University of São Paulo

A successful first Global Dialogue on AI Governance should move beyond high-level discussions and deliver concrete, actionable outcomes that can guide global coordination on AI. First, it should produce a shared baseline of principles that are not only normative, but also operationalizable, particularly around accountability, transparency, and fairness. These principles must be grounded in real-world technical constraints and existing methodologies, ensuring they are implementable across diverse contexts. Second, the Dialogue should initiate the development of a structured, lifecycle-oriented governance framework for AI systems. Current approaches often address isolated aspects (e.g., fairness or interpretability), but lack integration across the full AI lifecycle—from data collection to deployment and monitoring. Establishing a roadmap toward such a framework would be a critical step forward. Third, success would include fostering sustained multi-stakeholder collaboration, especially between technical researchers, policymakers, and actors from the Global South. Inclusive participation is essential to avoid governance models that are misaligned with local realities and to ensure equitable representation in shaping global standards. Finally, the Dialogue should define clear next steps, such as working groups, pilot initiatives, or evaluation benchmarks, to maintain momentum beyond the event. Without continuity and accountability mechanisms, even strong initial alignment risks losing impact. In this sense, success is not only measured by consensus, but by the creation of durable structures that translate dialogue into coordinated global action.

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

My priorities center on advancing trustworthy AI through technically grounded approaches that directly address societal risks in real-world deployments. I selected "Safe, secure and trustworthy AI" as a foundational priority, as ensuring reliability and robustness is a prerequisite for any meaningful governance effort. However, achieving trustworthiness requires more than high-level principles; it depends on actionable mechanisms. For this reason, I also prioritize "Transparency, accountability, and human oversight." My work focuses on explainability methods for NLP systems in sensitive domains, where understanding model behavior is essential for auditing decisions, mitigating harms, and enabling meaningful human oversight. "Protection and promotion of human rights" is equally central. AI systems increasingly mediate access to information and public discourse, particularly in areas such as content moderation. Without careful design and evaluation, these systems risk reinforcing discrimination, silencing marginalized voices, or amplifying harmful content. Finally, I emphasize the "Social, economic, ethical, cultural, linguistic and technical implications of AI." Many current approaches are developed in limited linguistic and cultural contexts, which creates significant gaps when applied globally. My work with Portuguese-language data highlights the importance of inclusive, context-aware solutions that reflect diverse realities, particularly in the Global South. Together, these priorities reflect a commitment to bridging technical innovation with governance needs, ensuring that AI systems are not only performant, but also accountable, inclusive, and aligned with fundamental rights.

While the listed themes cover many critical dimensions of AI governance, an important cross-cutting gap lies in the lack of a lifecycle-oriented perspective that connects these areas into a coherent, operational framework. Current discussions often address issues such as fairness, transparency, and safety in isolation. However, in practice, these challenges are deeply interconnected and emerge at different stages of the AI lifecycle, from data collection and model development to deployment and post-deployment monitoring. Without an integrated approach, governance efforts risk remaining fragmented and difficult to implement consistently. A related emerging issue is the gap between technical methods and governance requirements. While there has been progress in areas such as explainability, bias mitigation, and robustness, these methods are not yet systematically aligned with policy needs or regulatory expectations. This creates challenges for translating high-level principles into actionable and auditable practices. Additionally, there is a lack of standardized evaluation frameworks and benchmarks for assessing whether AI systems meet governance objectives in real-world settings. Defining measurable criteria for concepts like accountability, transparency, and fairness remains an open challenge, particularly across different cultural and linguistic contexts. Addressing these cross-cutting issues would strengthen the connection between principles and practice, enabling more effective and globally applicable AI governance.

Governance gaps in trustworthy AI, transparency, and human rights have direct and tangible impacts in Brazil and, more broadly, in the Global South, particularly in high-risk domains such as online content moderation and information ecosystems. One of the most significant challenges is the lack of context-aware and linguistically inclusive AI systems. Many models and evaluation frameworks are developed primarily for English, which leads to reduced performance and higher risks of misclassification in Portuguese. In sensitive applications, such as detecting hate speech or harmful content, this can result in both under-enforcement (allowing harmful content to persist) and over-enforcement (silencing legitimate or marginalized voices), raising serious human rights concerns. Another key challenge is the limited availability of standardized mechanisms for transparency and accountability. While technical methods for explainability and bias mitigation exist, they are not yet systematically integrated into deployed systems or aligned with regulatory expectations. This creates barriers for auditing AI systems and ensuring responsible deployment in practice. At the same time, these gaps present important opportunities. Brazil has a growing research community and increasing engagement in AI governance discussions, positioning it as a potential leader in shaping more inclusive and globally relevant standards. Efforts to develop datasets, benchmarks, and evaluation methodologies in Portuguese and other underrepresented languages can contribute to more equitable AI systems worldwide. Overall, addressing these challenges requires bridging technical innovation with governance frameworks that are sensitive to local realities while contributing to global coordination.

The AI Dialogue can advance international cooperation by providing a structured platform for multi-stakeholder engagement, aligning technical and policy perspectives, and fostering the sharing of best practices across regions. It can promote consensus on operational principles for transparency, accountability, and fairness, and support the development of benchmarks and tools that guide trustworthy AI deployment. Additionally, it can enable collaborative initiatives and capacity-building, particularly in underrepresented regions, ensuring that governance approaches are inclusive, culturally aware, and globally relevant. By connecting high-level discussions with actionable outcomes, the Dialogue can help establish a coordinated ecosystem for safe, responsible, and socially beneficial AI.

The AI Dialogue should build on initiatives such as UNESCO's Recommendation on the Ethics of AI, the OECD AI Principles, the EU AI Act, and the Global Partnership on AI. While these provide essential ethical and policy guidance, they often operate in silos or focus on specific regions, languages, or sectors. The Dialogue can add value by connecting these efforts, fostering cross-regional learning, and promoting multi-stakeholder collaboration. It can support standardized benchmarks for transparency, accountability, and fairness, enable pilot projects and joint research, and ensure inclusivity for underrepresented languages and contexts. In doing so, the Dialogue can translate high-level principles into actionable, globally relevant governance practices.

Different stakeholders can contribute to the AI Dialogue through their unique expertise. Governments provide regulatory guidance and policy priorities, academia contributes evidence-based research and technical innovations, the private sector shares deployment experience, and civil society highlights societal, human rights, and cultural considerations. A multi-tiered format could maximize impact: thematic working groups addressing safety, transparency, fairness, and social implications; periodic plenary sessions for cross-group coordination; interactive workshops or pilot projects to test governance principles; and open calls for contributions to ensure broad, inclusive participation. Documenting outcomes—frameworks, benchmarks, and best practices—would create lasting value. This approach fosters collaboration, evidence-informed decisions, and globally relevant AI governance.

Voices from the Global South, Indigenous communities, marginalized populations, underrepresented languages, and women in technology remain limited in global AI governance discussions. This underrepresentation leads to governance frameworks that often fail to capture diverse socio-cultural realities and perpetuate biases. The AI Dialogue could address this by ensuring targeted outreach, inclusive consultation processes, and support for participation from these groups. Multi-lingual engagement, advisory boards including underrepresented stakeholders, and partnerships with local research institutions and civil society organizations would amplify grassroots perspectives. By systematically including these voices, the Dialogue can foster equitable, culturally sensitive, and globally relevant AI governance that reflects diverse experiences, needs, and contexts.

To foster meaningful engagement during the AI Dialogue, a combination of structured and interactive formats is key. Thematic working groups can bring together technical, policy, and civil society experts to tackle specific issues, while multi-stakeholder plenary sessions allow cross-group sharing and consensus building. Interactive workshops or hackathons can translate governance principles into practical solutions, producing tangible outputs. Open calls and virtual consultations ensure participation from underrepresented regions, languages, and communities. Finally, documentation and feedback loops capture outcomes and enable iterative improvement. Combining focused deliberation, collaborative experimentation, and inclusive outreach creates a dynamic, results-oriented, and globally representative engagement framework.

Effective AI governance can be promoted through a combination of policies, technical practices, and collaborative platforms. Regulatory frameworks such as the EU AI Act, OECD AI Principles, and UNESCO's Recommendation on the Ethics of AI provide standards for safety, fairness, transparency, and accountability. Technical practices like explainability methods, fairness audits, and robustness testing enable stakeholders to operationalize these principles, particularly in sensitive domains like content moderation. Collaborative initiatives, including the Global Partnership on AI and open-source platforms for datasets and models, facilitate knowledge sharing, reproducibility, and inclusion of underrepresented languages and contexts. Organizational measures such as AI ethics boards, pre-deployment risk assessments, and continuous monitoring help translate principles into practice, ensuring responsible and socially beneficial AI deployment.