RENCI at UNC Chapel Hill

Here's the version refined for an international, multilateral audience—the tone is diplomatic, inclusive, and geared toward global collaboration (e.g., UN, OECD, or Global Partnership on AI contexts). A successful first Global Dialogue on AI Governance would mark a turning point in collective efforts to shape artificial intelligence for the global public good. Its strongest achievement would be a shared commitment among governments and institutions to a core set of guiding principles—transparency, accountability, safety, equity, and respect for human rights. Such principles would establish a common framework for cooperation, ensuring that progress in AI development advances human well‑being, sustainable growth, and international stability. Equally important, the dialogue should produce a roadmap for interoperable governance—a practical blueprint for aligning policies, standards, and oversight mechanisms across jurisdictions. This roadmap would identify key areas of convergence, such as risk classification systems, model evaluation and audit processes, and ethical norms for cross‑border data use. By clarifying how diverse governance approaches can connect, it would strengthen regulatory coherence, facilitate trade and research collaboration, and reduce global disparities in AI oversight. Participants should also endorse commitments to capacity building that enable all regions, including developing economies, to engage fully in AI governance. These efforts might include global knowledge‑sharing networks, technical exchange platforms, and targeted support for institutions building responsible AI infrastructure. Ensuring that every nation can contribute to and benefit from AI governance is essential to long‑term equity and legitimacy. Finally, true success will depend on institutional continuity and accountability. Establishing ongoing working groups, transparent progress reviews, and mechanisms for updating shared norms will ensure that dialogue outcomes translate into durable action. If the Global Dialogue fosters trust, shared direction, and sustained cooperation across regions, it will have succeeded in laying the groundwork for a coordinated, human‑centered international architecture for AI governance.

• Interoperability of governance approaches
• Protection and promotion of human rights
• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI

From my perspective, these thematic areas together capture the conditions under which AI can truly serve the public interest rather than undermine it. First, "safe, secure and trustworthy AI" is foundational because it addresses immediate, tangible risks to people and institutions. Without basic assurances that systems are robust, resilient to attack, and behave as intended, all other aspirations for AI governance become aspirational rather than actionable. Safety and security are also prerequisites for public trust, which is essential for the legitimate deployment of AI in critical domains such as health, justice, and public administration. Second, "interoperability of governance approaches" is urgent because AI systems, data flows, and commercial actors are inherently transnational. Fragmented or conflicting regulatory regimes create gaps that can be exploited, while also placing disproportionate burdens on smaller states and institutions with fewer resources to navigate multiple frameworks. Interoperability does not require uniformity, but it does require shared reference points, mutual recognition where appropriate, and mechanisms that allow different regimes to "speak to each other" in practice. Third, the "protection and promotion of human rights" must be a central pillar rather than an afterthought. AI systems can scale discrimination, surveillance, and exclusion as easily as they can scale beneficial services. Anchoring the Dialogue in human rights ensures that governance debates are not confined to technical parameters or economic competitiveness, but are grounded in dignity, autonomy, and equity for individuals and communities. Finally, "transparency, accountability, and human oversight" are critical for converting principles into enforceable practice. Transparency enables scrutiny and informed consent; accountability ensures that harms have remedies and that responsibilities are clearly allocated; and human oversight maintains a meaningful role for human judgment, especially in high-stakes decisions. Taken together, these themes define a coherent agenda: building AI ecosystems that are safe, rights-respecting, aligned across borders, and governed through structures that answer to people, not just to markets or technologies.

First, data governance and infrastructure equity underpins every other objective. Questions of data quality, representativeness, access, and cross-border flows determine whose realities are visible to AI systems and who bears the risks of error. Shared evaluation datasets, secure data environments, and strategies to address structural data gaps-especially for marginalized populations-are central to both safety and human rights, yet often treated as a technical afterthought rather than a governance priority. Second, concentration of compute, models, and market power is an emerging structural risk. A small number of firms and countries increasingly control the computational resources, talent, and foundational models that others must rely on. This raises concerns about dependency, lock-in, and the potential for regulatory capture that are not fully addressed by transparency or interoperability alone. Governance needs to grapple explicitly with competition, open ecosystems, and safeguards against excessive centralization of power. Third, the environmental and resource impacts of AI require more systematic attention. Training and deploying advanced models is energy- and resource-intensive, with implications for climate commitments, water use, and e-waste. Embedding sustainability metrics and lifecycle considerations into AI governance is essential to avoid exporting environmental burdens to already vulnerable communities. Finally, global equity and capacity asymmetries cut across all themes. Many states and communities lack the technical, regulatory, or institutional capacity to participate meaningfully in standard-setting, risk assessment, or enforcement. Without deliberate investment in shared tools, capacity building, and inclusive multilateral processes, AI governance risks becoming "governance by import," where rules and technologies are effectively set elsewhere and merely received rather than co-created.

As a US based academic researcher the most immediate challenge is regulatory and policy fragmentation. In the absence of a clear, stable federal framework for safe, secure and trustworthy AI, researchers must navigate a patchwork of institutional policies, state laws, funder expectations, and publisher standards that often conflict or change quickly. This increases compliance burden, creates uncertainty about acceptable practices (especially with generative models and sensitive data), and can chill high‑risk but socially valuable research. A second challenge is access and dependence. Interoperability gaps and concentration of compute and data in a few firms mean that many academics rely on corporate platforms and partnerships to do cutting‑edge work. That dependence can constrain research agendas, complicate transparency and reproducibility, and raise questions about independence when studying the impacts of those same systems. Human rights, transparency, and oversight gaps also show up as institutional strain. Universities are under pressure to adopt AI tools in teaching and administration without robust governance processes that include faculty. This can erode academic freedom, alter evaluation and hiring practices, and shift labor conditions in ways that academics experience directly but do not fully control. At the same time, these gaps create significant opportunities. Researchers are uniquely positioned to design evaluation methods, auditing tools, and socio‑technical studies that operationalize safety, rights protection, and accountability. Interdisciplinary teams can help define practical standards for transparency and human oversight, and empirically document harms in real‑world deployments. Fragmented governance also opens space for academics to shape norms: by drafting institutional policies, informing standard‑setting bodies, partnering with civil society, and providing independent evidence to courts and regulators. For US researchers, the current moment offers a rare chance to influence how AI is governed, not only to comply with rules but to help write them.

We need AI Dialogue to serve as practical engine for international cooperation rather than a one‑off discussion forum. It can translate broad principles into shared reference standards. By convening diverse states and stakeholders, the Dialogue can help crystallize common baselines for concepts like "high‑risk AI," meaningful transparency, and human oversight. Even non‑binding agreement on definitions, documentation practices, and risk‑management expectations can make national frameworks more interoperable and reduce regulatory fragmentation. The Dialogue can build institutional infrastructure for cooperation. This includes establishing standing expert groups, technical working parties, and joint observatories to monitor emerging risks, share incident reports, and coordinate responses. It can promote mechanisms for mutual learning—such as model evaluation consortia, cross‑jurisdictional audit pilots, and shared benchmarks—that support safe, secure, and trustworthy AI across borders. Finally, the Dialogue can connect AI governance to existing international regimes on human rights, trade, development, and security—so that AI is not governed in a silo. It can encourage alignment with existing obligations, promote policy coherence across forums, and create regular review cycles where states report on progress and challenges. In doing so, the AI Dialogue can function as a coordinating hub that turns scattered initiatives into a more coherent, human‑centered international architecture for AI governance.

At NSF, the AI Dialogue should build on the National AI Research Institutes, including those focused on trustworthy AI, AI and society, and cyberinfrastructure for data‑intensive science. These institutes are developing methods for transparency, robustness, human‑in‑the‑loop design, and cross‑institutional data stewardship. Bringing their frameworks, benchmarks, and socio‑technical findings into the Dialogue would ground international conversations in tested practices rather than abstract principles. At the Department of Energy, the Dialogue can draw on large‑scale AI initiatives such as Genesis, which leverage DOE's advanced computing and data infrastructure to develop and evaluate powerful models in high‑consequence scientific and energy contexts. Programs like Genesis will confront questions of safety, reliability, and security at scale, including controls on access to compute, rigorous evaluation of model behavior, and governance of dual‑use capabilities. Their approaches to verification, model governance, and secure compute environments could inform global norms on "safe and secure" AI, particularly for frontier‑scale systems. From NIH, the Dialogue should connect with programs at the intersection of AI, health, and sensitive data—for example, initiatives in AI for biomedical and behavioral research, large data‑commons and cloud platforms, and precision medicine or genomics programs that rely on strict data governance. These efforts have advanced practical models for consent, de‑identification, controlled access, and oversight of algorithmic tools in clinical and public‑health contexts. In particular, the large NIH data ecosystems such as NHLBI BioData Catalyst are pushing the boundaries on existing policies, security and governance.

To address unequal power dynamics among stakeholders in AI ethics dialogues, while countering AI's novel speed, prioritize equitable participation, agile decision-making, and accountability. While it can be hard to do, results are improved is underrepresented groups are compensated for time, travel and other requirements. Given the speed of AI evolution, agile co-governance is the best approach, although this can be difficult. This means: shared agenda-setting via async tools, weekly sprints for principle updates, and community escalation for urgent risks. Embed speed clauses mandating real-time reviews before large releases or policies. Use public dashboards tracking consultations, feedback, and outcomes in real time. Conduct rolling external audits of AI systems and processes, with automated alerts for rapid changes. Engagement across stakeholders will likely require upskilling. For a diverse group such as this just-in-time training via modular online modules, neutral advisors, and AI literacy fellowships is likely to reach the community. Pair communities with dedicated rapid response experts for ongoing input. Measurement of the success of these approaches is critical and also must be considered.

AI underrepresentation mirrors underrepresentation in other technologies. Those communities often excluded from design tables, facing harms in hiring, policing, and healthcare via biased systems are: Global South nations and smaller organizations, marginalized communities (e.g., women, low-income, immigrants), and non-western ethical frameworks. Strategies include holding slots for underrepresented experts, building capacity, co-creation initiatives, and requirements for diverse data (e.g., multilingual, non-western data, bias auditing).

Start with lightning rounds followed by AI-facilitated breakouts: Each stakeholder cluster gets two minutes to share, then AI tools sort participants into complementary breakout rooms—pairing ethicists with engineers on live-voted priorities. Real-time transcription and sentiment analysis highlight consensus gaps, sparking immediate synthesis. For hands-on momentum, host hackathon dialogues: 48-hour sprints where civil society teams up with tech experts to prototype solutions, such as bias-detection tools or community veto dashboards, with prizes for viable pilots. Sustain energy through living playbooks on collaborative platforms like Notion with AI co-pilots, enabling async updates on emerging risks and rapid-response threads for urgent issues, like new model drops. Add depth with peer shadowing, rotating underrepresented voices into tech or policy teams for 1-3 months to influence decisions from the inside. Voice-activated async input, multilingual voice submissions auto-translated into shared dashboards, and gamified feedback with points for contributions, tracking real action to prevent participation-washing.

Singapore's Model AI Governance Framework promotes human-centric AI through voluntary guidelines on transparency, explainability, and accountability, enabling rapid adoption while building trust. The EU AI Act classifies systems by risk tiers, mandating audits, human oversight, and bans on high-risk uses like real-time biometric ID, with compliance deadlines phased through 2027. Cleveland Clinic's governance model uses multidisciplinary committees (doctors, ethicists, patients) for monthly AI diagnostic reviews, catching biases early and retaining physician veto power. Google's AI Principles prohibit harmful applications (e.g., weapons) and require fairness assessments, integrated into development workflows. Credo AI centralizes model risk management, tracking lineage and regulatory alignment (e.g., NIST AI RMF) to prevent shadow AI. Holistic AI inventories systems lifecycle-wide, automating bias/drift monitoring for LLMs. Monitaur enforces policy-to-proof workflows, generating audit-ready evidence. Human-in-the-loop for high-risk decisions, continuous model registries, and public dashboards link input to outcomes, countering participation-washing while matching AI's pace.