UN AI Governance Roadmap Urges Safety Testing Before Release
Co-leads of the first UN Global Dialogue on AI Governance call for independent safety testing, published protocols, and inclusive global participation ahead of next May's talks.
Governments should require that new artificial intelligence systems be independently tested against agreed safety thresholds before they are released to the public, according to the co-leads of the first United Nations Global Dialogue on AI Governance. The recommendation is part of a broader roadmap for coordinated international oversight of AI that emerged from the Geneva meeting held this past July, where 170 countries joined industry and civil society representatives to find common ground on governing the fast-moving technology.
The call for binding pre-release testing comes amid mounting warnings from within the AI industry itself. Several leading chief executives have publicly urged a slower pace of development, citing risks that range from loss of control over advanced systems to cyberattacks and bioterrorism. Days later, Bank of England Governor Andrew Bailey, in his role as chair of the Financial Stability Board, told G20 finance ministers and central bankers that the global financial system faces a new threat from AI. Because so many institutions now depend on the same small group of AI models and cloud providers, a single exploited weakness could spread rapidly across markets and borders.
The Geneva dialogue produced a set of mechanisms that participants described as a first step toward coordinated governance rather than a finished framework. Among the core proposals: developers should publish their safety protocols, and any incidents that occur after deployment should be clearly monitored and reported. The co-leads argue that these measures are «no regret» actions — steps that carry no downside and that the UN could take immediately, even given the constraints of working at the international level.
Three priorities stand out for the months ahead. The first is connecting science and policy. At present, researchers studying AI risks and diplomats negotiating rules operate on separate tracks with no structured way to communicate. The UN's Independent International Scientific Panel on AI presented its first evidence-based assessment in Geneva, and the co-leads say that evidence needs a formal channel into the Global Dialogue's negotiating process before the group reconvenes next May, rather than sitting alongside it. With most AI research and development happening behind closed industry doors, citizens and policymakers need the best available scientific advice to remain informed and act.
The second priority is building on existing frameworks rather than starting from scratch. The co-leads call on the UN to develop a common reference baseline for safety, grounded in international human rights law and assembled from what already exists. They point to the OECD AI Principles, the G7 Hiroshima code of conduct, the Frontier AI Safety Commitments from the 2024 Seoul Summit, and UNESCO's Recommendation on the Ethics of AI, alongside standards from the International Organization for Standardization and the National Institute of Standards and Technology that engineers actually build to. A shared understanding of current regulatory floors would let individual nations prioritize what to build in their own national and regional contexts. Costa Rica's National AI Strategy, developed with government, civil society, industry, and academia at the same table and anchored in international frameworks, is cited as a model worth replicating.
The third priority is inclusivity, both geographic and financial. The co-leads argue that governments, researchers, and civil society groups with less financial capacity must be in the rooms where decisions are made, which requires funding real participation through a dedicated trust fund similar to the one that supports the Intergovernmental Panel on Climate Change. Inclusivity also means proving that different regulatory systems can work together on equal footing. The co-leads propose a handful of cross-border regulatory sandboxes, piloted by a small group of countries and shared openly, modeled on the Global Financial Innovation Network that lets financial regulators test new products across borders.
By the time the Global Dialogue reconvenes in May, the co-leads say three things should be in place: an evidence assessment from the scientific panel, a Dialogue agenda that takes up the safety baseline, and a first cross-border sandbox pilot. They warn that the urgency is not abstract, noting that the Financial Stability Board's letter and other recent events make clear these risks are already at the world's front door. The summer talks demonstrated that a wide range of countries and organizations can agree under pressure, the co-leads write, and the next step is delivering results the world can actually see. If a handful of powerful actors decide how AI is governed, they caution, the rest of the world loses out.
