Lowy Institute

Three outcomes would make this Dialogue meaningful rather than merely procedural. First, developing countries and small island states must also shape the outcome, not endorse one written without them. Governance frameworks for AI have so far been drafted primarily by technologically advanced nations and large private sector actors. If this Dialogue produces another framework in that mould, it will have replicated the problem it was convened to address. Second, the Dialogue must connect AI governance to the conditions that make it real. Such as, in much of Southeast Asia and the Pacific, fewer than 40 percent of households have reliable internet access, and costs routinely exceed the UN affordability benchmark. Governance frameworks that do not address this are frameworks for the already-connected. A concrete outcome would be an agreed commitment that AI governance strategies must include baseline access assessments and financing targets. Third, the Dialogue must take seriously what ordinary users are already experiencing. People encounter AI-generated fake images, fabricated video, and false information daily – in messaging apps, social media, and news feeds – with no reliable means of identifying what is real. This is eroding trust in digital systems broadly. The Dialogue should agree on concrete standards for synthetic content disclosure and commit to investing in user-facing verification tools and literacy programmes. These three outcomes – inclusive process, equity-linked commitments, and user-centred safeguards – form a coherent foundation. Each addresses a distinct failure mode of AI governance as currently practised.

• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

Existing programmes are not reaching the people who need them most. Rural communities, women, low-income, and Indigenous peoples face the steepest barriers and are consistently underserved by initiatives that tend to reach urban, formally-employed populations. Beyond individual skills, there is a critical shortage of institutional capacity - governments often lack the technical expertise to regulate, procure, and audit AI systems, leaving them dependent on the companies they are meant to oversee. Capacity-building must be embedded in national strategies, locally tailored, resourced through development finance, and measured by outcomes rather than inputs. AI systems built on English-dominant data produce concrete harms in linguistically diverse regions: mistranslation, cultural misrepresentation, and exclusion from services affecting healthcare, finance, and education. Economically, productivity gains from AI are currently concentrated among large, data-rich enterprises; micro and small businesses - which represent the majority of employment in the developing nations - remain largely outside the frame. Governance frameworks must be evaluated against these distributional effects, not only their technical specifications. Users interacting with AI systems are frequently unaware of how those systems affect consequential decisions about them such as in credit assessments, recruitment, content visibility, and access to services. They have no meaningful way to challenge outcomes they cannot see. Human rights protections must be grounded in the user's position: the right to know when AI is used, the right to an explanation of consequential decisions, and the right to effective remedy - regardless of location, language, social origin, or income. Transparency must be operational, not declaratory. It should require disclosure of training data, known limitations, and decision logic for high-stakes systems. Accountability must include accessible redress mechanisms for affected individuals, not only institutional-level regulatory frameworks. Oversight requires regulators with technical knowledge - so that scrutiny reflects how systems actually work, not how developers describe them.

The proliferation of fake images, fabricated video, and AI-written content is generating a harm that existing governance frameworks are not designed to address: the progressive breakdown of shared epistemic ground within communities. People make decisions - about health, about who to trust, about what is happening around them - based on what they believe to be true. When fabricated content is indistinguishable from real content, and circulates through the same channels as authentic information, it distorts those decisions at scale. The absence of trusted technology standards for content authentication means that users currently have no reliable, accessible means of distinguishing real from synthetic and no institution they can turn to when fabricated content causes them harm. Selecting and assessing technology providers requires technical expertise that most developing nations currently lack, creating an asymmetry between those deploying AI and those responsible for governing it. Establishing criteria for trusted technology providers is an emerging priority in international AI governance, but no agreed multilateral framework yet exists. Developing nations and Least Developed Countries are increasingly receiving AI/ICT infrastructure through bilateral financing arrangements. These arrangements carry conditions, often implicit, that shape recipient nations' choices on standards, procurement, regulation, and long-term vendor dependency. The Dialogue should examine these dependency relationships and develop multilateral norms and practical guidance - including on provider assessment, procurement criteria, and contract conditionality - that protect nations' ability to make independent governance decisions.

For Southeast Asia and the Pacific island countries, these governance gaps are not future risks, they are present realities. On connectivity, national AI strategies are being written while large parts of the population remain offline. In several Pacific island states, fewer than half of households have reliable internet access. You cannot govern what people cannot reach, and you cannot include people who are not connected. On institutional capacity, most governments in the region are being asked to regulate, procure, and audit AI systems they do not yet have the technical expertise to evaluate independently. It reflects how fast AI has moved relative to how fast public institutions can build specialised knowledge. In practice, it means regulatory decisions often rely on what vendors say about their own systems. That is a structural problem, not an isolated one. On technology financing and provider dependency, the region is already living with the consequences of infrastructure arrangements made without full awareness of their long-term implications. Connectivity systems, cloud platforms, and e-government tools built under bilateral financing agreements have created technical and contractual dependencies that now constrain what governments can choose. For Least Developed Countries in the Pacific, where alternative financing is scarce, the negotiating position is particularly narrow. On language and cultural exclusion, Southeast Asia and the Pacific are among the most linguistically diverse regions on earth. Most of that diversity is invisible to the AI systems being deployed here. Communities interacting with AI-mediated healthcare, education, and financial services regularly encounter systems that do not understand them. When AI systems cannot serve a community, that community also cannot meaningfully participate in governing them.

The AI Dialogue has genuine potential – but whether it advances international cooperation or simply adds to an already crowded landscape of declarations depends entirely on how it is structured and who it genuinely includes. The most valuable role the Dialogue can play is one that existing mechanisms have not yet filled: bridging the gap between global norm-setting and the conditions facing developing countries and small island states. International cooperation on AI governance has so far produced principles that are broadly agreed and poorly implemented. Smaller nations lack the capacity to translate global frameworks into domestic policy, and global frameworks are rarely designed with their constraints in mind. The Dialogue can address this in three ways. First, by ensuring that nations from Southeast Asia, the Pacific, and other underrepresented regions participate as co-authors of outcomes, not as endorsers of positions formed elsewhere. Second, by producing cooperation mechanisms that are implementation-oriented – focused on financing, technical assistance, and institutional capacity. Third, by establishing a regular accountability mechanism: a process through which commitments made at the Dialogue are tracked, reported on, and revisited.

Several existing mechanisms have laid important groundwork that the AI Dialogue should build on rather than duplicate. The UN Secretary-General's AI Advisory Body report identified governance fragmentation and the exclusion of developing nations as the central weaknesses of current international AI governance. The Dialogue should treat its findings as a baseline and demonstrate concrete progress against them, rather than revisiting the same diagnosis. The ITU's AI for Good platform has built a practical community of practice around AI deployment in developing contexts. The Dialogue should draw on ITU's regional presence and technical expertise, particularly for implementation support in Least Developed Countries and small island developing states. The G20 AI Principles established early international consensus on responsible AI, but the G20 represents neither the majority of the world's nations nor the populations most affected by AI governance gaps. The Dialogue's added value lies precisely here – it can extend the conversation to actors and regions that G20 processes structurally exclude. The added value of the AI Dialogue is not in producing new principles. It is in creating an inclusive process that connects existing frameworks to the nations that need implementation support most, coordinates financing and technical assistance, and establishes accountability for commitments already made.

The UN system has developed strong principles for inclusive multilateral processes. The question for the AI Dialogue is whether those principles are reflected in its design from the outset – not added later as a correction. A more effective structure starts with regional preparatory processes. Delegations from Southeast Asia, the Pacific, and other underrepresented regions need space to develop consolidated positions before the global meeting, not during it. Smaller nations are consistently outpaced in plenary settings by better-resourced delegations. Regional caucuses with real outputs address this directly – and are already an established feature of UN preparatory processes. Within the Dialogue itself, working sessions should be built around specific unresolved questions rather than general themes. Participants should be responding to each other, not delivering positions at each other. This is a small design change with significant consequences for what gets said and decided. Civil society and community representatives must have roles, not courtesy slots. This requires funded participation, translated materials, and adequate preparation time. Their involvement in standard-setting processes should be subject to transparent disclosure of commercial interests – consistent with conflict of interest frameworks applied in other UN technical bodies.

The Leaving No One Behind principle, central to the 2030 Agenda, applies as directly to AI governance as it does to any other development priority. The people most affected by AI governance decisions are the least represented in the conversations that produce them. This is not a peripheral problem – it directly shapes the quality and legitimacy of whatever the Dialogue produces. Communities from Least Developed Countries and small island states face practical barriers that are rarely acknowledged honestly: the cost of travel, visa difficulties, the capacity to prepare technical positions, and the reality that most global forums operate in English. Fixing this requires money, logistical support, and a genuine willingness to accept contributions in forms other than an English-language statement delivered in a plenary room. Indigenous peoples and local communities are almost entirely absent from AI governance forums. Inclusion means designing processes that their own decision-making protocols can actually engage with, not simply extending an invitation. Women from rural and marginalised communities, speakers of low-resource languages, and young people all face versions of the same structural problem: the forum was not designed with them in mind.

Many of the formats that would make the AI Dialogue more effective are already used within the UN system. The ask is that they be applied here as standard practice, not exception, and designed from the outset with inclusion in mind. Sessions built around specific unresolved problems – how to assess technology providers, how to set synthetic media standards, how to structure financing conditionality – produce more than sessions built around general themes.