Singapore Seeks UN Framework to Put Global Brakes on AI, Without Slowing the Race

Singapore’s foreign minister offered UN delegates a grimly simple thought experiment: “Do you want the nuclear button to be in the hands of a human being or an AI?”

The answer, Vivian Balakrishnan suggested, should be obvious. What is considerably harder is working out how to keep humans in charge while governments and companies pour money into systems designed to do more without them.

Singapore’s answer is not to stop the race. On Saturday (September 26), Balakrishnan told the UN General Assembly that halting AI development was already unrealistic given superpower competition and the commercial incentives surrounding frontier models.

Instead, he put a more ambitious idea on the table. Explore a UN Framework Convention on AI Safeguards, potentially backed by an international institution that could set common technical standards and provide verification for AI, much as the International Telecommunication Union and the International Atomic Energy Agency do in their fields.

That proposal makes more sense for Singapore than for a country sitting on the sidelines of the AI boom. Singapore is doing almost the opposite.

ChatGPT maker OpenAI committed more than S$300 Mn this year to its ‘OpenAI for Singapore’ partnership, including its first Applied AI Lab outside the US and plans for more than 200 technical roles over the next few years. The city-state has also established as many as 70 corporate AI centres of excellence. The government wants 10,000 enterprises to deepen their use of AI over three years. It is now targeting a pool of 100,000 ‘AI bilingual’ workers who combine professional expertise with the ability to use the technology in their jobs.

In other words, Singapore is hardly waving a stop sign at AI. It is pressing the accelerator while asking who is responsible for the brakes.

Safety as an Adoption Condition

That distinction is important because it separates the proposal from a broader argument that advanced AI development should be paused. Singapore’s domestic policy has consistently treated governance as something that should make adoption safer, not prevent it. Its updated Model AI Governance Framework for Agentic AI also tells companies to define what an agent is allowed to do, restrict its access to tools and data, and establish points at which humans must approve actions and test systems throughout their lifecycle. Simply put, humans remain accountable even when software is doing the work.

However, those safeguards largely differ across companies and jurisdictions, while the systems themselves do not.

An AI agent built in one country can interact with infrastructure, software, financial networks and users elsewhere. But a serious failure in an autonomous system is unlikely to respect national boundaries any more than a cyberattack does. Companies, meanwhile, face the opposite problem. Different testing, reporting and liability rules across markets can produce more compliance without necessarily producing more safety.

That is the gap Singapore wants the UN to address.

A Warning from the UN’s Own Scientists

The timing has given the argument more weight. Just five days before Balakrishnan’s speech, the UN’s Independent International Scientific Panel on AI published a brief examining an incident in which AI agents being evaluated by OpenAI bypassed network restrictions, communicated across runs intended to remain separate, cheated an evaluator and compromised parts of the systems of OpenAI and Hugging Face, the AI model-sharing platform.

The panel did not claim that catastrophic loss of control was imminent. It did conclude that increasingly capable agents can discover loopholes and conceal behaviour, while no individual company or country is likely to see enough incidents to identify every emerging pattern.

Its broader preliminary assessment, published in July, reached a similar conclusion from a different direction. AI capabilities are advancing faster than existing safeguards.

Singapore is not alone in worrying about that mismatch. Prime Minister Lawrence Wong was among the leaders backing a Finland- and Norway-led call this month for frontier-AI developers to conduct pre-deployment testing and independent evaluations, for governments to share reports of serious incidents and for UN members to consider an international institution capable of setting standards and supporting verification.

Some international machinery already exists to build upon. The UN established its Independent International Scientific Panel and Global Dialogue on AI Governance after the Global Digital Compact, a diplomatic and co-operative framework rather than a binding international law. It relies on a multi-stakeholder model—bringing together governments, the private sector, civil society and technical communities—to guide responsible digital development.

The first Global Dialogue was held in Geneva in July, bringing governments, industry, researchers and civil society into the same process, with a second session scheduled for New York in May 2027.

Nevertheless, the UN still lacks a universally negotiated rulebook with the institutional heft Singapore is now asking countries to consider.

Nor would such a framework be the first international AI treaty. The Council of Europe opened its Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law for signature in 2024, describing it as the first legally binding international treaty in the field. It has attracted 21 signatures, including the US, Canada and Japan, and the European Union became the first to ratify it in May. The convention has not yet entered into force, however. That requires five ratifications, including at least three from Council of Europe member states.

Where the Proposal Meets Reality

A UN convention would attempt something broader and much harder: rules negotiated within an organisation whose membership stretches from the countries building the most powerful models to those worried they will have little say over technologies that could reshape their economies.

That is also where the proposal runs into reality.

The same competitive forces that make common safeguards useful make them difficult to agree on. Frontier AI is now part of the contest for technological and economic advantage. Governments may share an interest in preventing catastrophic failures but still disagree on testing thresholds, access to models, disclosure requirements, verification and how much authority any international institution should have.

For companies, those details will matter far more than another statement of principles. Common testing standards could reduce regulatory fragmentation. Mandatory independent evaluation or incident reporting could add substantial obligations for frontier-model developers. Rules around autonomous systems could influence everything from AI agents making payments to software operating parts of critical infrastructure.

The Business Case for Common Rules

Singapore appears to be making a wager familiar from its approach to finance, trade and technology. When something important crosses borders, common rules are preferable to regulatory chaos.

That position also has commercial logic. Singapore wants companies to deploy AI across finance, advanced manufacturing, healthcare and connectivity, four sectors that together account for more than 40% of its GDP. Its AI strategy depends on businesses trusting the technology enough to put it inside real operations, not merely running pilots on the side.

The faster those systems become autonomous, the more trust stops being an abstract ethical concern and starts becoming part of the economics of adoption.

That makes Balakrishnan’s nuclear-button question intentionally extreme, but the business version is already much closer to home. How much authority should a company give an AI agent over money, data, software or customers, and who answers when it gets the decision wrong?

Singapore has spent years trying to answer that question domestically without smothering experimentation. Its UN proposal tries to carry the same argument across borders.

The city-state is not asking the world to choose between AI innovation and AI safety. It bets the race will continue regardless. The question is whether the rules can catch up before the machines become considerably harder to stop.