Federal State Budgetary Institution «N.N. Blokhin National Medical Research Center of Oncology» of Ministry of Health of the Russian Federation

The first outcome should be a consolidated summary of discussions that captures key areas of agreement, divergence, and priority challenges in AI governance. Such a document should structure participants' positions across thematic areas and clearly identify points of convergence and divergence. This would contribute to establishing a foundation for further work, enabling a transition from general understanding to more focused and substantive discussions, and supporting the development of practical conclusions. The second outcome should be the establishment of a sustainable platform for ongoing engagement, with clearly defined formats for continued work, including regular meetings, thematic tracks, and mechanisms for the exchange of practices. This would provide a structured framework for future discussions. A key condition in this regard is inclusiveness: the Dialogue should take into account the perspectives of countries at different levels of economic and technological development, including developing countries, for which access to technology, infrastructure, and data remains critical. This would contribute to a more balanced and genuinely multilateral agenda and support efforts to reduce the widening gap in AI capabilities. The third outcome should be a clearer understanding of potential areas and stakeholders for future collaboration. It is important to identify not only topics for cooperation but also areas of alignment among participants where joint work is feasible. This may be reflected in a willingness to continue engagement through working formats or practice exchange. Such interaction should contribute not only to the sharing of experience but also to the development of complementary initiatives and practical outcomes applicable in subsequent work.

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

Safe, secure, and trustworthy AI: Safety and reliability constitute fundamental requirements for medical AI solutions. Errors in models may directly affect clinical decision-making; therefore, data quality, model validation, and interpretability are of critical importance. AI capacity-building: AI enables addressing tasks that are either beyond human capability or would require significantly greater time and resources. It facilitates the identification of weak signals and hidden patterns in large and heterogeneous datasets, thereby supporting earlier disease detection, forecasting disease progression, identifying risk groups, optimizing patient pathways, and strengthening clinical decision support through integrated data analysis. AI also supports the transition toward personalized medicine by enabling therapy selection based on individual patient characteristics, predicting treatment response, and contributing to the development of targeted and personalized vaccines. In addition, it accelerates medical image analysis, automates routine processes, reduces clinician workload, and enhances diagnostic accuracy through the integration of multiple data sources. Social, economic, ethical, cultural, linguistic, and technical implications: The implications of AI deployment directly shape its practical application and impact on healthcare systems. These include improvements in the quality and accessibility of care, more accurate and timely diagnostics, optimization of processes, and more efficient resource utilization. AI deployment also supports the development of new approaches to healthcare organization and requires due consideration of ethical, linguistic, and cultural factors to ensure sustainable and scalable implementation. Transparency, accountability, and human oversight: These are essential for the integration of AI into clinical practice, as decisions supported by such systems directly affect patient health and safety. Clinicians should be able to understand how model outputs are generated, critically assess them, and retain final control over decision-making. This ensures clinical validity, strengthens trust among medical professionals and patients, and supports clear allocation of responsibility.

First, data governance requires dedicated attention. While UNGA Resolution 79/325 highlights dataset publication as a key driver of AI development, this approach remains associated with significant risks. These include limited control over data use, application outside the original context, unequal distribution of value, privacy risks, and legal uncertainty, including issues of ownership, licensing, and liability. These factors may reduce incentives to create and share high-quality datasets and exacerbate imbalances in their use. It is therefore important to develop aligned approaches to data governance, including access conditions, attribution, licensing, and principles of use. An international classification system and registry of datasets are also needed to improve transparency, reduce duplication, and facilitate discovery and reuse, thereby creating a more predictable environment for cross-border collaboration. Second, the cross-border use of digital services represents a significant issue. Differences in regulation, security requirements, and levels of technological development create barriers to the import and application of digital solutions. More consistent and harmonized rules are needed, including requirements related to data, security, and interoperability. This is particularly important for enabling inclusive participation by countries with varying levels of technological development, as reducing such barriers would expand access to AI solutions and help bridge existing gaps. Third, it is advisable to structure discussions by sector. Requirements for AI systems, acceptable levels of risk, and conditions for deployment vary significantly across domains. A sectoral approach allows these differences to be taken into account and leads to more practical and applicable outcomes than general recommendations. Fourth, the use of large language models requires separate consideration. Their rapid deployment is outpacing the development of requirements related to reliability, limitations, and risk management. This creates uncertainty for both developers and users, particularly in sensitive sectors. More systematic approaches are needed to assess their quality, predictability, and safe use.

Key gaps in AI governance in healthcare: 1. Insufficiently developed and fragmented regulation. Requirements for the development, clinical validation, registration, and monitoring of AI remain heterogeneous and insufficiently specified. This creates uncertainty and slows down the adoption of technologies. 2. Fragmentation of data and infrastructure. Medical data are distributed across multiple information systems, stored in heterogeneous formats, and often include large volumes of unstructured data. The absence of unified standards, data-sharing mechanisms, and access rules limits their use for AI development, validation and constrains scalability. 3. Limited institutional coordination of AI implementation. There is a lack of stable mechanisms at the system level for evaluating, deploying, scaling AI solutions. Therefore, many solutions remain at the pilot stage, not integrating into broader practice. 4. Insufficient development of managerial and interdisciplinary competencies. There is a shortage of specialists capable of working at the intersection of medicine, data, and management, who can assess the applicability of AI solutions and support their implementation throughout the lifecycle. 5. Limited sharing of unsuccessful AI implementation experiences. Existing knowledge-sharing mechanisms, particularly peer-reviewed publications, tend to focus on successful cases. Therefore, much practical experience – failures, limitations, and unsuccessful solutions – remains outside the public domain. This limits learning and leads to repeated mistakes, especially in countries with limited experience and resources. Hence, there is a need for platforms, enabling systematic exchange of both successful and unsuccessful AI implementation experiences. Opportunities: 1. Improving healthcare efficiency. AI is already used for diagnostics, forecasting, and resource management, which can help reduce the burden on healthcare systems. 2. Shaping new standards. The current stage offers an opportunity to establish sustainable governance approaches, from risk assessment to the full lifecycle of AI models. 3. International coordination. Aligning approaches to data and regulation can accelerate the deployment of solutions and reduce barriers.

The Global Dialogue can serve as a platform for comparing national approaches to AI governance, reducing fragmentation of the international agenda, and enhancing coherence and mutual understanding between approaches. In the context of differing strategies and levels of technological development, it enables the identification of common ground and areas where coordination is possible. A key role of the Dialogue is also to establish a shared conceptual and methodological foundation. Aligning terminology, core principles, and approaches to risk assessment creates a basis for more comparable regulation and facilitates interaction between countries and organizations. The Dialogue can also support a transition from discussions to more sustained forms of cooperation, including the creation of thematic working groups, exchange of practices, and the development of joint initiatives in priority areas. This is particularly important for issues requiring collective solutions, such as data governance, safety, and building trust in AI. Another important function is supporting countries at different levels of AI readiness. Sharing experience, providing access to existing knowledge, and coordinating capacity-building efforts can help reduce existing gaps and foster a more inclusive and accessible platform for cooperation

Firstly, the Dialogue should take into account the findings of the UN report Governing AI for Humanity (2024) that the current AI governance landscape remains fragmented: numerous initiatives exist, but gaps persist in representation, coordination, and implementation. None of the existing mechanisms is truly global, which limits their effectiveness. Secondly, the UNESCO Recommendation on the Ethics of Artificial Intelligence (2021) establishes requirements for transparency, accountability, and human oversight, as well as mechanisms for assessing ethical risks throughout the AI lifecycle. Thirdly, the report "AI Standards for Global Impact" (2025), prepared under the auspices of the ITU, presents standards as a tool for applying principles in practice through requirements for testing, evaluation, safety, and the use of AI systems, helping to ensure comparability of approaches and build trust in technologies. Moreover, the BRICS Leaders' Statement on AI Governance emphasizes the need for an inclusive global system and coordination through the UN; the Shanghai Declaration on Global AI Governance highlights the importance of international cooperation, equitable access to technologies, and the development of regulatory mechanisms; and the G20 AI principles establish a coordinated agenda among major economies around human-centered and trustworthy AI. This helps incorporate diverse perspectives and reduce the risk of fragmentation. Furthermore, international standards, including ISO/IEC 42001, can serve as a tool for alignment by establishing common requirements for AI governance. At the same time, ensuring inclusiveness requires considering a wider range of approaches across different levels of technological development and institutional contexts; otherwise, frameworks risk being shaped by a limited set of countries and becoming less applicable elsewhere. The added value lies in its ability to connect existing principles, policy frameworks, and technical standards into a coherent framework, ensuring alignment, accommodating diverse contexts, and supporting the transition from general approaches to practical and implementable solutions.

Different stakeholder groups, including countries with varying levels of technological development, bring different types of expertise. Governments establish regulatory frameworks and define the conditions for technology deployment. Research institutions develop methods and assess their applicability. Businesses are responsible for creating and scaling solutions. Practitioners provide insight into the real-world use of technologies. The structure of the Dialogue should reflect this distribution of roles. In addition to plenary sessions, it should include focused working groups and thematic discussions where participants engage with specific tasks, cases, and challenges directly related to their professional activities and applicable to their work. Equally important is enabling direct interaction between these groups, ensuring that discussions reflect not only positions but also practical experience. This helps produce more applicable and actionable outcomes.

First, participants working at the level of practical implementation of technologies are underrepresented. Their experience helps identify potential risks and real constraints related to data quality, infrastructure, system interoperability, and process organization. Without their input, discussions may remain at the level of general principles with limited practical applicability. Second, participation from countries and organizations with different levels of technological and institutional development remains limited. As a result, approaches may fail to reflect diverse contexts and can reinforce existing gaps in access to AI and the ability to use it. Third, groups directly affected by the deployment of AI are insufficiently represented. Their exclusion reduces the sensitivity of discussions to the social impacts of technologies, including issues of fairness, accessibility, and potential forms of implicit exclusion. Finally, specialists working with data and infrastructure are comparatively underrepresented, even though key issues of quality, interoperability, security, and data lifecycle management are addressed at this level. To include these groups more effectively, interaction formats need to be adapted. This may involve creating thematic and sector-specific working groups, engaging participants in discussions of concrete use cases, and developing open, multi-level formats (including regional consultations and online platforms) that lower participation barriers and better reflect diverse contexts and practices.

The most effective formats are those where participants work on specific tasks with a clear, fixed outcome. This can include short working cycles on focused topics, structured around a predefined problem and progressing through three stages: clarifying the context, discussing possible solutions, and consolidating conclusions. Such a structure makes discussions more focused and manageable, while producing results that can be used for further development or alignment. Panel discussions are also effective when speakers focus on a single question. The moderator sets the framework, and participants respond to each other in sequence, refining their positions. This helps maintain coherence and reach substantive conclusions more quickly. Digital tools can further enhance discussions, such as quick pre-session polls to capture participants' positions, and collaborative boards to document ideas, agreements, and disagreements in real time. Cross-sectoral groups are particularly useful, bringing together different perspectives on the same issue (e.g., regulators, developers, and users). This helps identify constraints more quickly and find applicable solutions. It is also important to include formats focused not only on successful cases but also on problematic implementations. Dedicated sessions or working groups can analyze real constraints, errors, and reasons for failure. This increases the practical value of discussions and helps participants better account for risks and avoid repeating mistakes, especially in countries at earlier stages of AI adoption. Finally, it is to ensure continuity beyond the sessions. The same groups can continue refining their proposals in smaller formats, helping maintain momentum and gradually move from discussion to more structured outcomes.

In Russia, a key role is played by the National Strategy for the Development of AI up to 2030. The Strategy sets out long-term goals, principles, and priorities for AI development, including requirements for safety, data quality, and the practical application of technologies. It provides the foundation for national policy and ensures alignment of actions across stakeholders, from public authorities to research institutions and businesses. The Strategy is implemented through the federal project "Artificial Intelligence," within which research centers are established across sectors. For example, the Research Center for Artificial Intelligence in Healthcare, based at the N.N. Blokhin National Medical Research Center of Oncology, carries out projects across multiple areas. In particular, the Center develops predictive models based on multimodal medical data for the early detection of cancer, including conditions without existing screening programs. These models achieve performance comparable to international benchmarks and enable the identification of risks several years prior to diagnosis. In parallel, computer vision solutions are being developed for the analysis of medical imaging (CT, MRI, mammography, PET/CT), improving diagnostic accuracy and reducing clinician workload. Research is also underway in personalized anticancer vaccines and radiopharmaceuticals, contributing to improved treatment effectiveness, more precise diagnostics, reduced recurrence risk. A key feature of the Center's work is its full-cycle approach, from clinical problem definition and dataset development to model validation, regulatory approval as a medical device, deployment in healthcare practice. This ensures scalability and practical applicability of solutions. Such initiatives demonstrate an effective model of AI governance, in which strategy is supported by institutions, technologies, and implementation, and may serve as a basis for replication in other sectors and countries. Furthermore, experimental legal regimes, implemented under Federal Law No. 258-FZ, allow the testing of AI technologies in real-world conditions under temporary regulatory frameworks, supporting their validation prior to large-scale deployment.