and contributing to AI is not simply a matter of development status. She asks whether the mechanisms being built are designed to genuinely close this gap or merely manage it.

Chami described Latvia as a digitally advanced EU member that continues to fight for its own language to gain representation in AI systems like many others, and asked from that vantage point what it actually takes to move from consumer to contributor, and whether the mechanisms being built are designed to close that gap or merely to manage it .

Chami characterises the UAE's contribution as offering a responsible AI perspective on agentic AI governance, and acknowledges that policymaking in this area is necessarily iterative and evolving. This framing validates the approach of building governance frameworks incrementally rather than waiting for comprehensive regulation before acting. It also signals the importance of practical, institution-level governance alongside national and international policy.

Chami thanked the UAE speaker for reflections on a responsible AI perspective to the governance of agentic AI systems and policymaking as a work in progress .

Chami explicitly links the discussion of AI governance to Internet governance, framing them as inseparable issues that must be addressed together. She acknowledges the reflections on openness in AI before transitioning to the European Commission's contribution on the IGF and AI. This framing reinforces the importance of treating AI not as a standalone governance challenge but as an extension of existing internet governance concerns.

Chami noted that the session was moving from reflections on openness and innovation in AI to another important issue, and introduced the European Commission's contribution in the context of the WSIS Plus 20 review's recognition of the importance of strengthening the IGF and improving the follow-up of its outcomes .