Google

To ensure the first Global Dialogue on AI Governance is a success, the dialogue's outcome should focus on building accountability infrastructure, not just declarations. I would outline three suggested outcomes that I believe would drive our success in AI Governance. First, establish a global living glossary for key definitions. Terms like "Safe AI", and " Human oversight" have varied technical and political definitions depending on who is speaking. Without agreed and standardized definitions, Governance frameworks would be impossible to reach because the goal post can easily be shifted. Secondly, we need consensus and a formal acknowledgment that data quality is foundational for AI Governance. AI systems are only as trustworthy as the data on which they are built. Any governance framework that evaluates models without addressing the integrity, representativeness, and documentation of training and operational data is auditing the output while ignoring the input. The Dialogue should call for data quality standards to be embedded in international AI governance frameworks. Lastly, we need meaningful representation from the Global South and underrepresented communities as co-architects. Many of the populations most exposed to AI-driven decisions, in public services, financial access, and labor markets, are least represented in the rooms where governance norms are being set. A successful first Dialogue begins to structurally correct that imbalance.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight

Safe, secure, and trustworthy AI is a core prerequisite, especially given the prevalence of AI tools in our day-to-day lives. Many users are sharing private and intimate data with AI models. Trustworthiness is not a property of the AI/ML model alone; it is a function of the entire data-to-deployment pipeline. As a Technical Program Manager at Google working on data quality programs for large-scale infrastructure, I have seen how upstream data-integrity failures silently propagate into downstream decisions. Safety frameworks must reach into data governance, not just model evaluation. AI capacity-building is where global equity is either built or abandoned. The countries and communities with the least AI infrastructure are often those most affected by AI systems built elsewhere. Capacity-building is rapidly increasing, and decisions on where this infrastructure is built must go beyond access to tools. It should include the technical and institutional capacity to audit, contest, and govern AI systems deployed within a country's borders. Social, economic, ethical, cultural, linguistic, and technical implications are most frequently treated as soft and therefore deprioritized. Language and cultural representation gaps in training data are not peripheral concerns; they are the mechanism by which AI systems systematically underserve entire populations. This area must be treated with the same urgency as technical safety. Transparency, accountability, and human oversight close the loop. Governance frameworks without enforcement mechanisms are voluntary guidelines. Accountability requires that someone, a person, an institution, or a government, can be held responsible when an AI system causes harm. Transparency is the precondition for that accountability. Human oversight, particularly for high-stakes decisions, must be treated as non-negotiable rather than aspirational. Together, these four themes form a governance stack: data integrity feeds safety; capacity enables participation; cultural inclusion shapes equity; and transparency enables accountability.

Two emerging issues deserve dedicated key attention in the Dialogue. Data quality and data governance as an AI governance layer. The current framework focuses primarily on model behavior, deployment risk, and user rights. It does not adequately address the governance of the data systems that feed AI models. Data quality rules, lineage documentation, representativeness standards, and data provenance are governance instruments, but they remain classified as technical infrastructure rather than policy concerns. This creates a structural gap: organizations can satisfy AI governance requirements while deploying models trained on incomplete, biased, or undocumented datasets. The Dialogue should establish a working stream specifically on data governance standards as a precondition for trustworthy AI, including minimum documentation requirements for datasets used in high-stakes AI applications. The representation pipeline problem. AI governance discussions frequently address bias as an output problem. The deeper issue is a participation pipeline problem: who is building these systems, who is setting the quality and safety benchmarks, and whose lived experience is treated as the default against which AI performance is measured. This is not only an ethical question; it is also a question of technical accuracy. Systems built without diverse practitioner perspectives are more likely to have undetected failure modes for underrepresented populations. Governance frameworks should address workforce representation in AI development as a technical governance concern, not merely a diversity and inclusion objective. These two issues are cross-cutting because they affect every thematic area already identified. Data quality shapes safety, capacity, transparency, and equity simultaneously. Representation in the development pipeline determines whose risks are visible and whose remain undetected. Neither is adequately addressed by any single thematic cluster, which is precisely why they require explicit, cross-cutting recognition in the Dialogue's structure.

My perspective spans two contexts: the United States, where I work as a Technical Program Manager at Google, building data-quality programs for large-scale infrastructure, and Nigeria, where I have led data-driven workforce initiatives that impact hundreds of thousands of young people. The governance gaps in my four priority areas manifest differently across these contexts but share a common root: the systems making consequential decisions about people's lives are not accountable to the people most affected by them. In the United States, the absence of a federal AI governance framework has produced fragmentation. Texas's Responsible AI Governance Act represents a meaningful state-level step, but organizations operating nationally face inconsistent compliance obligations and little incentive to proactively build trustworthiness into their data pipelines. Within the tech sector specifically, data quality standards are often treated as engineering concerns rather than governance instruments. This means AI systems can satisfy internal review processes while trained on undocumented, unrepresentative, or historically biased datasets. There is no current mechanism requiring organizations to demonstrate data integrity as a precondition for deployment in high-stakes contexts. In Nigeria and across the broader Global South, AI deployment in public services, financial access decisions, and labor markets is accelerating, often driven by external vendors and platforms. What has not kept pace is the institutional infrastructure for auditing, contesting, or governing those systems locally. Regulatory frameworks are nascent, technical expertise is concentrated in the private sector, and civil society has limited visibility into how these systems are making decisions that affect citizens. The result is a governance deficit that tracks the deployment curve but trails it dangerously: the populations with the most exposure to AI-driven decisions have the least structural ability to interrogate them. The opportunity, however, is significant. Nigeria's large youth population, growing technology sector, and existing data governance frameworks in financial services provide a foundation for leapfrogging fragmented Western models toward rights-based, data-quality-anchored AI governance from the outset.

The Dialogue can advance international cooperation in three ways. First, by functioning as a body for policy convergence. Bilateral and regional AI agreements are proliferating, but they are not interoperable. The EU AI Act, the US Executive Orders on AI, and the African Union's Continental AI Strategy each reflect legitimate governance priorities, yet organizations operating across these jurisdictions face contradictory compliance obligations. The Dialogue can identify the common floor beneath these frameworks, the minimum standards on transparency, data integrity, and human oversight that all jurisdictions can anchor to, without requiring regulatory uniformity. Second, by creating a legitimate escalation pathway. When an AI system deployed by a multinational corporation causes harm in a lower-income country with limited regulatory capacity, there is currently no international mechanism for redress. The Dialogue can lay the groundwork for a multilateral accountability structure that smaller states can access without needing bilateral leverage. Third, by institutionalizing technical assistance as a governance obligation. Countries that lack the regulatory and technical infrastructure to audit AI systems deployed within their borders are effectively ungoverned in this domain, regardless of their domestic AI development activity. The Dialogue should establish that building this oversight capacity is a core deliverable of international AI cooperation, funded proportionally by states and companies that benefit most from global AI deployment. The Dialogue's value could be immense in the durable structures it builds for cooperation that outlasts any single summit.

Several existing mechanisms have laid important groundwork. The Dialogue has an opportunity to connect them into a coherent architecture rather than duplicate their work. The OECD AI Principles and the accompanying OECD.AI Policy Observatory represent the most comprehensive existing cross-jurisdictional framework, with adoption across over 40 countries. The Dialogue should treat these as a baseline and focus on extending their reach to non-OECD member states, particularly in Africa, Southeast Asia, and Latin America, where AI deployment is accelerating fastest. The UNESCO Recommendation on the Ethics of AI, adopted by 193 member states, provides a rights-based foundation. Its implementation, however, remains uneven. The Dialogue can add value by developing a monitoring mechanism to track how member states translate the Recommendation into domestic policy, thereby making implementation visible and comparable. The Global Partnership on AI (GPAI) connects research with policy. The Council of Europe's Framework Convention on AI and Human Rights represents the first binding international instrument in this space. Each of these operates in a distinct lane. The Dialogue's added value is in its universality. The OECD's reach is limited to wealthier economies. UNESCO implementation is voluntary. GPAI membership is selective. The Dialogue is the only forum where every country has standing, which means it is uniquely positioned to surface governance gaps that are invisible to forums dominated by high-capacity states. Specifically, the Dialogue should map existing initiatives, identify gaps in their collective coverage, and commission targeted workstreams to fill them, rather than building parallel structures that compete for the same institutional attention and funding.

For effective stakeholder contribution, there should be a structural design taht encourage pre-work and active engagement. Member states should be required to submit data governance self-assessments prior to each session, documenting the state of AI oversight infrastructure within their jurisdictions. This creates accountability and generates the comparative data needed to identify where international support is most needed. Civil society and affected communities should have designated speaking slots in substantive sessions, not only in side events. Organizations working on the front lines of AI-driven harm, in credit scoring, predictive policing, content moderation, and automated public-benefit decisions, hold evidentiary knowledge that policymakers do not. That knowledge belongs in the main room. Technical practitioners from industry and academia should be engaged as working group contributors, tasked with translating governance principles into implementable technical standards. Policy fluency and technical depth rarely coexist; the structure should engineer their collaboration. For format, the Dialogue should adopt a tiered structure: plenary sessions for norm-setting and political commitment; technical working groups for standards development; and regional implementation hubs that localize outputs for different legal and infrastructural contexts. Between formal sessions, an open digital consultation process should allow broader participation from individuals, researchers, and community organizations who cannot travel to Geneva or New York. Submissions should be synthesized and formally tabled, not archived. The Dialogue's legitimacy will ultimately be determined by whether its outputs reflect the priorities of those most affected by AI, not only those with the most resources to participate.

The gap between who governs AI and who is governed by it is one of the most structural problems in this space. Several communities remain systematically underrepresented. People in the Global South are among the most exposed to AI systems in public services, yet their governments are among the least represented in the rooms where governance norms are set. Inclusion requires more than open invitations: it requires funded participation pathways, advanced translation of working documents, and scheduling that accounts for time zone distribution across plenary and working group sessions. Frontline workers and labor communities are living the economic disruption that AI is driving in real time, yet labor perspectives are rarely present in AI governance forums beyond high-level union representation. Workers in logistics, content moderation, data annotation, and customer service have direct, granular knowledge of how AI systems behave in practice that governance frameworks currently lack. Indigenous communities hold governance traditions centered on collective accountability, intergenerational responsibility, and relationship to land and data, frameworks that offer genuine alternatives to the individual rights-based models that dominate current AI governance discourse. Their inclusion is not symbolic; it is intellectually necessary. Young technologists from underrepresented regions, particularly those building AI applications in contexts that Western frameworks did not anticipate, bring both technical knowledge and contextual legitimacy that established institutions cannot replicate. Structural inclusion mechanisms should include: reserved seats for civil society organizations from low and middle-income countries on all working groups; a youth delegate program modeled on existing UN climate mechanisms; and a formal process for translating community-level impact evidence into policy submissions that the Dialogue is required to consider and respond to.

The standard conference format, keynotes, panels, and side events, produces declarations instead of actionable decisions. The Dialogue should experiment with formats designed to generate durable outputs from participants. Deliberative working groups with binding deliverables. Rather than panels where participants share perspectives, structured working groups should be tasked with producing specific outputs: a draft standard, a gap analysis, a set of implementation criteria. Assigning real work to diverse participants yields more equitable outcomes than open discussion, in which eloquence and institutional backing determine whose views prevail. Red team sessions. Invite practitioners, particularly from civil society and technical communities in the Global South, to actively stress-test proposed governance frameworks against real deployment contexts. A data governance standard developed in Geneva should be tested against how it would function in a Nigerian state benefits system or an Indonesian labor platform before it is adopted. Structured adversarial review is a governance tool, not just an engineering one. Asynchronous digital participation tracks. Not everyone who has relevant knowledge can travel to Geneva. A structured asynchronous track, with real deadlines, synthesis processes, and formal tabling of outputs, would extend participation without reducing rigor. This is distinct from open comment periods, which are rarely read and never binding. Implementation showcases. Each session should include dedicated time for practitioners to present concrete examples of AI governance in action: what worked, what failed, and why. Case-based learning from real deployments is more generative than abstract principle-setting. Rotating regional hosts for intersessional work. Hosting working group meetings in Nairobi, Sao Paulo, or Jakarta, not only Geneva and New York, signals that governance is being built for global application, and surfaces contextual knowledge that headquarters-city forums consistently miss.

Several existing models offer genuine lessons for the Dialogue to build on. Singapore's Model AI Governance Framework is one of the most implementable practitioner-facing governance documents currently in existence. It translates high-level principles into sector-specific guidance with concrete decision trees. Its voluntary adoption model has limits, but its translation methodology is replicable at the international level. The EU AI Act's risk-tiered approach establishes a precedent for proportionate governance: higher-stakes applications face stricter requirements. The data governance provisions, including requirements for training data documentation in high-risk systems, are the closest existing international model to treating data quality as a governance obligation rather than an engineering choice. Rwanda's national AI policy represents a Global South model worth amplifying. Developed with explicit attention to local context, linguistic diversity, and leapfrog potential, it demonstrates that high-capacity AI governance is not exclusive to wealthy economies. The NIST AI Risk Management Framework in the United States offers a voluntary but technically rigorous approach to organizational AI governance. Its limitation is enforceability; its strength is practitioner adoption. The Dialogue should examine how frameworks with high technical credibility can be connected to accountability mechanisms with actual consequences. From my own sector, data quality programs structured as compliance frameworks, with defined rules, documented lineage, and regular audit cycles, demonstrate that technical governance infrastructure can be built at scale inside large organizations. The same design logic applies internationally: governance without measurement infrastructure is aspiration; governance with it is accountability. The Dialogue should commission a living registry of national and organizational AI governance implementations, documenting what works, at what scale, under what conditions, so member states can learn from each other rather than building in isolation.