Singapore has called for countries to consider a United Nations convention on artificial intelligence (AI) safeguards, warning that geopolitical rivalry and commercial pressures have made halting the technology’s development unrealistic.
To continue reading, subscribe to Eco‑Business.
Access all our exclusive content and archive — with a 7-day free trial. Then just US$5/month. Cancel anytime.
- Exclusive access to training across business and policy topics
- Direct access to industry leaders at Eco-Business events
- A seat among the decision makers shaping the region's sustainable development
Foreign Minister Vivian Balakrishnan told the UN General Assembly in New York on 26 September that governments should pursue common safety rules while continuing to invest in AI, citing threats ranging from the misuse of the technology to develop biological weapons to humans losing control of autonomous systems.
He also raised the possibility of an international institution with responsibilities for AI comparable to the technical standard-setting functions of the International Telecommunication Union or the verification work of the International Atomic Energy Agency (IAEA).
“We must explore all options, including the idea of a UN framework convention on AI safeguards,” Balakrishnan said.
His remarks outlined an initiative for countries to consider, without setting out a negotiating timetable, enforcement arrangements or detailed treaty provisions.
The proposal comes as researchers document advances in AI’s ability to perform complex tasks, while cautioning that evidence of potentially dangerous capabilities does not establish how readily they could cause harm outside controlled settings.
The International AI Safety Report 2026, an assessment drawing on contributions from more than 100 independent experts, found that general-purpose AI systems can provide technical information and help troubleshoot laboratory procedures that could assist attempts to develop biological or chemical weapons.
However, practical barriers to producing such weapons mean there remains substantial uncertainty about how much those capabilities increase real-world risks, the report said. Some of the same capabilities also support legitimate medical research, complicating efforts to restrict harmful uses.
Balakrishnan said Singapore’s approach included testing systems before deployment, restricting the actions autonomous systems can take and ensuring people can intervene when they behave unexpectedly.
“Most importantly, human beings must ultimately remain in control and accountable,” he said.
He argued that a pause in development was no longer realistic given competition between major powers and the financial incentives facing companies developing the most advanced models.
Research points to risks, but also uncertainty
Autonomous AI agents, which can carry out tasks with limited human intervention, present particular oversight challenges because people have fewer opportunities to interrupt errors, according to the safety report.
The assessment said researchers differed widely over the likelihood of a future loss of control. It found early warning signs in laboratory tests, including models disabling simulated oversight, but said systems assessed had not demonstrated capabilities at levels that would enable such a loss of control.
The economic implications extend well beyond those security concerns.
A separate 2025 study by the International Labour Organization and Poland’s National Research Institute, NASK, estimated that one in four workers worldwide held an occupation with some exposure to generative AI – technology capable of producing text, images and other content.
Exposure reached 34 per cent of employment in high-income countries, compared with 11 per cent in low-income countries. Clerical occupations were the most exposed, although the technology’s expanding capabilities were also affecting professional and technical tasks.
The study measured the potential for tasks to be affected, rather than forecasting actual job losses. It concluded that changes to jobs were more likely than their wholesale elimination because most occupations still required human input.
Trust remains an obstacle
Balakrishnan said international safeguards would require cooperation among countries competing for technological advantage, including confidence that risks would be disclosed and safety testing would be credible.
Any lasting arrangement should give smaller and less technologically advanced countries a meaningful role, he said, rather than leaving decisions to those with the most powerful models, financial resources or military capabilities.
The UN has already established mechanisms intended to support wider participation. In August 2025, the General Assembly agreed to create an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance, combining scientific assessment with international policy discussions.
Balakrishnan cited those initiatives as foundations for further cooperation, alongside the International Telecommunication Union’s AI for Good Global Commission, an international initiative to define practical pathways for trusted AI and responsible global innovation.
Even where governments disagreed on regulation, they could begin by developing a shared understanding of AI capabilities, comparable safety tests and arrangements for rapidly reporting serious incidents across borders, he said.
