University Hospitals of Leicester NHS Trust

A successful first Global Dialogue on AI Governance would do three things. First, it would move beyond general principles alone and identify a small number of practical governance gaps that need global attention. One of these gaps is the distance between model-level governance and the baseline data (e.g minimum core data) and standards that AI systems depend on in real-world settings. In safety-critical domains, AI can only be as safe, fair, and interoperable as the information it is able to see and act upon. The Dialogue's concept note already points to safe, secure and trustworthy AI, interoperability, and transparency, accountability and robust human oversight as core objectives. A useful outcome would be to connect those objectives more explicitly to the underlying data and standards layer. Second, success would mean producing a practical roadmap for July that identifies thematic clusters and concrete outputs, not only broad aspirations. For example, the Dialogue could recommend guidance on baseline data requirements for safety-critical domains, traceability and provenance expectations, and approaches for identifying where local variation or underrepresentation may be scaled by AI. This would respond directly to the consultation's request for actionable themes and outputs. Third, the first Dialogue would be successful if it demonstrates a genuinely multistakeholder implementation logic- governments and multilaterals for legitimacy and coordination, professions and civil society for real-world meaning and accountability, and technology and academia for standards, testing, and implementation. That would help ensure AI governance is not only principled, but also implementable across different countries and levels of capacity. In short, success would mean making AI governance more operational, governing not only the models, but also the baseline information and structures they rely on.

• Interoperability of governance approaches
• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI

I selected these priorities because, in my view, trustworthy AI depends not only on the behaviour of models, but also on the quality, structure, and governance of the data and workflows those models rely on. Safe, secure and trustworthy AI is essential because AI systems in safety-critical settings can only be trusted if the underlying information they act upon is reliable, sufficiently complete, and governed appropriately. Interoperability of governance approaches matters because fragmented governance can reproduce fragmented implementation. If countries, sectors, or organisations adopt AI under incompatible assumptions about data, safety, or accountability, the result may be uneven standards and increased inequity. Transparency, accountability, and human oversight are critical because AI systems must remain understandable, contestable, and subject to meaningful professional and institutional responsibility. This is especially important where AI influences documentation, decisions, or workflows that affect people directly. I also selected social, economic, ethical, cultural, linguistic and technical implications of AI because governance must reflect real-world diversity. AI systems do not operate in a vacuum; they act within social and professional contexts. If the underlying data structures and assumptions do not reflect those contexts, exclusion and bias can become embedded in design and scaled through implementation. Taken together, these priorities help shift AI governance from general principles alone toward practical, implementable governance that connects model oversight with data quality, interoperability, inclusion, and real-world accountability.

One important cross-cutting issue not yet captured clearly enough is the governance of the baseline data and standards that AI systems depend on. Much AI governance discussion focuses on models, compute, transparency, and oversight. These are essential. But AI also acts on whatever information is structured, visible, and computable underneath it. If that baseline is incomplete, inconsistent, or locally fragmented, those weaknesses can be scaled by AI. This matters across many of the listed themes. It affects safety and trustworthiness, transparency and accountability, interoperability, human rights, and the wider social and technical implications of AI. It is also an emerging issue because new forms of AI-especially generative and agentic systems- are beginning to reason over, summarise, and trigger workflows from existing records and data structures. That means governance needs to pay more attention not only to the behaviour of the model, but also to the quality, representativeness, and consistency of the underlying data architecture. A related issue is the gap between profession-agnostic or sector-wide frameworks and the domain-specific information needed for safe implementation in practice. Broad governance frameworks are necessary, but they do not always define the minimum core information that must be consistently represented in safety-critical settings. Without that, local variation can become embedded design, and inequality can be reproduced through implementation. In my view, the Dialogue could usefully address this by recognising baseline data requirements, provenance, and standards governance as cross-cutting enablers of trustworthy AI. This would help connect high-level principles with real-world implementation.

In my sector - health and care, particularly within digital health and nursing informatics in the UK - the governance gaps are already having practical effects. AI adoption is moving faster than agreement on the underlying data, standards, and accountability structures needed to support safe implementation. This is especially visible in areas such as AI-assisted documentation, decision support, and workflow automation. The most significant challenge is that governance often focuses on the model or tool, while the underlying data remains inconsistent, fragmented, or under-structured across organisations. In practice, this means local templates, varying documentation approaches, and uneven data quality can become embedded into AI-enabled workflows. That creates risks for interoperability, transparency, professional accountability, and equity. It also makes it harder to evaluate whether systems are genuinely improving care or simply automating inconsistency. A second challenge is the gap between broad, profession-agnostic frameworks and the specific information needs of safety-critical practice. In healthcare, if the baseline data needed to represent clinical reasoning, care continuity, and handover is not consistently defined, then AI may act on a partial picture. At the same time, there is a major opportunity. In the UK and more broadly, stronger attention to data standards, interoperability, and accountability could make AI governance more practical and more equitable. This is particularly important for public health systems and for low- and middle-income settings, where limited resources make it even more important to avoid duplication, vendor lock-in, and fragmented implementation. The opportunity is not only to govern AI models better, but to strengthen the foundations they rely on- baseline data, shared standards, provenance, and meaningful human oversight. If done well, this can support safer innovation, stronger interoperability, and more inclusive digital transformation across systems.

The AI Dialogue can play a vital role by helping translate broad global principles into practical cooperation mechanisms that countries, sectors, and institutions can actually use. Its value is not only as a forum for discussion, but as a bridge between policy, implementation, and trust. First, it can create a shared space for identifying common governance gaps across regions, including where current frameworks are strong at the principle level but weaker at the level of implementation, interoperability, accountability, and real-world oversight. Second, it can support convergence without forcing uniformity. Different countries and sectors will adopt different regulatory and institutional models, but the Dialogue can help establish shared reference points around safety, transparency, human oversight, provenance, and interoperability. That would make international cooperation more realistic and inclusive. Third, it can elevate voices that are often underrepresented in global AI governance- including public sector practitioners, professions, civil society, and actors from low- and middle-income settings- so that governance is shaped not only by states and major technology companies, but also by those working closest to implementation and impact. Fourth, it can help surface practical outputs, such as shared guidance, implementation principles, common terminology, and examples of good governance practice that can travel across borders and sectors. Finally, the Dialogue can reinforce the idea that AI governance is not only about governing models, but also about governing the data, standards, and institutional conditions that make safe and fair AI possible. In that sense, the Dialogue can become a mechanism for building international cooperation through shared understanding, shared vocabulary, and shared implementation pathways- not only shared concern.

The AI Dialogue should build on existing initiatives that already provide global principles, implementation tools, and multistakeholder networks. In particular, it should connect with the Global Digital Compact, which established the mandate for a UN Global Dialogue on AI Governance and provides the broad political framework for inclusive digital cooperation. It should also build on UNESCO's Recommendation on the Ethics of Artificial Intelligence, which remains the only global normative instrument on AI ethics across 194 Member States and already includes policy action areas such as data governance, health, inclusion, and accountability. The Dialogue should also connect with the OECD AI Principles and the integrated GPAI/OECD.AI partnership, which provide practical policy reference points, implementation tracking, incident monitoring, and multistakeholder expert communities. In safety-critical sectors, it should draw on domain-specific guidance such as WHO's Ethics and Governance of AI for Health, which addresses accountability to health workers and affected communities. Finally, it should connect with the UN High-level Advisory Body's recommendations, especially the emphasis on common understanding, common ground, capacity development, and a more coherent international AI governance ecosystem. The added value of the AI Dialogue would be to connect these initiatives, rather than duplicate them. It can help translate broad principles into shared implementation pathways, identify gaps between frameworks, and create a space where governments, multilaterals, civil society, professions, academia, and technology actors can align around practical governance needs. It can also elevate underrepresented implementation perspectives, including those from public services and low- and middle-income contexts, and help connect model-level governance with the underlying data, standards, and institutional conditions that trustworthy AI requires.

Different stakeholders should contribute to the AI Dialogue in ways that reflect both their expertise and their role in implementation. Governments and multilateral organisations can contribute legitimacy, coordination, and links to public policy. Technical communities and academia can contribute standards expertise, evidence, testing methods, and analysis of emerging risks. Civil society and affected communities can bring accountability, human rights, inclusion, and lived experience. Professional groups and public sector practitioners can contribute implementation realities, especially in safety-critical domains where AI affects workflows, documentation, decisions, and service delivery. This kind of multistakeholder participation is consistent with the Dialogue's purpose and structure as set out in the concept note. In terms of format, the Dialogue would be most useful if it combined three elements: First, plenary discussion to build shared understanding of priority issues and common vocabulary across regions and sectors. Second, smaller thematic working sessions organised around practical clusters, such as safety and trustworthiness, interoperability, accountability and human oversight, and the governance of underlying data and standards. This would help move beyond general statements into more actionable discussion. Third, structured written input mechanisms before and after meetings, so that participants who cannot speak live- including those in different time zones or with limited resources- can still contribute meaningfully. This is especially important for inclusive participation. I would also recommend that each Dialogue produce a small number of clear outputs, such as areas of convergence, unresolved questions, and practical recommendations for future work. That would help ensure continuity between meetings and make the process cumulative rather than purely consultative. Overall, the Dialogue will be strongest if it is genuinely multistakeholder, implementation-oriented, and designed to connect high-level principles with real-world governance needs.

Several voices remain underrepresented in global AI governance discussions. First, public sector practitioners are often missing, especially those working inside health, education, social care, justice, and local government systems. These are the people closest to implementation, safety, and public accountability, yet they are often not included in high-level governance design. Second, professional communities in safety-critical domains are underrepresented. AI governance is often discussed at a general level, but professions such as nursing, teaching, social work, and public health can see where governance gaps appear in real workflows, documentation, and service delivery. Third, civil society actors and affected communities from low- and middle-income countries remain insufficiently heard. Global discussions can still be dominated by states, major technology companies, and well-resourced institutions, even though many implementation risks are most acute in settings with fewer resources, weaker infrastructure, or less bargaining power. Fourth, data and standards practitioners are often less visible than model developers and policy experts. Yet many governance failures emerge not only from the model, but from the quality, representativeness, and structure of the underlying data. These voices could be better included through a combination of: -dedicated seats or quotas in multistakeholder processes -stronger support for participation from low-resource settings -written input channels alongside live speaking opportunities -thematic working groups that include frontline practitioners, professions, and community representatives -and support for participation that goes beyond consultation, so underrepresented groups are involved in shaping outputs, not only reacting to them In my view, the Dialogue will be stronger if it includes not only those who govern AI at a distance, but also those who experience its implementation up close.

To foster meaningful and dynamic engagement, the AI Dialogue should combine formal multistakeholder diplomacy with more practical and participatory formats. First, it would benefit from short thematic roundtables rather than only long plenary statements. Smaller, moderated sessions can help participants engage more directly across sectors and regions, especially on issues such as safety, interoperability, accountability, capacity-building, and data governance. Second, the Dialogue could use challenge-based discussion formats. For example, participants could be invited to respond to a small number of concrete governance dilemmas or implementation scenarios. This helps move discussion from abstract principles to practical trade-offs and real-world questions. Third, written and asynchronous participation mechanisms are essential. Many stakeholders, especially from different time zones or lower-resourced settings, may not be able to engage fully in live sessions. Structured written submissions, short post-session feedback forms, and digital comment periods can widen participation and improve inclusivity. Fourth, it may be helpful to include cross-sector "bridging" sessions where technical experts, policymakers, practitioners, and civil society respond to the same issue from different perspectives. This can reveal where assumptions differ and where common ground exists. Fifth, the Dialogue could experiment with lightweight visual synthesis tools during meetings- for example, real-time mapping of themes, governance gaps, or areas of convergence- so that participants can see how the discussion is developing. Finally, continuity matters. Each session should end with a brief summary of: -key areas of agreement -unresolved tensions -proposed next steps This would help ensure that engagement is cumulative, not one-off. In my view, the most effective format is one that combines accessibility, structure, practical problem-solving, and visible continuity across meetings.

Several existing policies, practices, and approaches already offer useful building blocks for effective AI governance. At the global level, UNESCO's Recommendation on the Ethics of Artificial Intelligence is important because it provides a shared normative foundation across Member States, including principles on transparency, accountability, fairness, human oversight, and data governance. At the implementation level, the NIST AI Risk Management Framework is valuable because it translates broad trustworthiness goals into a practical risk-management approach for the design, development, use, and evaluation of AI systems. Its companion playbook is also a useful example of how high-level governance can be made more operational. In safety-critical sectors, WHO's guidance on ethics and governance of AI for health is a strong example of domain-specific governance. It addresses ethical risks, accountability, public benefit, and the role of health systems and health workers in ensuring safe implementation. At the regional regulatory level, the EU AI Act offers an important example of a risk-based legal framework that translates governance principles into obligations, including differentiated treatment of high-risk systems and governance/enforcement mechanisms. Beyond formal policy, practical approaches such as impact assessments, incident reporting, traceability/provenance requirements, and structured risk reviews are especially useful because they create feedback loops between governance and real-world deployment. The broader lesson from these examples is that effective AI governance requires more than one instrument. It works best when principles, risk-management methods, sector-specific guidance, legal accountability, and operational practices are connected. The AI Dialogue could add value by helping align these layers, identifying where gaps remain, and making practical solutions more visible and transferable across countries and sectors.