Imperial College London

Convergence on a shared governance baseline: A clearly articulated set of globally recognised principles (e.g., safety, accountability, transparency, human oversight) that move beyond high-level ethics toward operational definitions usable across jurisdictions. - see our papers (preprints) on this here: https://www.researchsquare.com/article/rs-8759109/v1 https://www.researchsquare.com/article/rs-9004371/v1 Agreement on a risk-tiered governance approach: Endorsement of a common framework that stratifies AI systems by risk (e.g., minimal, moderate, high-risk, frontier systems), enabling proportionate regulatory responses & reducing fragmentation across regions. Establishment of an interoperability roadmap with concrete plan to align existing regulatory regimes (e.g., EU AI Act–type approaches, OECD principles, national strategies), including mutual recognition mechanisms & technical standards harmonisation. Creation of a standing multilateral coordination mechanism such as greement to institutionalise the dialogue (e.g., permanent UN-affiliated AI governance forum or observatory) with defined mandates for monitoring, coordination & rapid response. Commitment to measurable accountability instruments facilitated by the introduction of auditable mechanisms such as AI system registries (for high-risk systems), standardised impact assessments & independent audit protocols... Advancing technical standards for safety & evaluation including consensus on priority areas for standard-setting (e.g., robustness testing, bias evaluation, red-teaming of frontier models), with linkage to international standards bodies Inclusion of LMIC priorities with demonstrable integration of Global South perspectives, including capacity-building commitments, equitable access to data infrastructure & representation in governance bodies. Success requires avoiding a purely OECD-centric framework. Public interest safeguards & rights-based protections with clear articulation of protections against harms (e.g., misinformation, discrimination, privacy violations), aligned with international human rights frameworks, with pathways for enforcement. This is super important for decisions support including triaging (eg. symptom checkers & related). See our preprint on this here: https://www.researchsquare.com/article/rs-8824194/v1 Mechanisms for evidence generation & policy feedback including commitment to ongoing evaluation of AI systems in real-world contexts (e.g., health, labour & education), including data-sharing frameworks and research collaborations to inform adaptive governance. + defined implementation pathway with timelines...

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

These priorities reflect a focus on ensuring that AI systems are safe, equitable & implementable within real-world public service contexts, particularly health systems. Safe, secure and trustworthy AI is foundational, as the rapid deployment of AI in high-stakes domains such as healthcare, public health surveillance & decision support requires robust validation, risk stratification &post-market monitoring. Without this, there is a material risk of harm at scale. The emphasis on social, economic, ethical, cultural, linguistic and technical implications recognises that AI systems are not neutral technologies. Their performance and impact are shaped by context, including data representativeness, digital literacy & structural inequalities. Addressing these dimensions is essential to avoid widening existing disparities, particularly across underserved populations. Protection and promotion of human rights is critical to ensure that AI deployment aligns with established international frameworks, including privacy, non-discrimination, and autonomy. This is especially important in contexts where AI systems influence access to care, health information, and resource allocation. Transparency, accountability, and human oversight are necessary to operationalise trust. This includes explainability standards, auditability of systems, clear lines of responsibility & mechanisms for redress. In practice, this enables regulators, clinicians, and citizens to interrogate and challenge AI-driven decisions. These priorities support a shift from principle-based discussions toward enforceable, system-level governance that is responsive to real-world implementation and population health outcomes.

1- implementation & real-world evaluation. Existing discussions emphasise principles & upstream design, but there is limited focus on how AI systems perform once deployed in complex, resource-constrained environments. This includes post-deployment surveillance, adaptive regulation & continuous performance monitoring across diverse populations. 2- system-level impacts & unintended consequences. AI is often assessed at the level of individual tools, yet its cumulative effects on health systems, labour markets & information ecosystems remain under-examined. This includes workflow disruption, over-reliance & shifts in professional accountability. 3-While trust & rights are often 'discussed', there is insufficient focus on data ownership, cross-border data flows &concentration of data assets among a small number of actors - with impositions or equity, sovereignty & innovation capacity. 4- Access to computational resources and advanced models is highly uneven globally. Without explicit consideration, this risks entrenching a two-tier AI ecosystem, limiting participation from LMICs 5- There is a need for internationally agreed methodologies for assessing effectiveness, safety & value in real-world settings, particularly in high-stakes domains such as healthcare. 6- Governance of general-purpose and frontier AI supert important given the rapid advances in general-purpose models raise novel challenges that cut across all themes, including systemic risk, dual-use concerns & the pace of capability development relative to regulatory capacity.

In the UK health &public health sector, governance gaps in AI are already having material effects on both system performance & population outcomes. A primary challenge is the LTD availability of robust, real-world evaluation frameworks for AI-enabled tools. Many systems are deployed with evidence of technical performance but insufficient validation in routine clinical or community settings. This creates risks related to safety, effectiveness & unintended bias, particularly for underserved populations. Another challenge relates to transparency and accountability. In practice, there is often limited clarity on responsibility when AI-informed decisions influence clinical pathways or patient behaviour. This is compounded by low explainability in some systems and variable levels of digital & health literacy among users. Fragmentation of governance approaches across jurisdictions & institutions also creates barriers to implementation. At the same time, there are significant opportunities since AI has the potential to strengthen prevention-oriented health systems through earlier risk detection, personalised self-care support & improved system efficiency (this is a key area of focus for Imperial SCARU)... This is because of the potenital to enhance access to health information and services, particularly in resource-constrained settings (primarily by inflating the "pre-primary care bubble") & support workforce capacity through decision augmentation.

The AI Dialogue can play a critical role as a neutral, multilateral platform for aligning fragmented approaches to AI governance and accelerating practical international cooperation - primarily by facilitating convergence on shared frameworks and terminology, reducing fragmentation across national & regional regulatory approaches (inc. advancing interoperability of risk-based governance models, evaluation standards & accountability mechanisms). The Dialogue can act as a coordination hub linking existing initiatives across the UN system, OECD, standards bodies & regional regulators since mapping efforts and identifying gaps ca reduce duplication and promote coherence in areas such as safety testing, audit protocols, and data governance. The Dialogue can also support the development of internationally recognised technical & policy standards, particularly for high-risk and general-purpose AI systems. This includes promoting common approaches to impact assessment, post-deployment monitoring & independent oversight. Crucially the intiaitve can strengthen inclusive participation by embedding the perspectives of lLMICs & underrepresented communities as this is essential for addressing global disparities in data access, infrastructure & governance capacity.

We are doing something related in UK (specifically in terms of streamlining timely access o regulatory support for SaMD/AIaMD): Check out https://radiant-cersi.org/ Imperial SCARU also leading a huge Digital Health Validation & Regulatory Science WP. Specifically on development of methodological standards for evaluating AI-enabled triage systems, digital symptom assessment tools & SaMD integrating research using RWE & clinical vignette benchmarking, NHS utilisation data & behavioural response modelling to assess diagnostic safety, escalation accuracy & regulatory readiness of DHTs. I''d be delighted to get involved & support your initiative if invited.

Different stakeholders should have clearly defined, complementary roles within a structured, multi-layered dialogue. Governments should provide regulatory perspectives, commit to policy alignment, & ID implementation constraints. Industry should contribute technical expertise, safety practices & transparency on model development and deployment. Academia should provide independent evidence, evaluation frameworks & methodological standards. Public representatives should articulate lived experience, equity concerns & societal impacts. The Dialogue should be structured in three tiers: (i) high-level plenaries to set political direction, (ii) technical working groups focused on specific themes (e.g., safety, evaluation, data governance), & (iii) implementation labs to translate principles into actionable tools (e.g., audit templates, regulatory sandboxes). Outputs should be concrete and time-bound, including draft standards, policy roadmaps, and pilot collaborations. Continuity is essential; the Dialogue should operate as an ongoing process with defined milestones rather than a one-off event.

Key underrepresented groups include LMICss, frontline public service providers (e.g., clinicians, educators), patients & end-users + communities affected by digital exclusion. SMEs & local innovators are also often absent despite being critical to implementation. Inclusion requires more than representation. Mechanisms should include funded participation, regional consultations & multilingual engagement to reduce structural bariers. Deliberative approaches (e.g., citizen panels) can ensure meaningful input from affected populations. There should also be formal integration of these perspectives into decision-making processes, not only consultation. This includes representation within working groups, co-design of policy outputs & feedback loops to demonstrate how input influences outcomes.

Traditional panel formats are nice but should be complemented by more interactive & output-oriented approaches. Fro ecxampel scenario-based simulations can test governance responses to real-world cases (e.g., AI in healthcare or crisis settings). Policy "sprints" or design labs can bring diverse stakeholders together to co-develop specific tools or standards within defined timeframes. Red-teaming & stress-testing exercises can be used to examine risks and governance gaps in frontier systems wehreas structured debates between regulators, developers, academics & affected communities can surface trade-offs more transparently. Digital platforms should be used enable as/ynchronous global participation, including crowdsourced input & iterative feedback on draft outputs. Hybrid formats combining in-person & virtual engagement will be important to maximise inclusivity & continuity.

Check out: https://radiant-cersi.org/