UN AI Safeguards Convention: Singapore's Proposal Explained

Singapore told the United Nations on 26 September 2026 that the world should start work on a UN Framework Convention on AI Safeguards. The idea came from Foreign Affairs Minister Vivian Balakrishnan, in Singapore's national statement at the general debate of the 81st UN General Assembly session in New York. The Ministry of Foreign Affairs published the summary and the full transcript the next day.

"We must explore all options, including the idea of an UN Framework Convention on AI Safeguards," he said. He added that the world may also need an international institution doing for AI some of what the ITU does for technical standards and the IAEA does for verification.

It was offered as a proposal, not an announcement. In the same statement Singapore confirmed it had joined the call for greater safety and security of frontier AI models led by Finland and Norway. Prime Minister Lawrence Wong signed that joint statement on 22 September, together with 21 other world leaders including Canada's Mark Carney and Australia's Anthony Albanese. The letter asks AI companies for transparent safety protocols, mandatory pre-deployment testing and independent evaluation, with qualified evaluators given enough access to actually check the work.

The full national statement is on the MFA website, with the AI section in paragraphs 20 to 31, and CNA's report carries the summary.

The speech is more measured than the headlines that followed it. Balakrishnan's own metaphor was a high-performance car. Going fast needs a powerful engine, but it also needs good brakes, and once you cross out of your own neighbourhood you need shared rules: traffic lights, insurance, seatbelts. Innovation is the engine, safeguards are the brakes.

"The genie is already out of the bottle. It is too late to call for a halt, especially given the superpower rivalry and the huge financial incentives for companies who already have frontier models," he said. He named three categories of risk: losing control of autonomous systems; rogue actors using AI to build bioweapons or other weapons of mass destruction; and the economic, social and political disruption that is only starting to show. Then he asked the room a blunt question. Should the nuclear button be in the hands of a human being or an AI?

Singapore's own answer is the part local employers should read. Its stated approach: rigorous testing and evaluation before deployment, clear limits on what autonomous systems are allowed to do, mechanisms to step in when systems behave unexpectedly, and human beings who stay accountable. That is close to what a decent IT risk register already looks like.

A framework convention is not a ban and not a regulator. It is a broad treaty that sets out principles and creates the institutions to act on them, the same structure the UN Framework Convention on Climate Change used before countries negotiated detailed commitments on top of it. The Straits Times made the same comparison. Think of it as agreeing on the shape of the table before anyone argues about the menu.

What makes this interesting is Singapore's diagnosis of the obstacle. "The scarce resource today is not legal architecture or technology. It is trust," Balakrishnan said. Any arrangement has to give every country a meaningful stake, not only the states with the most advanced models, the deepest pockets or the largest arsenals. Since governments disagree about governance, Singapore suggests starting where agreement is easier: a shared scientific understanding of what AI models can and cannot do, comparable methods for testing and evaluation, and fast cross-border reporting of serious AI incidents.

Three existing UN platforms are named as the foundation: the Independent International Scientific Panel on AI, the Global Dialogue on AI Governance, and the ITU's AI for Good Global Commission. Singapore also offered to keep making its own governance frameworks and testing tools available to others.

The honest caveat: nothing has been negotiated. In a wrap-up interview with Singapore media the same day, he said it is "too early to conclude yet, but at least the discussions have begun". Getting 193 member states to sign one treaty takes years.

No law changed this week, and no Singapore company picked up a new obligation because of a speech in New York. Singapore's approach stays voluntary and framework-based: IMDA's Model AI Governance Framework for Agentic AI, updated to version 1.5 in May 2026, plus the AI Verify testing toolkit, which lets a firm self-assess its systems against 11 recognised principles.

But the direction of travel is now easy to read. If a convention is ever agreed, the obligations it would start from are the ones Singapore already tests for: evidence that you evaluated a system before deploying it, a named human who is accountable, limits on what an autonomous agent may do on its own, and a way to record what happened when it does something you did not intend. A UN treaty would make that international; your bank customers and your insurer are already moving that way. The practical job today is being able to answer those questions about your own AI use, which is what our State of AI in Singapore report tracks across the wider policy picture.

For workers, the speech points at a familiar direction. Balakrishnan argued the real productivity gain "comes not from the IT department, but from our businesses and workers having the skills, the tools, and the wherewithal to re-engineer their work processes". That is the same message behind the public service training push we covered in our 150,000 officers guide. SkillsFuture AI courses and the National AI Impact Programme remain the local levers, treaty or no treaty.

Replace the square brackets with your own details, and keep NRIC numbers, customer contacts and client records out of anything you paste into a public tool.

1. A one-page AI use policy for your team

Write a one-page AI use policy for a [11-person F&B business / 25-person logistics firm / tuition centre] in Singapore. Cover which AI tools staff may use, what information must never be entered into a public AI tool, who must review AI output before it reaches a customer, and what staff should do when an AI tool gives a wrong answer. Plain English, one page, no legal jargon.

2. A risk review of one AI workflow

Act as an IT risk reviewer. Here is one AI workflow we use: [describe it, for example an agent that reads customer enquiries from our inbox and drafts replies]. List the three worst ways it could fail in a way that costs us money or breaks PDPA, the human checkpoint we should add at each one, and what we should log so we can explain later what happened.

3. An internal note when an AI tool gets it wrong

Help me write a short internal note about an AI tool producing a wrong [quote / delivery date / fee figure] that reached a customer. Cover what happened, when we noticed, who checked it, what we told the customer, and what we changed so it does not repeat. About 200 words, neutral tone.