High-level multi-stakeholder plenary segment: Harnessing the benefits of AI for all through inclusive and interoperable approaches - part 2

Bosun Tijani argues that AI governance must move beyond abstract principles and become practical national systems that countries can actually implement and assess. He says this requires country strategies, measurement, capable public institutions, trust in governance bodies, and sustained investment in infrastructure and talent.

He explicitly says the first step is to move away from governance principles into practical systems because AI is already affecting everyday life and disrupting sectors and business opportunities . He adds that each country needs clear strategies for AI adoption and for measuring what is being governed . He further identifies public-sector capacity, trusted institutions, impact measurement, and investment in compute and talent as necessary conditions for meaningful implementation , .

Tijani stresses that governance capacity is not only about training engineers or AI developers. Governments also need officials who understand the technology and can keep pace with rapid developments in order to implement governance frameworks effectively.

He asks how to ensure that government officials responsible for implementation have the understanding needed and can keep up with developments in the field . He then states that capacity building must cover not just talent to build AI, but also public-sector capacity, which he describes as extremely important .

Tijani presents trust as a foundational requirement for AI governance. In his view, even well-designed rules will fail if the institutions responsible for implementing them are not trusted by the public and stakeholders.

He says there will be a need for trusted institutions because governance systems cannot really function if people do not trust the institutions implementing the governance framework . He presents this point alongside political leadership and impact measurement as part of the practical conditions needed for effective governance .

Tijani argues that countries cannot benefit from or govern AI well unless they have the basic capacity to absorb and use it. This includes connectivity, infrastructure, and the ability to participate in AI development rather than remaining passive consumers.

He says governance must be linked to nations’ absorptive capacity to put AI to good use . He explains that where connectivity is insufficient, AI application is limited and people’s ability to govern it is also weakened . He adds that lack of infrastructure constrains contextual AI development and appropriate governance, while countries that can only consume rather than develop AI face clear limitations .

In his closing remarks, Tijani frames inclusion as meaningful access to AI benefits, not mere access to imported tools. He argues that countries need the infrastructure and capabilities to build AI suited to their own contexts if inclusion is to be real.

He says inclusion should centre on access to the benefits of AI and that such access will not come from merely adopting what has been built elsewhere . He then argues that nations need capabilities, especially infrastructure, so they can build for context, which he sees as essential to realising AI’s benefits .

Tijani warns that unequal access to core AI infrastructure, especially compute, could create a deeper divide than the internet divide. He calls for practical clarity and investment in shared compute and other prerequisites so that more countries can participate meaningfully.

He says the UN platform can help clarify access to shared compute within regions and across countries . He also argues that nations need critical prerequisites in place and warns that if investment and support for infrastructure are not mobilised, the AI gap will become larger than the divide seen during the emergence of the internet .

Marwala argues that the biggest implementation gap is fragmentation across multiple AI governance systems. Because international bodies, states, and industry are developing separate approaches, actors can exploit weaker jurisdictions, making harmonisation and practical coordination essential.

He defines the problem as 'AI governance arbitrage', explaining that governance models are emerging at the international level, in industry, and among member states, but they are not coordinated . He illustrates this by saying that someone interested in data will go to a country where regulations are weaker, and that the same logic applies to compute capability . He concludes that a mechanism is needed to harmonise these approaches and that institutions must be brought together not only for dialogue but also at the working level , .

Marwala sees education as the central mechanism for making AI governance work. He argues that broad literacy is needed across technical subjects, standards, policy, and lawmaking so that both practitioners and legislators can act responsibly and effectively.

He repeatedly says that people need education about data, algorithms, computing, and how AI can be applied . He adds that standards should be brought into classrooms so that people designing systems are exposed to them early . He also stresses the need to educate legislators so they can craft sensible AI legislation .

Marwala argues that the next step for global AI governance is practical action rather than repeated discussion. He also frames governance as a balancing exercise, where institutions must help decide how to manage trade-offs between competing values and technical choices.

He says plainly that he wants to see a move from talk to action, especially around taking technology to the Global South and democratising access . He then explains that governance is a balancing problem between opportunity seeking and risk aversion . As examples, he raises the need to determine the balance between transparency and security, and between synthetic and authentic data use .

Yi Zeng argues that agreement on ethical principles has not translated into dependable implementation. He says testing shows that current language models often fail when ethics and safety are examined rigorously and in varied real-world forms.

He says there is already broad consensus on AI ethical principles, citing UNESCO’s ethics recommendation and the advisory body’s report . He explains that when high-level principles were translated into about 90 finer considerations and tested, performance dropped sharply when the same question was asked in many different ways, from around 98 to 58 . He also reports safety testing results showing that even the best systems could be successfully attacked 60% of the time and the worst 40% of the time, indicating serious practical weaknesses .

Yi Zeng argues that global dialogue must deliver tangible outcomes, not just discussion. He says ethics and safety should define red lines and steer AI development in a constructive direction, rather than being treated as anti-innovation constraints.

In his closing remarks, he says the dialogue needs outputs and that AI development definitely requires red lines if ethics and safety are being taken seriously . He then argues that ethics and safety are not a backlash against AI development, but instead help identify the right direction and support healthy, steady progress .

Yi Zeng argues that better public understanding of AI is essential for responsible governance. He wants scientists to communicate the actual capabilities and limitations of current systems so that society does not mistake them for human-like intelligence.

He says the general public needs higher literacy and that scientists should help the world understand the current status of AI more clearly . He emphasises that there is 'no magic' in current systems and describes AI as a machine that processes information without truly understanding it .

Yi Zeng argues that systematic testing reveals serious weaknesses in present-day AI systems. Because models can appear strong on simplified tests yet fail under more robust questioning and adversarial challenge, he sees stronger evaluation as urgently necessary.

He describes a testing method in which around 10 high-level ethical principles were broken into 90 finer considerations and then posed to language models . He reports that although models appeared to score very highly at first, their performance fell to 58 when the same question was reframed in 100 different ways . He also cites adversarial safety testing showing successful attacks against the best systems 60% of the time and against the weakest 40% of the time .

Whitney Baird presents multistakeholder inclusion as a core premise of AI governance. She argues that because both the promise and the risks of AI affect many actors, effective governance requires all stakeholders to be involved rather than leaving decision-making to any single country or sector.

At the outset, she says that with the promise and risks of AI development and implementation, everyone who has a stake in the process needs to be at the table . In her closing summary, she reinforces this by stating that AI governance cannot be achieved by any single stakeholder or country and will depend on sustained collaboration across governments, industry, academia, civil society, and the international community .

Baird argues that AI oversight cannot be left solely to technical experts or machines. She stresses the importance of human judgement, critical thinking, and broad social perspectives, including from non-engineers and the public, in deciding how AI should be deployed and trusted.

After Yi Zeng’s remarks, she highlights the critical need for humans to remain in the 'chain of command' . She adds that society needs not only engineers but also philosophers and others who can bring the broadest possible perspective to questions of deployment, use of information, and trust .

Baird frames evaluation as a practical mechanism for making AI governance more trustworthy. In her moderation, she links evaluation with greater certainty for stakeholders and presents trust as a recurring benchmark for whether governance systems will be accepted.

After Rumman Chowdhury’s intervention, she says evaluation is a way to build trust and build certainty . Earlier, she also highlighted trust as enormously important in response to Bosun Tijani’s comments, showing that she sees it as a central governance concern .

Rumman Chowdhury argues that existing global AI governance has tended to consolidate the influence of actors who already hold power. She believes the UN dialogue offers a chance to reset this pattern by intentionally bringing in Global South and global majority perspectives.

She says that while AI still desperately needs global governance, what she has seen is a form of governance built on consolidating people who are already powerful . She then argues that these dialogues create an opportunity to reset the direction and specifically calls for concrete engagement with Global South and global majority countries .

Chowdhury supports AI sovereignty, but rejects the idea that sovereignty should lead to fragmentation or disengagement. She argues that countries should maintain ownership over AI development while continuing to cooperate internationally, especially in a tense geopolitical climate.

She says she wants to see a discussion of sovereignty that does not lose the concept of cooperation . She notes that current geopolitics feel fraught and that some international cooperative agreements appear to be rolling back . She concludes that while she strongly supports sovereignty and ownership of AI models and modalities, countries should still engage in international cooperation rather than closing doors to one another .

Chowdhury argues that genuine AI sovereignty rests on people, not just infrastructure or models. She says only humans can interpret context and decide what counts as acceptable or successful AI performance in society.

She defines true AI sovereignty as a combination of good governance and domain expertise, which requires cultivating human capital . She then says only humans can understand context and define model performance, because people must shape how AI is used in the world .

Chowdhury argues that the most important power in AI governance is the power to evaluate systems. For her, evaluation is where societies decide normative questions about whether systems are appropriate, beneficial, or harmful in real-world use.

She states directly that the power of AI lies not in development but in evaluation . She explains this by asking what greater power there is than defining what is good and bad and right and wrong, linking evaluation to the social authority to judge AI outcomes .

Chowdhury proposes concrete institutional steps for scaling public participation and accountability in AI governance. She argues that governments should create independent evaluators, localise standards, and use procurement rules to ensure AI providers meet evaluation requirements.

She says that as countries invest in data centres, models, and talent, the evaluations layer and the infrastructure needed to scale individual perspectives are often overlooked . She then recommends three policy steps: creating independent evaluation bodies, operationalising evaluation standards from regional and domain perspectives, and linking evaluation requirements to procurement so government and social-sector contracts go only to compliant organisations .

Kate Kallot argues that prevailing ideas about what AI should look like are being set by too few actors from a narrow part of the world. She wants the UN to challenge this by elevating different models, experiences, and evidence from the Global South and global majority.

She says one desired outcome is a place where the AI narrative is deconstructed and reconstructed because it has been driven by a very specific part of the world and a very specific set of actors . She argues that these actors have promoted the idea that AI must mean large general-purpose models and gigawatt-scale factories, which she rejects . She adds that more stories and evidence from the Global South and global majority need to be surfaced .

Kallot argues that AI governance should empower the people building and using systems locally, not just elites in global forums. She says domestic developers, companies, and communities will determine how AI evolves in practice and therefore need visibility and support.

She calls for support for companies and builders who are trying to advance AI in their own countries and help their societies develop on their own terms . She also asks for platforms where they can surface their innovation and appear on stages like the UN dialogue, arguing that they will be the real guardians of what AI becomes in the coming years .

Kallot argues that consultation after the fact is insufficient and that communities should be genuine co-designers of AI systems. She says local users need to shape problems, use cases, and edge cases before coding and deployment begin.

She says Amini operates on the principle that consultation is not fit for purpose and that true sovereignty requires local ownership . She concludes that the most effective approach is to bring communities in at the design stage, before code is written or systems deployed, so they can define the problem, use cases, edge cases, and contribute local lived experience .

Kallot argues that meaningful AI sovereignty requires keeping critical state data under domestic control. She says countries should not have to choose between accessing powerful AI tools and retaining jurisdiction over sensitive information.

Using Barbados as an example, she says Amini digitised millions of government records, including bilateral agreements and cabinet papers dating back to 2018 . She explains that this intelligent knowledge base remains in Barbados under government-controlled infrastructure and jurisdiction, allowing ministers and civil servants to use frontier model capabilities without sending critical data outside the country .

Kallot argues that capacity building must be contractual and practical, not rhetorical. She believes external providers should train local government staff and engineers over time so that systems can ultimately be handed over and operated independently.

She says that in every government contract, Amini includes a transfer obligation rather than just offering a help desk or customer service . She explains that in-country teams train data stewards within government and engineers on operational pillars over a two- to three-year period, after which systems are handed over when the local team is ready .

Kallot argues that local ownership should include the ability to extend and build upon AI systems. Open-source public AI tools can enable domestic developers to contribute local knowledge, create new services, and retain more value within the country.

She says that in Barbados, Amini open-sourced 'ChatBB', an AI assistant aimed at lowering barriers to public services . She explains that because the system is open source, Barbadian developers can build on top of it, while training programmes for local youth and data scientists allow them to contribute new applications and services that bring more local context into the system and generate economic value locally .