Siegel Family Endowment

Success requires establishing new international infrastructure for ongoing AI governance—not just producing declarations, but creating mechanisms to stay current as AI rapidly evolves. Critical outcomes: (1) Adaptive coordination mechanisms: We need agile structures that can respond to technical developments in weeks, not years. This means creating smaller working groups of peer states—countries with comparable AI readiness and similar policy, infrastructure, and workforce challenges—that can learn and act faster, then share insights with the broader collective. (2) Knowledge-sharing infrastructure: Establish systems for states to continuously update each other on AI developments, regulatory experiments, and emerging challenges. No single country can track everything; we need structured ways to pool intelligence. (3) Concrete governance pilots: Move beyond principles to fund and document real governance experiments at city, regional, and national scales. Create accountability mechanisms to track what works and what doesn't. (4) Long-term capacity building: Success means establishing permanent structures that keep pace with AI advancement and help states build sovereign technical capacity. For most policymakers, the need for a "usability layer" to make sense of all of the levers that can be used for effective governance is paramount. The measure of success should not be consensus documents, but whether we leave with working mechanisms that help governments govern AI knowledgeably and adapt as technology changes.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• AI capacity-building
• Open-source software, open data and open AI models

Safe, secure and trustworthy AI: The focus on "facilitating access to AI applications" risks recreating digital colonialism. Simply providing access to dominant American and European models-like Facebook's "Free Basics" did with internet access-hooks populations into systems they don't control. We need technological sovereignty: supporting countries to build rights-respecting, privacy-protecting alternatives rather than deepening dependency on concentrated corporate power. AI capacity-building: Building genuine technical capacity means more than training people to use existing tools. Supporting countries to develop their own governance expertise, conduct independent evaluations (safety, bias, algorithmic auditing), and make informed technology choices rather than relying entirely on self-certification by developers. Interoperability of governance approaches: Current discussion focuses on interoperability of policies and regulations. We must expand this to include technical interoperability-oversight of code and models themselves. This means funding independent evaluation infrastructure, creating shared technical standards, and ensuring diverse governance approaches can actually be implemented technically. Open-source software, open data, and open AI models: The "open vs. closed" debate oversimplifies complex trade-offs. We need nuanced discussion of when openness advances equity (enabling local adaptation, reducing dependency) versus when it enables extraction (training on public data, concentrating benefits privately). Success means developing frameworks for responsible openness that balance access, safety, and sovereignty.

Process over performance: The agenda risks over-indexing on admiring the problem rather than responding to it. We need concrete solution proposals-even imperfect ones-for reaction and workshopping. Consider adding lightning talks presenting current or aspirational governance models so discussions start from tangible ideas rather than abstract principles. Governance infrastructure, not just governance principles: Missing from themes is how we organize ourselves to govern AI effectively. Key questions include: How do states stay current with rapid AI developments? How do smaller working groups of similarly situated countries learn faster and share back to the collective? What mechanisms enable ongoing adaptation rather than one-time consensus? Beyond access to alternatives: "Facilitating access to AI applications" appears in the document without interrogating what that means. Access to whose applications? Built on whose terms? We need explicit focus on technological sovereignty-supporting countries to build or meaningfully shape AI systems aligned with their values and needs, besides what dominant actors provide. Accountability mechanisms: The dialogue should establish concrete ways to track commitments, document governance experiments, and hold actors (governments and companies) accountable over time. Without this, we risk a cycle of declarations without implementation. Recommendation: Structure sessions to produce actionable next steps and accountability frameworks beyond shared understanding. The goal should be leaving with systems for ongoing response and adaptation, not just a summit report.

As a U.S.-based philanthropic foundation working globally on public interest technology, we see governance gaps creating three critical challenges: (1) Fragmented philanthropic coordination: Foundations funding AI governance operate without shared technical literacy or coordination mechanisms. This creates duplicative investments, missed leverage points, and inconsistent support for critical infrastructure (independent evaluation labs, data governance frameworks, Global South capacity). The result is a clear market failure.Commercial AI development is massively resourced while public interest alternatives and accountability mechanisms remain chronically underfunded. (2) Information asymmetry fueled by public interests: Philanthropic and policy actors are making decisions about governance based on information overwhelmingly controlled by frontier labs themselves. Without independent technical evaluation capacity, funders and governments cannot assess safety claims, identify where interventions matter most, or hold companies accountable. This power imbalance undermines effective progress. (3) Global South exclusion from governance design: Most AI governance conversations center U.S. and European perspectives, yet impacts are global. Countries lack resources to build sovereign technical capacity, conduct independent audits, or participate meaningfully in standard-setting. This risks governance frameworks that serve dominant markets while failing communities most vulnerable to AI harms. Opportunities we're pursuing: Through initiatives like Tech Together (my organization, Siegel Family Endowment's, annual convening of 200+ philanthropists, technologists, and advocates around the UN General Assembly in New York City) and "Let's Get Technical" (working groups building philanthropic technical literacy), we're creating coordination infrastructure the field lacks. We're demonstrating that foundations can move faster than governments to pilot governance models, fund critical gaps, and bridge sectors. The Global Dialogue presents an opportunity to formalize what we've learned: (1) Governance requires ongoing adaptive infrastructure. (2) Peer learning among similarly positioned actors accelerates progress. (3) Philanthropic capital can catalyze what markets won't fund and governments can't move quickly enough to build.

Create peer learning networks: Facilitate smaller working groups of countries with comparable AI readiness, infrastructure, and policy challenges. These cohorts can experiment, fail fast, share lessons, and develop governance models suited to their contexts, then contribute insights to the broader collective. Universal approaches often fail; peer-to-peer learning succeeds. Bridge fragmented efforts: Currently, philanthropic funders, national regulators, multilateral bodies, and civil society organizations operate in silos. The Dialogue can create structured channels for these actors to coordinate, sharing technical intelligence, aligning on standards, and avoiding duplicative efforts. This connective tissue doesn't exist today. Pilot governance experiments: Move beyond principles to fund concrete pilots at city, regional, and national scales. Document what works, what fails, and why. Create accountability mechanisms that track implementation. Make pilot learnings immediately available to other jurisdictions. Build rapid response capacity: Establish mechanisms for real-time information sharing when significant AI developments occur (new capabilities, safety incidents, regulatory innovations). Countries need ways to learn from each other's experiences quickly. Counter industry capture: Provide a space where governments can develop shared positions independent of industry influence. Currently, frontier labs shape narratives and set agendas at major conferences across the world, including at official UN convenings. The Dialogue can create forums where states build collective knowledge and negotiate from informed positions. Success looks like: Not a final report, but standing structures—working groups that meet quarterly, shared technical resources, rapid response protocols, and accountability dashboards tracking governance experiments globally. The Dialogue should serve as infrastructure, not an event.

BUILD UPON * Partnership on AI (PAI): Multi-stakeholder model with practical guidance, but industry-funded and therefore limited in challenging corporate practices. The Dialogue could provide independent space for harder conversations about regulation and accountability. * Office of the High Commissioner on Human Rights: Particularly the Human Rights Data Exchange and digital rights frameworks. The Dialogue should amplify this work and connect it to technical governance—showing how human rights principles translate into technical standards and evaluation criteria. * Regional efforts: EU AI Act, African Union's Continental Artificial Intelligence Strategy, ASEAN Guide on AI Governance and Ethics. The Dialogue can facilitate peer exchange between regions experimenting with different approaches, helping them learn from each other's implementation challenges. * Philanthropic coordination: bring funders, technologists, and advocates together to align resources. The Dialogue could formalize philanthropic roles in funding governance infrastructure that markets won't support and governments can't build alone. ADDED VALUE The Dialogue could bring inclusive speed and adaptation. Existing mechanisms are too slow. The Dialogue can create lightweight, agile structures that respond to developments in weeks, not treaty negotiation timescales.

STAKEHOLDERS Governments: Lead working groups on specific governance challenges (compute governance, data sovereignty, safety standards). Share regulatory experiments and implementation lessons that illustrate positions rather than simply stating those positions. Philanthropy: Fund governance infrastructure that markets won't support—independent evaluation labs, Global South capacity building, pilot programs. Coordinate resources across foundations to avoid duplication and maximize impact. Technical researchers: Translate complex AI developments into policy-relevant insights. Help governments understand technical feasibility constraints on proposed regulations. Conduct independent safety assessments. Industry: Participate in standards development, but not agenda-setting. Share technical information transparently. FORMAT RECOMMENDATIONS Avoid panel-heavy agendas. Use working sessions where participants collaborate on concrete problems: drafting model legislation, designing evaluation frameworks, mapping governance gaps. Use "lightning solutions" format: 5-minute presentations of existing governance models (even imperfect ones) to ground discussions in tangible approaches rather than abstract principles. End each session with accountability: What will each actor do next? By when? Who tracks progress?

UNDERREPRESENTED VOICES Global South technologists and policymakers: Not just as observers, but as leaders sharing governance innovations. Many African, Latin American, and Asian countries are experimenting with AI regulation in resource-constrained contexts—lessons Western countries need to learn. Workers and labor organizations: AI's most immediate impacts are on employment, yet labor voices are largely absent from governance conversations dominated by tech executives and policy elites. Indigenous communities: Particularly on data sovereignty, consent frameworks, and cultural preservation. Indigenous approaches to collective governance offer alternatives to individualized Western models. Artists and creative workers: Facing immediate AI displacement and appropriation, yet excluded from discussions about training data, copyright, and economic models. Small and medium states: Countries without major AI industries have different governance needs than tech superpowers, but lack resources to participate meaningfully in international processes. HOW TO INCLUDE THEM Redistribute decision-making power: Representation without power is performance. Ensure underrepresented voices have ways to shape agendas and outcomes, not just input or react. Shift timing and location: Hold regional sessions (or find ways to collect ideas) in different time zones. Not everyone can attend Geneva in July. Create dedicated working groups/tracks: Workers' forum, Indigenous data sovereignty session, artist roundtable, etc. One-time participation doesn't build influence; establish peer networks for sustained engagement.

"Show, don't tell" demonstrations: Instead of describing governance approaches abstractly, demonstrate them. Live red-teaming and algorithmic auditing exercise. Participatory AI impact assessment with affected community members. Rapid prototyping sprints: Give working groups 90 minutes to solve specific problems: "Design an evaluation framework for small countries without technical capacity." "Create a rapid response protocol for AI safety incidents." Present solutions; identify what's reusable. Governance prototyping workshops: Small groups draft model policies, technical standards, or accountability frameworks in real-time. Experts provide feedback; participants leave with working documents they can adapt for their contexts. Accountability mapping: Visualize who's committed to what, who's delivering, who's not. Create public dashboards tracking implementation in real-time (AI tools make this possible). Live case studies: Present real governance dilemmas countries are facing. Break into groups representing different stakeholders (government, civil society, industry, affected communities). Negotiate solutions together.

Policies and practices: EU AI Act's tiered risk approach: Imperfect but actionable. Demonstrates how to regulate AI based on use cases rather than technology itself. Provides templates others can adapt. UK AI Safety Institute: Independent government-funded entity conducting technical evaluations. Model for how countries can build capacity to assess AI systems without relying on industry self-certification. California's AI safety bill (SB 1047) debate: Though not passed, it surfaced critical questions about when regulation should focus on model developers vs. deployers. Valuable learning for other jurisdictions. Algorithmic impact assessments (Canada, EU): Require organizations to evaluate AI systems' potential harms before deployment. Creates accountability trail and documentation for oversight. Practices we're developing: Tech Together model: Siegel Family Endowment's Annual convening bringing 200+ philanthropists, technologists, and advocates to coordinate funding, share intelligence, and align on priorities. Demonstrates how to build field infrastructure outside government processes. Philanthropic working groups: Siegel Family Endowment's "Let's Get Technical" series builds funders' technical literacy so they can evaluate AI governance proposals critically and fund strategically. Addresses information asymmetry that undermines effective philanthropy. Peer learning cohorts: Small groups of similarly positioned countries sharing governance experiments in real-time. Faster learning than waiting for formal UN reports. What works across these examples: They're specific and actionable, and move beyond abstract principles. They're adaptable to different contexts. They include accountability mechanisms (who does what by when). They document and share learnings rather than treating each jurisdiction as starting from zero.