ANO HE Innopolis University

The success of the first Global Dialogue will not lie in the fact of the meeting itself, but in the establishment of a clear international framework for future work. From our perspective, a successful outcome must include four elements. Firstly, agreement on a basic minimum standard for international AI accessibility -an access to computing resources, foundational models, educational programs, evaluation tools, and multilingual services. Secondly, the launch of a transparent cycle of follow-up actions: an annual agenda, a public decision matrix, regular progress reviews, and linkage to the annual scientific report. Thirdly, recognizing that the AI challenge is not only about safety but also the risk of widening the global divide, as access to computing power, data, and expertise is already concentrated in a few countries and companies. Fourthly, a practical orientation: the dialogue must generate coalitions focused on capacity building, open models, multilingualism, and the application of AI in science, rather than being limited to declarative discussion. If the first meeting establishes these working guidelines, it will be possible to say that the Global Dialogue has begun to fulfil its purpose.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Open-source software, open data and open AI models

This choice reflects both our institutional practices and the most pressing global risks. Capacity building is a key priority because it is precisely the shortage of computing power, personnel, data, and expertise that determines who will develop AI and who will become merely a consumer of others' solutions. Open models and data, as well as open-source software, are important both as a tool for reducing dependence on a narrow circle of suppliers and as a prerequisite for the development of local languages, scientific schools, and applied ecosystems. This logic resonates with the position of Yann LeCun, who explicitly warned against a scenario in which AI would fall under the control of "a few corporate entities," and also emphasized the importance of an open approach for future systems through which people will interact with knowledge. Safe and trustworthy AI is essential, as its implementation in medicine, education, governance, and industry is impossible without managed risks, auditing, and human oversight. Finally, social, cultural, and linguistic implications cannot be ignored: if models normalize a single language, a single style, and a single set of interests, this undermines both the legitimacy of AI and the fairness of its distribution.

Yes, the list of topics is missing at least two key areas. The first is equal access to computing infrastructure and basic AI services. Today, it is precisely access to graphics processing units, cloud infrastructure, large models and application programming interfaces that is becoming a new form of infrastructural inequality. Export control regimes have already attempted to introduce a multi-tiered classification of countries regarding access to advanced computing chips, whilst commercial platforms publish lists of supported countries and block the use of services outside these lists. Such logic may be motivated by security or regulatory requirements, but its systemic effect is the fragmentation of the global AI landscape and the risk that the benefits of AI will be accessible only to the "club of the wealthy". The second missing block is AI for scientific discovery. The next stage of AI development must include a separate priority for interdisciplinary foundational models and models of the physical world in chemistry, physics, biology, materials science and medicine. Various world-class experts point to the same limitation: universal language models do not fully meet the needs of science and the physical world; a class of systems is needed that models causality, environmental dynamics and specialised scientific representations. Innopolis University considers this a priority and a distinct area for international cooperation.

For the Russian research and higher education sector, the main challenge is not regulation itself, but the combination of regulatory fragmentation with unequal access to computing power, data, cloud infrastructure and cutting-edge services. At the global level, this fits into a broader picture: differences in skills, connectivity, computing resources and governance will widen the international divide. For universities, this means a slowdown in fundamental research and fewer opportunities for long-term, high-risk research and development projects. However, the opportunities are no less significant. If international authorities support open models, regional computing power, compatible evaluation standards and the trusted use of AI, this will accelerate its integration into medicine, industry and education. For Innopolis University, such an architecture is particularly important because its portfolio already combines applied solutions and research areas – ranging from industrial AI services to models for chemistry and other scientific tasks.

The Global Dialogue can serve as a universal bridge between scientific expertise, policy decisions and practical capacity-building. Its uniqueness lies in the fact that it is conceived as the first international format under the aegis of the UN, where countries and other stakeholders discuss AI within a single institutional framework, rather than through a patchwork of disparate forums and regional regimes. For this role to be realised, the dialogue must fulfil three functions. The first is to translate the scientific panel's findings into politically viable, yet not excessively restrictive decisions. The second is to help countries coordinate minimum common guidelines on safety, openness, risk assessment and human oversight, without imposing a single regulatory model. The third is to launch specific international instruments: a capacity-building network, the exchange of standards and assessment methodologies, multilingual educational programmes, as well as mechanisms for shared access to computing infrastructure for science and socially significant tasks. It is particularly important that this dialogue becomes a platform for the systematic development of interdisciplinary research. This does not involve a mere declaration of the integration of disciplines, but rather the creation of new classes of models - fundamental models and world models, which integrate machine learning with physics, chemistry, biology and engineering sciences. Such systems require the integration of different types of data, physics-informed approaches, neurosymbolic methods and multi-agent architectures, which is already reflected in contemporary research programmes, including those at Innopolis University. Unlike general-purpose language models, they are capable of modelling causality, process dynamics and the behaviour of complex systems, which is critical for scientific discoveries and industrial tasks. In this context, the Global Dialogue can launch international consortia to develop such models, facilitate the exchange of scientific data and benchmarks, and agree on principles for open and equitable access to computational resources. This will help avoid fragmentation of the scientific landscape and accelerate the transition from AI as a tool for automatisation to AI as a tool for scientific discovery. The added value of the dialogue should lie not in competing with existing initiatives, but in linking them together, legitimising them, and opening them up to those countries and institutions that are currently outside the decision-making centre.

The global dialogue should not start from scratch, but should build on the already existing mechanisms. Firstly, on UNESCO's regulatory and applied ecosystem - its recommendations on AI ethics, its readiness assessment methodology, and its global observation on AI ethics. Secondly, on the standardisation and multilateral work of the ITU and the 'AI for Good' platform. Thirdly, on practical risk management frameworks. Fourthly, on infrastructure models for access to resources. The added value of the Global Dialogue lies in the fact that none of these initiatives individually possesses universal political legitimacy, global geographical coverage and a direct link to the sustainable development agenda all at once. The Dialogue can serve as a coordinating layer between them, particularly in terms of capacity building, open models, multilingualism and AI for science.

• Different groups should make a practical, rather than symbolic, contribution. Governments should establish the policy framework and allocate resources. The academic community should provide verifiable expertise, benchmarks and impact assessments. Business should provide data on real-world implementation scenarios, incidents and technical limitations. Civil society, trade unions and professional associations should document the impact on rights, employment and public trust. Specialist scientists and engineers from the fields of medicine, chemistry, energy, education and industry should play a particular role, as they are the ones who understand where universal models fall short and where specialised scientific systems are required. In terms of format, Innopolis University would support a multi-tiered structure: a high-level plenary segment
• thematic working groups
• implementation laboratories for pilot solutions
• an open register of written contributions
• regional preparatory consultations
• and a short summary document setting out specific tasks for the next session. This format is consistent with the architecture of the first dialogue already under discussion, but requires a stronger focus on measurable outcomes.

Five groups are currently under-represented in global discussions on AI governance. The first comprises universities, research centres and regulators from countries with limited access to computing resources and services. The second are communities of low-resource languages and cultures. The third consists of developers of open-source solutions and socially significant projects which are not part of the major platforms. The fourth group comprises specialists in applied scientific fields. The fifth includes workers and their representatives across the entire AI value chain. Their inclusion requires more than just open registration. Support for travel and participation grants, regional hubs, mandatory multilingual materials and consultations, transparent quotas by geography and sector, as well as dedicated opportunities for written contributions are required. This is particularly important because differences in infrastructure, skills and digital access lead to deepening inequalities, whilst linguistic and cultural homogenisation can reduce diversity.

The most effective formats will be those in which discussion is directly linked to collaborative decision-making. Innopolis University would propose four tools. The first is sessions for rapid decision-making on specific topics: access to computing resources, open models, multilingualism, trusted deployment, and AI for science. The second is expert hearings, where members of the scientific panel present an analysis of specific risks, trade-offs and uncertainties. The third is practical implementation reviews: an analysis of real-world case studies from countries and institutions. The fourth is a public commitment panel, where voluntary actions are documented and agreed upon until the next meeting. It is also useful to utilise regional centres and hybrid participation format to reduce barriers related to time, language and travel.

There already exist some useful examples: - The UNESCO Framework is one of the best examples of the transition from principles to implementation. - AI initiatives for science, focused on creating foundational models for scientific tasks. For Innopolis University, it is precisely this combination of approaches - ethics, security, risk management, openness, access to computing resources and AI for science - that appears to be the most productive.