HerWILL

A successful first Global Dialogue on AI Governance should feel grounded in reality, not just well written. Right now, we are not lacking principles. We are lacking alignment between policy and how AI is actually built and used across the world. If this Dialogue can bring that gap into focus using real examples from different regions, that alone would be meaningful. It should also change who is shaping the conversation. Not in a symbolic way, but in a practical one. The people working with data, building models, and testing systems in different languages and contexts need to help define what good governance looks like. Otherwise, we will continue to design systems that work well in a few places and fail everywhere else. From our work at HerWILL, across six countries and now with 2,200 participants from 45 countries, one thing is very clear. Talent is not the issue. The real issue is the absence of structured pathways for people to move from learning to contributing to real systems. So success would mean a few concrete outcomes. A commitment to build multilingual datasets as shared infrastructure. Clear pathways for global talent to contribute to model development and evaluation. And a direct connection between governance discussions and workforce systems. If this Dialogue begins to shape who gets to build AI, not just how it is discussed, it will have done something meaningful. If not, it will sound good and change very little.

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

Our priorities come directly from what we are building and observing in real time. AI capacity building is central to our work. Over the past five years, HerWILL has engaged more than 7,500 individuals across six countries. Our current global program includes 2,200 participants from 45 countries. The focus is not just learning tools, but moving people toward real contribution through applied work, mentorship, and collaboration across borders. Social, economic, ethical, cultural, linguistic and technical implications of AI are not abstract for us. In our current program, participants are working on detecting online harm across English, Bengali, and Arabic. This includes identifying subtle toxicity such as sarcasm, coded language, and cultural references that are often missed by existing systems. What we are seeing is that harm is deeply contextual. Models trained primarily on English data do not transfer well across languages or cultures, which creates real gaps in safety and reliability. Open source software, open data and open AI models matter because data is the foundation of everything else. We are working with real, multilingual datasets collected from diverse environments. These datasets reflect how people actually communicate, not idealized versions of it. Making this kind of data more accessible and representative is critical if AI systems are to work globally. Transparency, accountability, and human oversight connect directly to this work. By involving participants in evaluating models and identifying bias and failure points, we are creating a layer of human judgment that is grounded in lived context, not just technical metrics. Across all of this, one thing is clear. The gap is not talent. It is who gets to participate in building and shaping these systems.

Yes. There is one issue that cuts across all of these themes but is still not being addressed directly - who is structurally positioned to build AI systems. Right now, most discussions focus on how AI should be governed once it exists. Much less attention is given to how participation in building these systems is distributed in the first place. That gap shows up everywhere. In the data that is used, in the languages that are prioritized, in the benchmarks that define performance, and in the people who are funded to do the work. From our experience at HerWILL, working with over 7,500 individuals across six countries and currently engaging 2,200 participants from 45 countries, the issue is not a lack of talent. It is that there are no clear pathways for people to move from learning into real contribution. As a result, large parts of the world remain users of AI systems, not contributors to them. This has direct implications across all themes. It affects safety, because systems are not tested in diverse contexts. It affects accountability, because those impacted are not part of evaluation. It affects open data, because most datasets are still concentrated in a few regions and languages. Another emerging issue is the role of data as infrastructure. Data is still treated as an input to models, rather than as a shared system that requires long term investment, stewardship, and governance. Without this shift, efforts around fairness, transparency, and inclusion will remain limited. If these two issues are not addressed, governance will continue to operate downstream, reacting to systems that were never built with broad participation in mind.

The gaps show up most clearly at the point where interest meets reality. We are seeing a surge of engagement across regions. People are motivated, capable, and ready to work on real AI problems. But when they try to move beyond learning, the system thins out quickly. There are very few entry points into meaningful work unless you are already inside a well resourced institution or geography. This creates a distorted pipeline. On one end, there is growing access to tools and training. On the other, there are concentrated hubs where models, datasets, and standards are actually built. What is missing is the bridge between the two. In practical terms, this affects outcomes. Talent from emerging regions is either underutilized or redirected into low value tasks. At the same time, systems are being developed without sufficient exposure to diverse contexts, which limits their reliability and relevance. There is also a coordination gap. Efforts around open data, model development, and capacity building are happening, but often in isolation. Without shared frameworks or collaboration across regions, progress remains fragmented. At the same time, the opportunity is significant. The level of readiness we are seeing now did not exist even a few years ago. If structured pathways are created, this talent can contribute meaningfully to model development, evaluation, and applied use cases across sectors. For organizations like ours, the challenge is not generating participation. It is sustaining momentum and converting it into long term contribution. That is where governance, if aligned properly, can make a real difference.

The AI Dialogue can play a meaningful role if it moves beyond being a convening space and becomes a connector. Right now, efforts are happening in parallel. Governments, academia, civil society, and industry are all working on pieces of the same problem, often without enough coordination. The Dialogue has the opportunity to bring these strands together in a way that is practical, not just diplomatic. One role it can play is aligning priorities across regions. Not by forcing consensus, but by identifying where there is already overlap and where collaboration would have immediate value. For example, shared work on multilingual data, model evaluation across contexts, or common approaches to risk assessment. It can also help create direct pathways between those shaping policy and those working on implementation. In our experience, there is a gap between high level governance discussions and the realities of building and testing systems. Bridging that gap would make cooperation more grounded and effective. Another important role is continuity. International conversations often lose momentum after the event. If the Dialogue can support ongoing collaboration, even in small, focused working groups, it would lead to more tangible outcomes over time. Finally, it can help expand the circle of contributors. Not just by inviting more voices, but by creating ways for those voices to participate in real work across borders. If the Dialogue can connect people, align efforts, and sustain collaboration beyond the room, it can move international cooperation from discussion to action.

There is no shortage of activity in this space. The challenge is that most efforts are still operating in silos. On one side, there are strong technical and research ecosystems. Initiatives connected to leading universities and research institutes are advancing model development, optimization, and evaluation. On another side, there are policy and governance platforms convened by international organizations that are shaping principles and frameworks. There are also growing open source and open data communities working to make AI more accessible. And across regions, smaller organizations are building local talent pipelines and applied programs tied to real world problems. All of these are valuable. What is missing is connection. The AI Dialogue can add value by linking these layers in a more intentional way. For example, connecting policy discussions with live programs that are generating data, testing models, and training new contributors. Or aligning open data efforts with those who are actually collecting and working with multilingual, real world datasets. It can also help reduce duplication. Many groups are solving similar problems in parallel without visibility into each other's work. Even light coordination would increase efficiency and impact. Another area is validation. There is a need for credible pathways that allow work coming out of smaller or distributed programs to be recognized and integrated into larger systems. The Dialogue can help create that bridge. Finally, it can provide continuity. Not by launching new initiatives, but by strengthening what already exists and helping it connect across borders. If it focuses on linking ecosystems rather than adding another layer, it can create real leverage.

Different stakeholders should contribute based on what they actually do, not just what they represent. Governments should bring real policy questions they are struggling with, not just prepared positions. Where are systems failing? Where are regulations unclear or difficult to implement? That level of honesty would make the Dialogue more useful. Academia and research institutions should contribute methods, benchmarks, and evidence. Not just papers, but what has been tested, what works, and what does not across different contexts. Industry should share where deployment meets friction. This includes limitations, tradeoffs, and unintended outcomes that are not always visible in controlled environments. Civil society and organizations like ours should bring lived experience from the ground. What happens when these systems are used across languages, cultures, and economic realities? Where do they break? For the structure, the Dialogue should move away from long panels and general statements. Small, focused working groups would be more effective. Each group should be centered around a specific problem, such as multilingual data, model evaluation across regions, or pathways from training to contribution. The goal should be to leave with a few concrete next steps, not a summary of viewpoints. There should also be continuity built in. The same groups should reconvene over time to track progress, refine approaches, and stay accountable. Finally, participation should not end with the event. There should be simple ways for contributors outside the room to stay involved through shared work, not just feedback. If the structure is practical and sustained, the Dialogue can move from discussion to coordination.

The gap is not just who is invited into the room. It is who is positioned to contribute to the work that shapes AI. Right now, global discussions are dominated by institutions that already have access to data, compute, and funding. What is underrepresented are the people working closest to where these systems are actually used. This includes contributors from non dominant language communities, practitioners outside major research hubs, and those building or testing systems in resource constrained environments. From our work, another group that is consistently missing is early career talent that is already capable but not yet recognized. These are individuals who can work with data, build models, and identify failure points, but have no pathway into the systems where those skills matter. As a result, they remain users of AI rather than contributors to it. There is also a gap in perspectives shaped by lived context. Harm, bias, and system behavior do not look the same across regions. Without those perspectives, governance frameworks risk being incomplete or misaligned. Including these voices requires more than invitations. There need to be structured ways to contribute. This could include involving distributed teams in dataset development, model evaluation, and real-world testing across languages and contexts. It also means recognizing work that happens outside traditional institutions and creating pathways for it to connect to larger systems. If participation remains limited to discussion, the gap will persist. If it extends into contribution, the quality of both systems and governance will improve.

Most global forums rely on panels and prepared remarks. They sound good, but they do not move the work forward. If the goal is real engagement, the format has to shift from talking to doing. One effective approach would be live working sessions built around specific problems. Small groups made up of government, research, industry, and practitioners working together on one issue such as multilingual data gaps or model evaluation across regions. The outcome should be a short set of actions, not a summary. Another format is "show, don't tell." Instead of presentations, participants bring a real case. What was built, where it failed, what was learned. This grounds the conversation and avoids abstract discussions. Cross region pairing would also be powerful. For example, matching teams working in different languages or contexts to compare how the same system behaves. This quickly exposes gaps that are not visible in a single environment. Short, timed interventions could replace long panels. Give participants five minutes to present one insight or one unresolved challenge, followed by direct responses from others in the room. Finally, there should be continuity beyond the event. The most productive groups should continue working together after the Dialogue with light structure and accountability. The goal is avhievable. Less performance, more problem solving. If participants leave having worked on something together, even in a small way, the Dialogue will be far more effective.

Several existing efforts offer concrete building blocks for more effective AI governance. The EU AI Act (Regulation 2024/1689) matters because it moves beyond principles into an enforceable, risk based framework, the first of its kind globally. It classifies AI systems into four tiers and ties obligations to use cases, not vague ethical claims. Fines for non compliance can reach €35 million or 7% of global annual turnover. Governance becomes operational. At the same time, its implementation will test whether regulatory ambition can keep pace with rapidly evolving systems. The NIST AI Risk Management Framework (released January 2023) gives institutions a practical way to identify, map, measure, and manage AI risk across the full lifecycle. In July 2024, NIST extended it with a Generative AI Profile (NIST-AI-600-1), addressing risks such as confabulation, data privacy, and harmful bias. It is one of the more usable tools available today, though its impact depends on how widely it is adopted beyond well resourced institutions. The UNESCO Recommendation on the Ethics of AI, adopted by all 193 member states, sets a strong global baseline. Its value lies in implementation tools such as the Readiness Assessment Methodology and Ethical Impact Assessment. But like many global frameworks, the challenge is translation into practice, especially in contexts with limited institutional capacity. The African Union Continental AI Strategy (adopted July 2024) offers a different model. It ties governance to infrastructure, local datasets, and talent development. It starts from local needs rather than importing external frameworks, which makes it more grounded and potentially more sustainable. At the practice level, Masakhane, a grassroots community of over 2,000 African NLP researchers, demonstrates what community led dataset development looks like. They are building open, African language datasets that major model developers have largely overlooked. This kind of work is critical, but remains underfunded and disconnected from mainstream AI development. What the AI Dialogue can add is not another framework. It should address the fragmentation across these efforts. Legal models, risk tools, ethical standards, and community driven data systems are evolving in parallel, often without connection. That fragmentation is not neutral. It shapes who has influence and who does not. The value of the Dialogue would be to align these layers into something that can function as shared global practice.