Cluster 3: safe, secure & trustworthy AI

Thematic discussion · Day 2 · 7 July 2026 · Geneva
Session
Thematic Cluster 3 on safe, secure and trustworthy AI: interoperability and compatibility of governance approaches
Co-chairs
Paula Bogantes Zamora (Costa Rica, Minister of Science, Innovation, Technology & Telecommunications) and Rebecca Finlay (CEO, Partnership on AI)
Scene-setting
USG Amandeep Gill (UN Office for Digital and Emerging Technologies)
Featuring
Panel moderated by Virginia Dignum (Umeå University), with Japan, Infosys, the UAE and the European Data Protection Supervisor

What happened

Cluster 3 took on the least glamorous but most practical question in AI governance: how do different national rulebooks work together across borders? The framing analogy set the tone. The world runs on different plug shapes and voltages, yet a laptop still charges safely anywhere, not because every country agreed on one system but because adapters and common standards let different systems connect. The goal for AI governance is the same, the co-chairs argued: not harmonisation into a single model, but interoperability, so that safety evidence and accountability can travel.

Paula Bogantes Zamora (Costa Rica): “The world does not need more AI principles. It needs a common way to prove they’re being implemented.”

Concentration is the backdrop

Costa Rica grounded the case in hard numbers on how concentrated the field is. In 2025, institutions in the United States produced 59 notable AI models and China 35, while the rest of the world produced 13. The US also held roughly 75% of the computing power among the 500 largest known AI clusters, with China holding another 15%. The consequence Bogantes drew is the one closest to SDGCounting’s concern: concentration of infrastructure becomes concentration of evidence. Whoever has the compute also shapes which risks get measured, which benchmarks are accepted, which languages are evaluated, and who is recognised as capable of certifying safety. The preliminary scientific assessment found 118 countries, mainly from the global South, not meaningfully engaged in the main international AI-governance discussions, and fewer than a third of developing countries have a national AI strategy. Costa Rica’s proposal: aim for “minimal viable interoperability” before the 2027 dialogue, starting with shared terminology, comparable risk classifications, consistent documentation, interoperable incident reporting, and multilingual evaluation.

Fragmentation, and the bridges across it

USG Amandeep Gill set the technical scene by naming the opposite of interoperability: fragmentation. It drives regulatory arbitrage (developers moving to jurisdictions with weaker oversight), accountability gaps across global value chains, heavier compliance burdens (worst for SMEs and low-income-country developers), and deeper dependency for lower-capacity states. His answer was a set of “pathways,” not prescriptions: for example, connecting national regulatory sandboxes so AI can be tested and its evidence shared across borders. “The objective is not to harmonise AI governance into a single model,” he said, “but to build practical bridges across diverse approaches.”

Co-chair Rebecca Finlay of the Partnership on AI turned that into three asks for the year ahead: strengthen the independent scientific evidence base, open up the science behind it, and turn evidence into progress the public can see and hold institutions to. She announced a Global AI Progress Hub for sharing progress against a public-interest framework, and urged a change of questions: instead of “how fast can we deploy this,” ask “who does it serve, who is accountable when it fails, and how do we know?”

The panel: what to build on

Why it matters for the SDGs

This cluster is where AI safety meets the AI divide (SDG 10). The compute-and-evidence concentration means the countries least able to shape the rules are the ones that most need them, and a governance architecture that assumes universal access to advanced compute and regulatory expertise would reproduce inequality rather than close it. The through-line is measurement, SDGCounting’s home turf: interoperability “succeeds when evidence can travel, accountability can be traced, and safety claims can be verified.” That connects to SDG 16 (accountable institutions) and SDG 17 (capacity and cooperation), including the multilingual-evaluation gap across the world’s 7,000+ languages.

Watch & read

Quotations are lightly edited from an automated (Otter.ai) transcript of the UN Web TV recording and should be read as close paraphrase; names and titles were reconciled to public records and reflect roles at the time.