MLCommons

For the first Global Dialogue on AI Governance to be a real success, it needs to deliver results that aren't just nice in theory but are implementable with real-world practicality and safeguards, trackable, and open to everyone. First, everyone involved should come away with a shared sense of what really matters when it comes to AI governance—things like safety, transparency, sustainability, and security—while also keeping in mind that different countries have different needs and capabilities. Second, the Dialogue should lead to clear, actionable recommendations or guidelines that can actually be put to use in different settings. That means figuring out common ways to evaluate AI systems—through things like benchmarks, risk assessments, and monitoring over their entire lifecycle—using open and repeatable methods. Open source ecosystems can be a big help here, making it easier to evaluate systems transparently, share tools, and collaborate on governance. Third, success also means making sure everyone gets a seat at the table. The Dialogue needs to actively bring in voices from underrepresented parts of the world, especially the Global South. And capacity-building shouldn't be an afterthought—it has to be a core part of the outcome, including better access to infrastructure, knowledge, and open technologies. Fourth, this can't be a one-off event. To keep the momentum going, we should set up ongoing working groups, platforms for sharing knowledge, and ways for stakeholders and the Scientific Panel to keep feeding back into the process. Finally, it's crucial to connect what comes out of the Dialogue with broader global goals—like the Sustainable Development Goals. In particular, as AI systems use more and more energy and resources, we have to make sure environmental sustainability is part of the conversation. In short, the real measure of success will be how well we connect high-level principles with real-world action—through inclusive, transparent, and collaborative approaches.

• Safe, secure and trustworthy AI
• Open-source software, open data and open AI models
• Transparency, accountability, and human oversight
• AI capacity-building

The choices I've made come down to one thing: pushing for practical, measurable, and globally inclusive ways to govern AI. For me, safe, secure, and trustworthy AI is at the heart of everything I do-whether that's security work, vulnerability analysis, or risk assessment across different AI systems. If we want people to trust AI, especially in critical or large-scale deployments, we have to make sure it's robust and resilient first. Transparency, accountability, and human oversight aren't just buzzwords-they're how you actually make AI trustworthy. My research focuses on a few key areas: reproducible evaluations, benchmarks, and explainability. Why? Because without those things, you simply can't audit AI systems or hold them accountable once they're out in the real world. That's where open-source software, open data, and open AI models come in-they're total game changers. They provide the kind of infrastructure we actually need for transparent evaluation, shared benchmarks, and working together on governance. Plus, they let people independently verify AI systems, which is key to building trust globally. We also can't overlook AI capacity-building. Every region deserves a real seat at the table when it comes to developing and governing AI. That means providing access to infrastructure, tools, and knowledge-especially for underrepresented communities and the Global South. Open-source ecosystems can help close the gap by lowering barriers and making it easier to share what we know. Taken together, these priorities point to a governance approach that's not just principled, but also technically sound, scalable, and inclusive-exactly what we need to make sure AI benefits everyone.

The themes already mentioned cover a lot of the key areas in AI governance, but there are a few cross-cutting and emerging issues that really deserve more attention. First, the environmental impact of AI isn't getting enough focus. Large AI systems consume huge amounts of energy and resources, which puts real pressure on both the environment and infrastructure. Any serious governance framework needs to include clear, lifecycle-based sustainability metrics-looking at energy use, carbon emissions, and how efficiently resources are being used. Second, we urgently need standardized ways to measure and evaluate AI systems. A lot of governance principles-like safety, fairness, and transparency-sound great on paper. But without clear benchmarks and concrete metrics, they're really hard to put into practice. If we want to turn those high-level ideals into something that can actually be enforced and compared across different systems, we need open, reproducible ways of evaluating AI. Third, we need to start paying more attention to supply chain security and model provenance. Current AI systems often rely on a messy tangle of datasets, pre-trained models, and third-party pieces. To reduce risks like data poisoning, tampering, or hidden vulnerabilities, we have to make sure those dependencies are secure, traceable, and something we can actually trust. Fourth, governance can't stop once an AI system is deployed. These systems change over time and operate in dynamic environments, so we need ongoing oversight-not just a one-time check before launch. Finally, open source ecosystems are often talked about as if they're just nice to have, but they're actually a critical piece of governance infrastructure. Open platforms enable transparency, independent auditing, and global collaboration. They're a key part of making AI governance both accountable and inclusive. Tackling these cross-cutting issues would go a long way toward keeping global AI governance efforts adaptable, measurable, and grounded in the real-world challenges of deploying AI.

In my worldview, which spans academic research, open-source communities, and global AI infrastructure, the biggest gaps in governance show up when high-level principles don't match up with what actually happens on the ground. One major issue is that we don't have standardized, measurable ways to evaluate AI systems. Everyone talks about safety, transparency, and accountability, but there aren't many benchmarks that actually work across different systems and contexts. That leads to fragmentation—different tools, platforms, and places all doing their own thing—which makes it really hard to compare AI systems or enforce any kind of consistent governance. Another issue is supply chain security—basically, not really knowing where AI models come from. These days, most AI development relies on pre-trained models, open datasets, and third-party parts. But we often have very little idea where they originated, whether someone's messed with them, or what hidden vulnerabilities they might have. That's a serious risk, whether you're working in the public sector or for a private company. Then there's the capacity gap. Places like North America have advanced infrastructure and expertise, but many parts of the Global South still struggle to access compute power, datasets, and technical know-how. If we don't put real effort into targeted capacity-building, these gaps are only going to get wider. That said, there are also some promising opportunities. Open-source ecosystems are becoming a powerful way to enable transparent, collaborative governance—through shared benchmarks, reproducible tools, and community-driven standards. Here in Canada and similar research environments, I'm seeing growing momentum to build sustainability, security, and accountability right into AI development workflows. So in the end, tackling these governance gaps is a real chance to move toward AI governance that's more measurable, more interoperable, and more inclusive—grounded in open collaboration and the realities of how AI gets built and used in practice.

The AI Dialogue has a real chance to move the needle on international cooperation by connecting big-picture principles with actual, workable governance approaches. First, it can help get everyone on the same page when it comes to shared frameworks and standards—especially around evaluating AI systems. If we can agree on common ways to benchmark safety, transparency, and performance, we can reduce the chaos of fragmented rules across different countries and make governance more interoperable. Second, the Dialogue can serve as a true meeting ground for governments, researchers, industry, and open-source communities. That mix matters because policy needs to be grounded in what's actually possible technically—and governance needs to be both realistic and effective on the ground. Third, it can boost capacity-building and knowledge sharing, particularly for regions that are often left out. When people have better access to open tools, datasets, and infrastructure, it helps level the playing field, both in building AI and in deciding how it should be governed. Fourth, the Dialogue can really build transparency and trust by sticking with open, repeatable practices. Things like open-source ecosystems, shared benchmarks, and collaborative evaluation efforts make it possible for people to independently check AI systems. And that's how you strengthen accountability. Finally, the Dialogue can help governance keep up with fast-moving tech by creating ongoing feedback loops between research, real-world deployment, and policy. Setting up working groups, technical exchanges, and long-term collaboration will be key—so the impact doesn't fade after the annual meetings end. By bringing technical innovation together with inclusive, coordinated policy efforts, the AI Dialogue can help shape AI governance that's globally consistent, practical, and worthy of trust.

We shouldn't start from scratch with the AI Dialogue. There's already a whole ecosystem of initiatives out there doing important work on AI governance, standards, and capacity-building—and the Dialogue needs to build on that. Think about the OECD AI Principles and UNESCO's recommendation on the ethics of AI. They've already given us widely agreed-upon norms. Then there's the Global Partnership on AI-GPAI for short, wth the aim of bringing different people and groups together to work on AI. And on the technical side, organizations like ISO/IEC JTC 1/SC 42 and IEEE are busy figuring out standards for AI. When it comes to actually evaluating AI, efforts like MLCommons and open-source communities under the OpenInfra Foundation and the Linux Foundation provide real, hands-on tools for measuring how well AI performs, how reliable it is, and how sustainable it might be. The Dialogue should also tap into newer conversations around governance and safety, like international talks on AI risk management, model evaluation, and secure AI supply chains, plus open datasets and model repositories that support AI development worldwide. So what's the real added value of the AI Dialogue? It has the potential to take all these scattered pieces and pull them together into something that actually makes sense and includes everyone. Right now, a lot of the work being done is focused on specific regions, industries, or technical niches. But the Dialogue could be different, a neutral space where policy, technical standards, and what's actually happening on the ground can finally start to line up. It can also help close the gap between high-level principles and on-the-ground implementation, by promoting evaluation methods that work well together, encouraging open and reproducible tools, and helping different regions share what they've learned. And by making inclusivity and capacity-building a real priority, the Dialogue can ensure that global south aren't just passive in the discussion, but actually helping steer where global AI governance heads from here.

For people to really participate effectively in the AI Dialogue, everyone needs a clear role—and the whole event needs to be structured in a way that's inclusive and easy to follow. Governments, for example, can help by aligning their national policies, sharing what's worked (and what hasn't) in regulation, and pointing out where international cooperation is most needed. Academics and the tech community can bring evidence-based insights to the table, help develop evaluation methods, and chip in on benchmarking, safety, and sustainability assessments. Private sector folks should share real-world practices, how they approach risk management, and what they've learned from deploying AI systems at scale. Civil society organizations need to be a key actor, since they represent public interests, raise ethical red flags, and make sure governance frameworks stay inclusive and respectful of people's rights. Open-source communities can contribute practical tools, datasets, and reproducible frameworks that make transparency and independent verification possible. To get the biggest impact, the AI Dialogue should have a few different layers: Plenary sessions to agree on big-picture priorities and keep everyone politically aligned. Thematic working groups (on topics like safety, transparency, sustainability, and capacity-building) where tech and policy experts can roll up their sleeves and produce actionable outputs. Technical organizations/communities focused on benchmarking, open tools, and real implementation practices—hands-on collaboration in other words. Ongoing engagement like online platforms and regular check-ins, so the momentum doesn't die between annual meetings. The outcomes should aim for practical guidelines, shared benchmarks, and policy reports that come with clear steps for implementation. A structure like this would help make sure the Dialogue stays inclusive, grounded in real technical know-how, and actually capable of turning high-level talk into concrete, measurable results.

There are still quite a few important voices that don't get enough say in global conversations about AI governance. That's a problem—not just for fairness, but because it makes the resulting rules and frameworks less effective. First, people from the Global South are often left out, especially those from regions that lack computing power, data resources, and funding. Their input really matters if we want governance approaches to work across different economic and social contexts, and not accidentally make existing inequalities worse. Second, open-source and technical communities aren't always brought into policy talks, even though they're the ones actually building and running much of today's AI infrastructure. Their hands-on experience is crucial for making sure governance frameworks actually work in the real world. Third, small and medium-sized businesses and grassroots innovators tend to get overshadowed by big tech companies. But they face their own unique struggles when trying to comply with new AI rules, and they often have smart ideas about solutions that are both scalable and accessible. Fourth, interdisciplinary researchers—especially those working at the crossroads of AI, sustainability, and cybersecurity—are still trying to find their place in policy circles. They can offer valuable insights on risks and trade-offs that cut across different areas. So, how do we fix these gaps? The AI Dialogue needs to make inclusiveness a real priority. That means offering financial support to help people participate, holding regional consultations, and using hybrid formats that make it easier to join in. Open platforms and open-source tools can also help lower the barriers. And finally, creating clear pathways for technical contributions—like working groups or open calls for input—can make sure that all this diverse expertise is systematically woven into the governance process.

If we want the AI Dialogue to truly spark meaningful and dynamic engagement, we need to move beyond the usual panel discussions and try formats that are more interactive and focused on real outcomes. For starters, we could set up thematic working labs—or "policy–technical sprints"—where policymakers, researchers, and practitioners roll up their sleeves and work together on concrete outputs like draft guidelines, evaluation frameworks, or benchmarking approaches. These sessions would go a long way toward closing the gap between high-level policy and actual implementation. Live demos and benchmarking showcases would be encouraged too. People could actually get their hands on open-source tools, datasets, and evaluation platforms, which really matters for domains such as AI safety, transparency, and sustainability. You can talk about that stuff all day, but you really need to see it in action to get it. We should also set up multi-stakeholder roundtables, but make sure they're actually structured to solve problems, like using scenario-based discussions or real-world case studies. That setup gets people talking more deeply about actual challenges, like rolling out AI in critical sectors or dealing with cross-border governance issues. And let's be real—not everyone can show up in person. So we need open consultation platforms and ways for people to participate asynchronously. Think open calls for input, shared online repositories, feedback forms, that kind of thing. That way, we can actually hear from voices all over the world, especially from regions that usually get left out of these conversations. And finally, we shouldn't forget about cross-disciplinary conversations—where we intentionally bring together technical experts, policymakers, and civil society. Those kinds of discussions help uncover a much wider range of viewpoints and bring important trade-offs to the surface, like the balance between security, sustainability, human rights, and other priorities. Interactive sessions should be mixed with hands-on activities in an inclusive formats, the AI Dialogue becomes a lot more engaging. People won't just show up and sit through presentations—they'll actually take part and help deliver real, measurable results.

There are already a number of policies, practices, and platforms out there that show what effective and actionable AI governance can look like. On the policy side, you've got things like the OECD AI Principles and the UNESCO Recommendation on the Ethics of AI. These are widely used frameworks that focus on transparency, accountability, human rights, and making sure everyone is included. Alongside them, the NIST AI Risk Management Framework offers a pretty practical way to identify, assess, and reduce risks throughout an AI system's entire lifecycle. If you look at it from a more technical, hands-on angle, benchmarking efforts like MLCommons show how open, standardized benchmarks can help assess AI systems in a way that's transparent and reproducible. These kinds of approaches take big-picture principles and turn them into things you can actually measure. Open-source platforms are also really important. Take the ecosystems supported by the OpenInfra Foundation, or analytics projects (Linux foundation or similar) that provide tools to track software sustainability, community health, and system reliability. These platforms make collaborative governance possible by accessibility to shared infrastructure and transparent metrics. We're also seeing newer practices emerge around AI supply chain security and model provenance, like documenting datasets, using model cards, and following reproducibility standards. These are helping make AI development more traceable and accountable. And on the environmental front, carbon tracking and sustainability tools are now being incorporated into machine learning workflows; energy monitoring and reporting frameworks. That's a big step toward addressing the environmental impact of AI systems. Taken together, these examples show that good AI governance really needs a mix of policy frameworks, technical standards, open platforms, and measurable practices-backed up by collaboration across different sectors and regions.