Future Train Ai

The first Global Dialogue on AI Governance would be a success if it moves beyond declarations and produces concrete, actionable commitments that countries and institutions are actually held to. Too often these forums generate momentum that fades without follow-through. Specifically, success would mean establishing a shared baseline understanding of AI risks that is accepted across geopolitical divides, not just among like-minded nations. It would mean ensuring that the Global South has genuine influence over the outcomes, not just a seat at the table while decisions are shaped elsewhere. A successful dialogue would also produce clear mechanisms for ongoing coordination, so this is not a one-time event but the foundation of a sustained governance architecture. This includes agreeing on how to handle rapidly evolving capabilities, where governance frameworks constantly risk being outpaced by the technology itself. Finally, success would mean centering human rights and dignity as non-negotiable principles, ensuring that AI governance does not become a race to the lowest common denominator driven by competitive pressures between major powers. The legitimacy of whatever emerges depends on whether people around the world, not just governments and corporations, feel that their interests were genuinely considered.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight

These four areas reflect what I see as the most urgent and interconnected challenges in AI governance today. Safe, secure and trustworthy AI is foundational. Without it, nothing else matters. AI systems that cannot be trusted undermine public confidence and create risks that disproportionately affect those with the least power to push back. AI capacity-building is equally critical because the governance conversation cannot be dominated by a handful of technologically advanced nations. If developing countries lack the technical expertise and infrastructure to understand, deploy and regulate AI, they will be subject to decisions made elsewhere that shape their societies without their meaningful input. The social, economic, ethical, cultural, linguistic and technical implications of AI are often treated as secondary to technical standards discussions, but they are where the real human impact is felt. AI systems that do not account for linguistic and cultural diversity, or that accelerate economic inequality, will cause lasting harm even if they are technically well-designed. Finally, transparency, accountability and human oversight are the mechanisms that make all other governance commitments credible. Rules without accountability are aspirational at best. Human oversight is particularly important at this stage of AI development, where we do not yet have the tools to fully verify how these systems behave in complex real-world conditions. Together these priorities reflect a view that good AI governance must be technically grounded, globally inclusive and genuinely accountable.

Yes, there are several cross-cutting issues that deserve explicit attention. The first is the concentration of AI power in a small number of private companies. Most of the listed themes assume a governance relationship primarily between states, but the reality is that a handful of corporations hold more influence over AI development than most governments. Any serious governance framework needs to address this power asymmetry directly. The second is the environmental impact of AI. The computational infrastructure required to train and run large AI systems consumes enormous amounts of energy and water. This is rarely discussed in governance forums but has direct implications for climate commitments and for communities living near data centre infrastructure. The third is AI's role in information ecosystems and democratic integrity. The capacity of AI to generate and amplify disinformation at scale poses a distinct risk that cuts across almost every other governance challenge. It affects elections, public health, conflict situations and social cohesion. It warrants its own dedicated focus rather than being folded into broader ethical implications. Finally, there is the question of enforcement and compliance. The listed themes address what governance should cover but say little about how commitments will be monitored and enforced across jurisdictions. Without a serious conversation about accountability mechanisms, including what happens when states or companies fail to comply, even well-designed frameworks risk becoming aspirational documents with limited real-world impact. These issues are not peripheral. They are structural, and leaving them unaddressed would be a significant gap in the dialogue's ambition.

From a UK and broader European perspective, the governance gaps in my four priority areas are creating real and measurable challenges. On safe, secure and trustworthy AI, the UK finds itself in an awkward position post-Brexit. It has chosen a lighter-touch, pro-innovation regulatory approach that diverges from the EU AI Act, creating friction for organisations operating across both jurisdictions and raising questions about whether UK standards will be seen as credible internationally. The opportunity here is for the UK to use its convening power and strong research base to help bridge transatlantic and global governance approaches rather than being caught between them. On capacity-building, the UK has significant assets in its universities and AI research institutions, but the benefits are not evenly distributed domestically or globally. There is a risk that capacity-building efforts remain extractive, drawing talent from the Global South without genuine technology transfer or institutional strengthening in those regions. On social, economic and cultural implications, the UK is already experiencing AI-driven labour market disruption in sectors from legal services to creative industries. The creative sector in particular has been vocal about the use of its work to train AI systems without consent or compensation, and this remains an unresolved governance gap with significant cultural and economic consequences. On transparency and accountability, the UK lacks a dedicated AI enforcement body with real teeth. Existing regulators are adapting their mandates to cover AI but without the resources or legal clarity to do so effectively. Human oversight of high-stakes AI decisions in public services, including welfare, immigration and criminal justice, remains inconsistent and poorly monitored. The opportunity is for the UK to lead by example, but that requires more ambition than is currently on display.

The AI Dialogue has a genuinely important role to play, but only if it is designed with that ambition from the outset rather than defaulting to the usual pattern of high-level statements followed by limited follow-through. Its most valuable contribution would be to serve as a legitimate space where the full diversity of national perspectives can be heard and taken seriously. Most existing AI governance initiatives have been shaped primarily by the US, EU and China, with other regions consulted at the margins. The Dialogue, sitting within the UN system, has the legitimacy to change that dynamic if it chooses to use it. Practically, it can advance cooperation in several ways. It can help map the existing landscape of national and regional AI governance frameworks, identifying where genuine convergence is possible and where differences reflect legitimate value choices rather than mere technical disagreement. This kind of structured comparison is necessary groundwork that is currently missing. It can also establish shared principles for how AI governance frameworks should interact across borders, particularly on issues like data flows, conformity assessment and mutual recognition. Companies and civil society organisations currently navigate a patchwork of overlapping and sometimes contradictory requirements, and this imposes the greatest costs on smaller actors in less resourced contexts. Perhaps most importantly, the Dialogue can create a standing forum and working relationship between nations that does not currently exist in any coherent form on AI. The technology will continue to evolve rapidly, and having an established channel for multilateral conversation is itself valuable, independent of any specific outcomes. The risk to avoid is that it becomes a talking shop. Concrete next steps and a clear follow-up mechanism should be agreed before the Dialogue concludes.

There is no shortage of existing initiatives, and the Dialogue's first task should be to take stock of them honestly rather than duplicating effort or creating yet another parallel process. The OECD AI Policy Observatory and the OECD AI Principles have provided a useful baseline, but their membership skews heavily towards wealthier nations. The Global Partnership on AI was a promising attempt to broaden participation but has struggled to translate dialogue into concrete outcomes. The EU AI Act represents the most comprehensive binding framework to date and will inevitably shape global norms through regulatory gravity, whether other countries choose to engage with it or not. The Bletchley process and subsequent AI Safety Summits have raised the profile of frontier AI risks but have been criticised for being too focused on catastrophic scenarios at the expense of near-term harms that are already affecting people. Within the UN system, UNESCO's Recommendation on the Ethics of AI and ITU's work on AI standards are relevant touchpoints. The AI Advisory Body's report provides a reasonable starting framework for what multilateral AI governance could look like. The added value the Dialogue can bring is something none of these initiatives has fully achieved: genuine universality combined with a mandate to move from principles to practice. Most existing frameworks are either inclusive but non-binding, or binding but exclusive. The Dialogue sits within a system that has the legitimacy to bridge that gap. It should also explicitly connect with civil society, academia and affected communities in ways that intergovernmental processes typically fail to do. That kind of structured inclusion would itself be a meaningful innovation in how AI governance is conducted.

The format and structure of the Dialogue will determine whether it is genuinely inclusive or merely appears to be. A few recommendations based on what tends to work and what tends to fail in these processes. First, the Dialogue should not be structured primarily around formal plenary sessions dominated by government delegations. These formats systematically advantage well-resourced states and large organisations that can afford to send representatives. Substantive thematic working sessions, with dedicated space for civil society, academia, technical experts and affected communities, should sit at the core of the programme rather than at the margins. Second, there needs to be meaningful pre-consultation at the regional level before the main Dialogue convenes. Views from Africa, Latin America, South and Southeast Asia are too often aggregated and flattened in global forums. Regional preparatory processes would allow more nuanced perspectives to be developed and carried forward with proper representation. Third, the language and accessibility of participation matters enormously. If documentation, sessions and deliberations are conducted primarily in English, that structurally excludes enormous constituencies. Interpretation and translated materials should be treated as non-negotiable rather than optional add-ons. Fourth, private sector participation should be structured carefully. Industry input is valuable and necessary, but tech companies should not be positioned as neutral technical experts. Their interests are specific and significant, and the format should reflect that clearly. Finally, the Dialogue should create explicit mechanisms for follow-up engagement, not just a one-time consultation. Stakeholders who contribute should be able to see how their input has been considered and responded to. Without that feedback loop, participation feels performative, and the same voices will disengage from future processes as a result.

The underrepresentation in global AI governance discussions is stark and systematic, and it goes beyond the obvious geographic imbalances. The most underrepresented are communities in the Global South who are already experiencing the consequences of AI deployment without having had any meaningful input into how those systems are designed or governed. This includes workers subjected to algorithmic management, communities affected by AI-enabled surveillance, and populations whose languages and cultural contexts are largely absent from the training data that shapes these systems. Indigenous communities deserve specific mention. Their knowledge systems, data sovereignty concerns and relationship to technology are almost entirely absent from mainstream governance conversations, yet AI systems increasingly interact with and affect their lives and their heritage. People with disabilities are another group whose perspectives are rarely centred. They are often positioned as beneficiaries of AI rather than as agents with distinct views on how it should be governed, what risks it poses and what safeguards they need. Young people, particularly from the Global South, will live longest with the consequences of the governance decisions being made now, but their structured participation in these forums is minimal. On the question of how to include them: funding and logistical support for participation is necessary but not sufficient. The more important shift is structural. These communities need to be involved in setting agendas and shaping outcomes, not just invited to comment on frameworks that have already been drafted by others. Civil society organisations working directly with affected communities should be given formal roles in the Dialogue, not just observer status. And the Dialogue should commission perspectives from underrepresented regions specifically, rather than waiting for submissions that well-resourced actors are better positioned to make.

The traditional conference format, panels of experts speaking at audiences followed by brief Q&A, is poorly suited to the complexity and stakes of AI governance. If the Dialogue wants genuine engagement rather than performance, it needs to be deliberately designed differently. A few formats worth considering: Deliberative dialogue sessions, modelled on citizens' assemblies, where mixed groups of participants from different backgrounds and regions work through specific governance questions together, guided by neutral facilitators. These create conditions for actual position change and mutual understanding rather than the restatement of prepared positions. Red team exercises, where participants are asked to identify the weaknesses and failure modes of proposed governance frameworks rather than simply debating their merits. This tends to surface more honest and substantive engagement than conventional debate formats. Problem-specific working groups that produce concrete draft outputs during the Dialogue itself, rather than feeding into a process that will produce recommendations months later. When participants know their work will result in something tangible, engagement quality improves significantly. Structured stakeholder exchanges that pair government representatives directly with civil society organisations or affected communities from their own regions, creating accountability and shared ownership rather than parallel conversations that never intersect. Asynchronous and digital participation options for those who cannot travel, with genuine mechanisms for those contributions to influence outcomes rather than being noted and set aside. And critically, the Dialogue should build in real-time public transparency, with open sessions, live documentation and mechanisms for the broader public to follow and respond. Governance processes that happen behind closed doors, however well-intentioned, struggle to build the legitimacy that effective global cooperation requires.

Several examples stand out as genuinely instructive, though none is without limitations. The EU AI Act is the most significant binding regulatory framework to date. Its risk-based approach, which applies stricter requirements to higher-risk applications, is a sensible model for proportionate governance. Its extraterritorial reach also means it is already shaping practices well beyond Europe. The challenge is that it was developed primarily by and for wealthy, technically sophisticated jurisdictions, and its compliance requirements may be disproportionately burdensome for smaller actors and developing countries. Canada's Algorithmic Impact Assessment tool is a practical example of how governments can build accountability into their own procurement and deployment of AI systems. It is not perfect, but the principle of requiring public bodies to assess and disclose the risks of AI tools before deploying them is one that should be adopted much more widely. The African Union's Continental AI Strategy represents an important attempt to develop a governance vision that reflects African priorities and contexts rather than simply importing frameworks designed elsewhere. It deserves more international attention and support than it currently receives. On the technical side, model cards and datasheets for datasets have emerged as useful transparency tools that allow developers and users to understand the limitations and intended uses of AI systems. They are voluntary and inconsistently applied, but the underlying practice of structured documentation is worth formalising. The Ada Lovelace Institute in the UK has done important work on algorithmic accountability and public deliberation around AI, demonstrating how civil society can contribute rigorous, independent analysis that complements rather than simply critiques regulatory processes. What links these examples is that they combine transparency with accountability. Good intentions and published principles are not enough. Governance needs teeth, and these approaches, imperfect as they are, point in the right direction.