University of Exeter, School of Law

• It is important to first underscore that the process itself has intrinsic value. This is because fostering global, inclusive dialogue, even where outcomes are incremental, is the Dialogue's goal - and this is the first time all nations and stakeholders will have a chance to come together to discuss global AI governance. That said, a modest outcome that would make the Global Dialogue a success would be the publication of annual reports summarising discussions and identifying areas of consensus and topics requiring further work. Beyond that, other markers of success include: 1) the implementation of a Global Fund for AI, in line with UN General Assembly resolution A/79/966 (particularly by fostering capacity-building commitments and contributions from developed countries and industry representatives)
• 2) the establishment of a global AI incident response mechanism, including a unified database of AI incidents and a list of national AI incident points of contact
• 3) consensus on the application of international law, including international human rights and humanitarian law, to AI technologies, and a report fleshing our how key rules/obligations apply
• 4) the adoption of new international rules to fill gaps in the application of existing international law (including imposing obligations on AI developers and deployers and agreeing on when strict state or corporate liability is warranted)
• 5) failing agreement on binding rules, the adoption of norms of responsible state and corporate development and deployment of AI.

• Safe, secure and trustworthy AI
• AI capacity-building
• Protection and promotion of human rights
• Interoperability of governance approaches

To address the most urgent gaps in global AI governance, we propose five priority clusters for discussion in July: a. AI risks and harms, including in the military domain. This should cover immediate or short-term harms to individuals, such as bias, discrimination, and manipulation; broader societal risks, such as threats to democracy, employment, and the environment; and longer-term or existential risks. The Dialogue can play a critical role in bridging different (and often conflicting) perspectives on which risks to prioritise, underscoring that efforts to mitigate these categories of risks are not mutually exclusive. b. Technical and policy responses to these risks. This includes measures to prevent harms before they arise (e.g., compute governance and legal prohibitions), mitigate harms that cannot be avoided (e.g., AI alignment, interpretability, and control), and strengthen resilience and response mechanisms when harms occur (such as critical infrastructure redundancy and PPE acquisition ahead of biosecurity risks). c. Confidence-building measures, such as verification strategies, global incident response systems, national points of contact, and shared incident databases, building on existing initiatives such as the OECD's incident database and the national cybersecurity points of contact. d. Capacity-building, not only for AI adoption and innovation, but also for safety, security, and governance. In this respect, the Global Dialogue should take an active part in supporting the implementation of General Assembly resolution A/79/966 on 'Innovative voluntary financing options for artificial intelligence capacity-building', particularly by fostering capacity-building commitments and contributions from developed countries and industry representatives. e. Norms and international law, including the extent to which existing frameworks, particularly international human rights and humanitarian law, already apply to AI, and whether/where new rules, norms, or institutions are needed to ensure AI risks are minimised, and benefits shared equitably.

Even though I am based in Canada and the UK, I am originally from Brazil, a developing country. So I would like to highlight that AI impacts tend to be amplified in the Global South for a number of reasons, including reduced AI safety and security capacity, lower levels of AI literacy, and societal resilience. For example, deep fakes and AI-generated hate speech can much more easily sway people. Likewise, the adoption of AI models and systems trained primarily on data originating from developed countries can lead to inaccurate and biased outputs. And the fact that AI evaluations and other assurance methods have been primarily designed by stakeholders in the Global North means that these risks can go undetected for years. Therefore, when we talk about closing the global AI divide, it is important to not only build the capacity of developing countries to innovate with AI but also to develop appropriate AI safety, security, and transparency mechanisms.

The Global Dialogue marks a historic opportunity to make AI governance a more democratic endeavour with the participation of all nations and all relevant stakeholders across industry, academia, and civil society. Therefore, it can facilitate dialogue and coordination between key actors based in the Global North - in particular, AI developers - and governments and stakeholders - in particular, civil society - in the Global South. We are witnessing an era of unprecedented corporate economic and political power, with billions of dollars being invested in AI labs. A few tech companies concentrate this power. Therefore, any progress in international cooperation will require bringing these companies to the negotiating table as well as holding them to account for irresponsible behaviour. As mentioned earlier, the Global Dialogue can also foster multistakeholder consensus or common understandings on appropriate mechanisms to prevent, mitigate, and respond to a range of AI risks and harms in both developed and developing nations, as well as areas that require further research. It can build confidence among nations by fostering transparency around AI incidents and devising mechanisms for appropriate responses. And it can help states agree on how existing international law applies to AI, as well as which new binding rules or non-binging norms are needed to fill accountability gaps.

First and foremost, the Dialogue should build on and align with existing UN processes that touch on different aspects of AI governance, such as the Group of Governmental Experts on Lethal Autonomous Weapons Systems and the OEWG on ICTs, ensuring coherence and avoiding duplication while strengthening ongoing work. Furthermore, the Global Dialogue should also connect with the AI Summit series and the OECD's efforts (including the Hiroshima Reporting framework and the Global Partnership on AI).

Drawing on our experience with other UN multistakeholder processes, including the Open-Ended Working Group (OEWG) focused on the security of and in the use of information and communications technologies (ICTs), we urge that any accreditation process be based on fair and transparent criteria, free from Member State veto. Beyond virtual consultations, there should be a dedicated in-person multistakeholder meeting held during formal sessions of the Global Dialogue, one that allows stakeholders to make oral submissions to Member States and engage formally and informally with them, including through side events. The value of multilateral processes lies precisely in those conversations that can only happen when state representatives and stakeholders are in the same room. An extensive number of participants joined and expressed an interest in speaking at the first round of virtual informal consultations this week. This signals the importance of the topic: AI's impacts span society as a whole. While it is important to give everyone an opportunity to speak and be heard at these consultations, it will be practically difficult, if not impossible, to give every single interested stakeholder a speaking slot. Therefore, the selection of stakeholders for oral interventions in virtual and in-person sessions should not be random or arbitrary, but based on fair and transparent criteria, such as equitable geographical, gender, sectoral, and subject-matter balance. During formal sessions of the Global Dialogue, expert briefings on thematic clusters should also be built into the programme to inform discussions and deliberations.

It is clear that stakeholders from the Global South have been excluded from mainstream global discussions on AI governance, such as the AI Summits and the OECD. To include, it is imperative that they benefit from funding to attend the in-person sessions of the Global Dialogue as well as long-term capacity-building efforts, including financial contributions and knowledge-sharing from Global North governments and institutions.

Virtual consultations should be longer to accommodate the number of interested parties.