Inria

The main stake is to preserve strong political steering and avoid risks of intellectual or operational capture. First, capitalising on existing initiatives remains essential. Building on ongoing programmes and infrastructures enables rapid deployment and effective use of accumulated experience. However, this should be accompanied by clear participation rules (including transparency and selection mechanisms) to prevent concentration of influence among a limited set of highly visible stakeholders. Second, interdisciplinarity must be deliberately structured. A purely technology-driven approach would risk overlooking key economic and societal dimensions. Strengthening links between AI, data, and economic value creation ("datanomics") is critical. This requires formal integration of economic and societal expertise in governance and project selection processes, building on approaches already explored in initiatives led by Inria in the Global Partnership on Artificial Intelligence (GPAI). Third, youth engagement should be embedded by design. Beyond symbolic participation, dedicated mechanisms should ensure meaningful involvement in agenda-setting and implementation, thereby strengthening both legitimacy and long-term relevance. Existing initiatives led by Inria and its partners could be leveraged for the benefit of the GPD. Fourth, regarding centres of excellence, priority should be given to open and collaborative infrastructures that support capacity building. In this context, actors such as Inria can play a structuring role as participants in an international network, contributing to skills development and knowledge sharing. To avoid fragmentation and dilution of responsibility, this network should be supported by a clear governance backbone, ensuring coordination, accountability, and strategic prioritisation at the political level. Fifth, the Geneva sequence requires a clear and operational political mandate. The articulation between political and expert phases should be grounded in explicit objectives and a shared impact measurement framework, including economic, societal, and capacity-building indicators. A structured feedback loop between these phases would reduce risks of misalignment and reinforce coherence. Finally, a project-based approach should be prioritised, within a coherent portfolio logic. Launching a pilot initiative this year, followed by structured evaluation, comparison, and scaling, would enable a shift from volume-based logic toward demonstrable value creation. Common evaluation criteria and periodic review cycles should guide resource allocation and ensure cumulative impact.

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

The selected priorities are consistent with preserving strong political steering and ensuring that AI-generated value returns to citizens and the broader economic fabric, rather than being captured by a limited set of actors. Placing open source, open data, and open AI models at the core is a structuring choice. Open infrastructures enable distributed innovation, reduce asymmetries of power, and support a wider circulation of value. They allow public authorities to retain strategic oversight while ensuring that citizens, researchers, and businesses can actively contribute, thereby reinforcing democratic agency and limiting value capture. This directly supports AI capacity building, as open ecosystems lower entry barriers and enable a broader set of actors, including SMEs and public institutions, to develop capabilities and appropriate technologies. The focus on social, economic, cultural, linguistic, and technical implications ensures alignment with public value objectives. In line with approaches developed by Inria, AI should be assessed through its capacity to generate inclusive economic value and to shape societal and environmental trajectories, avoiding concentration of benefits. On governance, it is essential to articulate two complementary logics. Open source ecosystems rely on decentralised, iterative, and merit-based governance, enabling rapid innovation, transparency, and collective oversight. By contrast, intergovernmental processes provide legitimacy, accountability, and strategic direction, but often operate through slower, consensus-based mechanisms. The priority is to bridge these models by leveraging the agility and openness of technical communities while anchoring them in clear political mandates and public interest objectives. This also requires avoiding duplication, building on existing initiatives, and using this unique convening platform to ensure coherence, coordination, and effective political arbitration. Together, these priorities support an open, accountable, and strategically steered AI ecosystem.

The main gaps relate to value distribution, governance of open infrastructures, operational implementation, interoperability, and environmental impact. First, the issue of value capture and redistribution is not explicitly addressed. While economic implications are mentioned, there is no clear focus on ensuring that AI-generated value effectively returns to citizens and the broader economic fabric, rather than being concentrated among a limited number of actors. Second, the governance of open infrastructures remains underdeveloped. Open source is referenced, but without addressing its specific governance models, such as distributed contribution mechanisms or their articulation with public authorities. This is critical to align open innovation with political steering. Third, interoperability of governance frameworks is insufficiently specified. Beyond general compatibility, there is a need to ensure that different political, economic, and technical systems can effectively operate together in practice, especially in a fragmented international context. Fourth, operational implementation is not fully reflected. While a project-based approach and impact measurement are central in your response, the list does not adequately capture this emphasis on execution, evaluation, and scaling. Finally, environmental impact remains implicit. Given its strategic importance, it should be more explicitly recognised as a transversal issue shaping AI trajectories.

From the perspective of the research and higher education sector, governance gaps and advances in AI reveal a structural imbalance between resources and value creation. A first challenge lies in the overemphasis on compute and infrastructure, often seen as the main drivers of competitiveness. While essential, they overshadow the strategic role of data and software, which remain less visible but hold a major share of value creation potential. Without more proactive governance, these assets risk being concentrated among a few actors, limiting both autonomy and the return on public investment. This is reinforced by the data paradox. On the one hand, AI accelerates infobesity, increasing noise and complicating validation and trust. On the other hand, access to high-quality, curated data remains restricted, creating bottlenecks for research and innovation. This calls for treating data and software as strategic commons, with clearer governance and access frameworks. A third challenge concerns the fragmentation between research, projects, and policy, which limits the scaling and operationalisation of scientific advances. From the perspective of the research and higher education sector, governance gaps and advances in AI reveal a structural imbalance between resources and value creation. A first challenge lies in the overemphasis on compute and infrastructure, often seen as the main drivers of competitiveness. While essential, they overshadow the strategic role of data and software, which remain less visible but hold a major share of value creation potential. Without more proactive governance, these assets risk being concentrated among a few actors, limiting both autonomy and the return on public investment. This is reinforced by the data paradox. On the one hand, AI accelerates infobesity, increasing noise and complicating validation and trust. On the other hand, access to high-quality, curated data remains restricted, creating bottlenecks for research and innovation. This calls for treating data and software as strategic commons, with clearer governance and access frameworks. A third challenge concerns the fragmentation between research, projects, and policy, which limits the scaling and operationalisation of scientific advances. Overall, the key issue is not only access to resources, but the capacity to steer and operationalise them through structured initiatives. This requires strategic clarity and political will to move beyond competition on means toward tangible, shared value creation.

Its role should be to enable practical cooperation around shared priorities, grounded in a realistic and inclusive common base. The AI Dialogue can play a pragmatic role by helping define a common baseline for cooperation, focused on what is broadly acceptable and clearly aligned with the public interest. Rather than aiming for full harmonisation, its value lies in identifying a lowest common denominator that allows countries and stakeholders with different approaches to work together. This helps reduce friction, build trust, and support gradual convergence over time. The Dialogue should also focus on what works in practice. By prioritising concrete initiatives and shared use cases, it can promote cooperation through projects rather than abstract principles. This includes supporting the development and dissemination of reusable resources, such as open data, open source tools, and shared methodologies, which benefit a wide range of actors. In addition, it can help surface and align existing efforts, avoiding duplication and strengthening complementarities across initiatives. This leads to more efficient use of resources and greater collective impact. Finally, the Dialogue can reinforce political agency by keeping the focus on outcomes that serve the public interest, ensuring that cooperation is not driven solely by technical or sectoral priorities, but remains anchored in societal value.

The AI Dialogue should build on and connect with a combination of multilateral initiatives and open, community-driven ecosystems, which play complementary roles. On the intergovernmental side, Global Partnership on Artificial Intelligence (GPAI) provides an established platform linking policy, research, and implementation through working groups and concrete projects. It offers tested mechanisms for multi-stakeholder cooperation and can serve as a foundation for aligning priorities and scaling operational initiatives. In parallel, the Dialogue should connect with open source ecosystems, such as FOSDEM, which embody decentralised, bottom-up innovation. These communities play a critical role in shaping software, standards, and practices, often ahead of formal policy processes. They also contribute to transparency, accountability, and the diffusion of innovation through reusable tools and shared resources. The value of the AI Dialogue lies in its ability to bridge these two spheres. It should not duplicate existing frameworks, but rather create a space where political actors, experts, and technical communities can converge around a pragmatic common baseline focused on the public interest. Concretely, the Dialogue can: - connect policy discussions with operational projects and open infrastructures - identify and promote reusable resources (open data, open source tools, methodologies) - support coordination across initiatives, avoiding fragmentation - reinforce political steering, ensuring that cooperation remains aligned with societal objective. By linking structured initiatives like GPAI with dynamic ecosystems such as FOSDEM, the Dialogue can foster a more coherent, open, and action-oriented approach to international AI governance.

Different stakeholders should contribute to the AI Dialogue through concrete, project-based engagement, rather than purely consultative or declarative formats. Governments should provide clear political mandates and priorities, define expected outcomes, and ensure alignment with public interest objectives. Their role is to arbitrate, prioritise, and create the conditions for scaling successful initiatives. Research institutions, such as Inria, can contribute by structuring and delivering projects, mobilising scientific expertise, and developing reusable tools, methodologies, and open resources. They are well positioned to bridge knowledge production and operational implementation. Private sector actors should engage through use cases and co-development, contributing technical capabilities while aligning with shared principles, particularly around openness and value distribution. Civil society and technical communities should play a role in ensuring transparency, inclusiveness, and real-world relevance, as well as contributing to open source ecosystems and participatory approaches. In terms of format, the Dialogue should be structured around a portfolio of pilot projects, each linked to clear objectives and measurable impact indicators. Rather than broad thematic discussions, working groups should be task-oriented, with defined deliverables and timelines. A two-level structure could be effective: a political layer setting direction and validating priorities, and an operational layer focused on implementation and iteration. Strong feedback loops between these levels are essential. The Dialogue should also include regular review cycles, enabling comparison of results, identification of best practices, and reallocation of resources toward the most impactful initiatives. Finally, priority should be given to reusable outputs such as open data, open source tools, and shared frameworks, ensuring that projects generate lasting and scalable value. Overall, the Dialogue should function as a platform for coordination, delivery, and scaling of concrete initiatives.

Several voices remain under-represented in global AI governance, often due to structural barriers rather than lack of relevance. First, actors from the Global South, particularly public institutions, researchers, and SMEs, are insufficiently represented. This limits the diversity of perspectives on development trajectories and reinforces dependency on externally defined models. Inclusion requires capacity building through projects, access to infrastructure, and participation in co-development rather than consultation only. Second, operational communities are often overlooked. This includes engineers, open source contributors, and practitioners who shape systems in practice. Spaces such as FOSDEM illustrate their importance. Their inclusion can be strengthened by integrating them directly into project-based workstreams and governance discussions. Third, social sciences, humanities, and local knowledge holders remain under-mobilised. Their perspectives are critical to understanding societal, cultural, and economic impacts, yet they are often peripheral to technical discussions. Structured interdisciplinarity should therefore be embedded in project design and evaluation. Fourth, youth and emerging professionals are still too often included symbolically. Their meaningful participation requires dedicated roles in implementation, not only in consultation processes. Finally, civil society organisations and local public actors are unevenly represented, despite their role in ensuring accountability and alignment with public interest. To address these gaps, the AI Dialogue should move toward project-based inclusion mechanisms. This means funding and structuring multi-stakeholder projects where under-represented actors contribute directly to design and delivery. It also implies supporting access to open resources such as data, software, and methodologies, lowering entry barriers. Overall, inclusion should be operational, not declarative, ensuring that diverse actors are involved in producing outcomes, not only in discussing them.

From the perspective of a research centre contributing to capacity building, the objective is to support the effective implementation of orientations defined within the Global Policy Dialogue, by translating them into operational projects and deployable resources. To foster meaningful and dynamic engagement, the AI Dialogue should therefore prioritise formats centred on co-production and delivery, directly aligned with this implementation logic. A first approach is to organise project sprints, where small, multi-stakeholder teams work on concrete challenges and deliver tangible outputs such as tools, datasets, or policy prototypes. This anchors discussions in practice and accelerates implementation. Second, open labs or sandboxes can provide environments to test solutions in real conditions, using shared resources such as open data and open source tools. These spaces facilitate collaboration while generating feedback loops between experimentation and policy. Third, challenge-driven calls can mobilise a wide range of actors around specific priorities, including research institutions, SMEs, and civil society, enabling broader participation and surfacing scalable solutions. Fourth, embedded collaboration formats such as fellowships or secondments can strengthen knowledge transfer and ensure continuity between policy and implementation. Finally, formats inspired by open source communities, such as those seen at FOSDEM, can encourage bottom-up contributions through informal, collaborative, and technically grounded exchanges. All these formats should be integrated into a structured project portfolio, with clear objectives, timelines, and evaluation criteria, ensuring that engagement translates into measurable outcomes and supports the effective deployment of policy priorities.

GPAI-associated projects led by Inria follow a structured approach combining three levels: research and analysis, experimentation through pilot projects, and large-scale dissemination of tools and practices. This continuum ensures that governance is not limited to recommendations but is translated into deployable resources and capacities. As an example, VIADUCT addresses a central governance challenge: access to data for AI training. It develops practical frameworks for ethical and equitable data sharing, moving beyond abstract principles to propose operational alternatives to current models of data capture. This contributes to more balanced cooperation and reduces structural dependencies. Similarly, the student community applies this logic to the renewal of knowledge production and participation. It creates an international community of students directly involved in analysing AI transformations and contributing to policy-relevant outputs. Through structured collaboration, mentoring, and integration into GPAI processes, it enables learning by doing while producing insights grounded in real concerns. This initiative strengthens inclusiveness, anticipates emerging skills needs, and supports the diffusion of shared frameworks, while ensuring that future generations actively contribute to shaping AI governance in practice.