Polytechnique Montreal and MLCommons

To bridge the gaps between high-level principles and operational implementation, a successful first Global Dialogue on AI Governance is required to deliver concrete, actionable, and measurable outcomes. In the following, I will break down my suggestions into five highlights: First, the Dialogue should establish a shared understanding of AI risk, trustworthiness, and sustainability, supported by common terminology and initial agreement on measurable indicators (e.g., safety, robustness, energy consumption, and lifecycle impacts). This would help reduce fragmentation across existing governance approaches. Second, it should initiate the development of global reference frameworks or benchmarks for AI systems, enabling consistent evaluation of safety, transparency, and environmental impact across jurisdictions and sectors. Such benchmarks would support evidence-based policymaking and foster accountability. Third, the Dialogue should catalyze multi-stakeholder collaboration mechanisms, including partnerships between governments, academia, industry, and open-source communities. These collaborations are essential to ensure that governance frameworks are both technically grounded and practically deployable. Fourth, the Dialogue should highlight inclusive and equitable participation, particularly by amplifying voices from underrepresented regions and supporting capacity-building initiatives that enable meaningful engagement in AI governance. Finally, a key outcome would be a clear roadmap for future Dialogues, including priority themes, expected deliverables, and mechanisms to translate scientific insights into policy-relevant actions. Overall, success will depend on the ability to move from discussion to implementation-oriented governance, where trust in AI systems is continuously assessed through transparent, collaborative, and globally aligned practices.

• Safe, secure and trustworthy AI
• Open-source software, open data and open AI models
• Transparency, accountability, and human oversight
• Interoperability of governance approaches

My selection is based on my expertise in the trustworthiness of AI safety-critical systems, which hugely involve applied research on cybersecurity and cyber-physical AI systems. Therefore, my selected priorities reflect the need to advance AI governance toward practical, interoperable, and evidence-based implementation. Safe, secure, and trustworthy AI is foundational, as increasing system complexity introduces risks related to robustness, security vulnerabilities, and unintended behaviors. Addressing these risks requires systematic approaches across the AI lifecycle. Transparency, accountability, and human oversight are critical to ensuring that AI systems remain aligned with societal values. However, transparency must go beyond documentation to include auditable and measurable mechanisms, enabling continuous evaluation of system behavior and risks. Open-source software, open data, and open AI models play a key role in fostering transparency, reproducibility, and global participation. Open ecosystems enable independent verification, accelerate innovation, and reduce asymmetries between regions, particularly benefiting capacity-building efforts in under-resourced contexts. Interoperability of governance approaches is essential to address the current fragmentation of AI policies and standards. Without alignment, differing regulatory frameworks may create barriers to collaboration, reduce effectiveness, and increase compliance complexity. Interoperable frameworks can support shared benchmarks, mutual recognition, and coordinated responses to emerging risks. Together, these priorities support a governance approach that is technically grounded, globally coherent, and inclusive, while enabling practical mechanisms to assess and improve trust in AI systems over time.

Indeed, several cross-cutting and emerging issues merit greater attention to ensure comprehensive and forward-looking AI governance. First, the sustainability (environmental and energy) impact of AI systems is an increasingly urgent concern. The rapid growth of large-scale models and infrastructure has significant implications for energy consumption and carbon emissions. Integrating sustainability metrics into AI governance frameworks is essential to align AI development with global climate objectives. Second, the AI supply chain and dependency ecosystem represents a critical but under-addressed dimension. AI systems rely on complex layers of data, models, software libraries, and infrastructure. Vulnerabilities or biases introduced at any stage can propagate across the system. Governance efforts should therefore adopt a lifecycle and ecosystem perspective, including supply chain transparency and risk management. Third, we cannot neglect the current rise of autonomous and agentic AI systems that introduce new governance challenges related to decision-making autonomy, coordination, and accountability. These systems blur traditional boundaries between tools and actors, requiring updated frameworks for oversight and responsibility. Finally, there is a need for continuous monitoring and adaptive governance mechanisms, rather than static compliance models. AI systems evolve over time, and governance approaches must support ongoing evaluation, feedback, and improvement. Addressing these cross-cutting issues will strengthen the ability of global AI governance to remain robust, adaptive, and aligned with both societal and environmental priorities.

In Canada, and particularly in Montréal as a global hub for AI research and innovation, governance gaps present both significant challenges and strategic opportunities. A key challenge lies in the fragmentation of governance approaches across jurisdictions and sectors. While Canada has taken leadership through initiatives such as responsible AI frameworks and standards development, there remains limited alignment between regulatory efforts, technical practices, and real-world deployment. This creates uncertainty for organizations and slows the operationalization of trustworthy AI. Another critical issue is the lack of standardized, measurable frameworks for AI safety, transparency, and accountability. In research-intensive environments such as Polytechnique Montréal, this gap limits the ability to systematically evaluate risks across the AI lifecycle, including vulnerabilities in data, models, and software dependencies. It also constrains reproducibility and independent auditing, particularly in complex or proprietary systems. The rapid growth of AI systems also raises concerns related to security and robustness, including exposure to adversarial threats and misuse. These risks are amplified in open and collaborative ecosystems, where dependencies across open-source components and datasets introduce additional layers of complexity. At the same time, Montréal's strong academic ecosystem and Canada's commitment to open science create important opportunities. The region is well-positioned to advance open, transparent, and interoperable governance approaches, leveraging collaboration between academia, industry, and public institutions. Open-source ecosystems, in particular, can support benchmarking, shared evaluation tools, and global participation. Finally, there is a growing opportunity to integrate sustainability considerations into AI governance, aligning AI development with climate objectives through energy-aware practices and lifecycle monitoring. Addressing these gaps can position Canada as a leader in operational, inclusive, and globally aligned AI governance.

The Global Dialogue on AI Governance can play a pivotal role as a neutral, multi-stakeholder coordination platform that bridges fragmented national, regional, and sectoral approaches to AI governance. First, it can facilitate the development of a shared understanding of AI risks, trustworthiness, and accountability, grounded in scientific evidence. By aligning terminology, concepts, and evaluation approaches, the Dialogue can reduce ambiguity and support more coherent global governance. Second, the Dialogue can promote interoperability across governance frameworks, enabling coordination between existing regulatory efforts while respecting different national contexts. This is particularly important to avoid regulatory fragmentation that may hinder innovation, collaboration, and cross-border deployment of AI systems. Third, it can act as a platform to translate technical knowledge into policy-relevant insights, ensuring that governance discussions are informed by the latest advances in AI safety, security, and system design. This includes integrating perspectives from academia, industry, and the technical community. Fourth, the Dialogue can strengthen inclusive participation and capacity-building, especially by supporting engagement from underrepresented regions and stakeholders. This will help ensure that AI governance reflects diverse priorities and does not exacerbate global inequalities. Finally, the Dialogue can catalyze practical cooperation mechanisms, such as joint initiatives, shared benchmarks, and collaborative platforms for evaluating AI systems. By focusing on implementation-oriented outcomes, it can move international cooperation from principles to actionable and measurable governance practices.

The Global Dialogue should connect existing international, multi-stakeholder, and technical initiatives advancing AI governance, standards, and best practices. These include global policy frameworks (OECD AI Principles, UNESCO Recommendation on AI Ethics); standardization efforts (ISO/IEC, IEEE); technical and benchmarking initiatives (MLCommons, open-source communities, AI safety research); and national/regional strategies (Canada, European Union). However, these efforts remain fragmented, limiting their global impact. The Dialogue's value lies in serving as a convergence layer to: Coordinate policy, standards, and technical communities. Promote interoperability and mutual recognition of governance approaches. Highlight best practices and scalable models (e.g., open-source platforms). Support shared evaluation frameworks and indicators. Crucially, it bridges principles and implementation by connecting high-level governance with operational tools and real-world use cases. By building on existing efforts while enhancing coherence and inclusivity, the Dialogue can strengthen international cooperation and accelerate globally aligned, evidence-based AI governance.

Different stakeholders can contribute to the AI Dialogue through complementary and clearly defined roles, supported by a structured and outcome-oriented format. Member States: Provide policy priorities, regulatory perspectives, and coordination mechanisms. Private sector: Share implementation practices, risk management approaches, and technical constraints. Academia and technical community: Contribute scientific evidence, benchmarks, and evaluation methodologies. Civil society: Ensure accountability, ethical considerations, and representation of societal impacts. Open-source communities: Enable transparency, reproducibility, and shared tools for governance. To maximize effectiveness, the Dialogue should adopt a multi-layered structure: Evidence Layer: Presentation of key findings from the Scientific Panel and technical experts. Policy Translation Layer: Multi-stakeholder discussions to translate evidence into governance implications. Action Layer: Identification of concrete outputs (e.g., frameworks, partnerships, pilot initiatives). The format should include short, focused interventions, moderated roundtables, and thematic breakout sessions aligned with priority areas. A clear emphasis should be placed on actionable outcomes, with documented recommendations and follow-up mechanisms. This structure would ensure that contributions are balanced, technically grounded, and directly linked to policy-relevant actions.

Several key voices and perspectives remain underrepresented in global AI governance discussions. First, stakeholders from the Global South often face barriers to participation due to limited resources, infrastructure, and access to technical expertise. Their perspectives are essential to ensure that AI governance frameworks are equitable and context-sensitive. Second, open-source communities and independent developers are frequently overlooked, despite their central role in building and maintaining widely used AI systems and infrastructure. Their inclusion is critical for transparency and practical implementation. Third, interdisciplinary experts, including those working at the intersection of AI, sustainability, and socio-technical systems, are underrepresented. This limits the ability to address cross-cutting issues such as environmental impact and systemic risks. Fourth, affected communities and end-users, particularly those disproportionately impacted by AI systems, are often insufficiently represented in decision-making processes. To improve inclusion, the Dialogue should: Provide targeted support for participation (e.g., funding, remote access, language accessibility) Integrate open calls for contributions from diverse stakeholders Include community-led sessions and participatory formats Ensure regional balance and representation in panels and working groups Strengthening inclusion will enhance the legitimacy, relevance, and global applicability of AI governance outcomes.

To foster meaningful and dynamic engagement, the AI Dialogue should adopt interactive, outcome-driven formats that go beyond traditional panel discussions. First, thematic breakout sessions can enable focused, small-group discussions on priority areas, encouraging deeper exchanges and more concrete outputs. Second, multi-stakeholder "policy labs" or co-design workshops can bring together policymakers, technical experts, and practitioners to collaboratively develop governance solutions, frameworks, or use cases in real time. Third, scenario-based simulations can be used to explore AI risks, trade-offs, and policy responses in realistic contexts. This approach helps bridge the gap between theory and practice. Fourth, case study showcases from different regions and sectors can highlight practical challenges, lessons learned, and scalable solutions, including examples from open-source ecosystems. Fifth, interactive digital platforms (e.g., live polling, collaborative documents, hybrid participation tools) can broaden participation and capture diverse inputs, including from remote stakeholders. Finally, the Dialogue should incorporate clear synthesis and feedback mechanisms, ensuring that discussions are translated into actionable insights and shared outputs. These formats would promote active participation, cross-sector collaboration, and practical problem-solving, making the Dialogue more inclusive, engaging, and impact-oriented.

Several existing policies, practices, and platforms provide concrete and scalable approaches to effective AI governance. First, risk-based governance frameworks, such as those reflected in national and regional strategies, offer structured approaches to categorizing AI systems based on their potential impact. These frameworks support proportionate oversight, enabling stronger requirements for high-risk systems while maintaining flexibility for innovation. Second, AI risk management and evaluation frameworks-including lifecycle-based approaches-promote continuous assessment of systems across data, model development, deployment, and monitoring. These practices enable governance to move beyond static compliance toward ongoing accountability and adaptive oversight. Third, technical benchmarking and evaluation platforms play a critical role in operationalizing governance. Shared benchmarks for safety, robustness, and performance enable consistent and transparent evaluation of AI systems across contexts. Open and collaborative benchmarking initiatives are particularly valuable for fostering reproducibility and evidence-based policymaking. Fourth, open-source ecosystems provide practical mechanisms for transparency, auditability, and global collaboration. By enabling access to models, datasets, and tools, open approaches support independent verification, accelerate innovation, and reduce barriers to participation-especially for under-resourced communities. Fifth, multi-stakeholder governance models-bringing together governments, academia, industry, and civil society-have proven effective in aligning technical development with societal values. These models support knowledge sharing, standard-setting, and coordinated responses to emerging risks. Finally, integrating sustainability considerations into AI governance-such as energy and carbon tracking-represents an emerging best practice aligned with global climate objectives. Together, these approaches demonstrate that effective AI governance requires a combination of policy frameworks, technical tools, and collaborative ecosystems, enabling measurable, transparent, and adaptive oversight of AI systems.