October 29: Singapore Prime Minister Lawrence Wong has proposed a pilot for countries to share lessons from serious artificial-intelligence incidents and near misses. Speaking at an October 28 virtual leaders’ meeting co-hosted by Norway and Finland, he called for practical, voluntary cooperation that protects legitimate security and commercial information.
Key highlights
- Wong proposed starting with a pilot rather than an elaborate new international structure.
- Participants would work on common incident categories and information to share.
- The proposal includes lessons from serious failures and near misses encountered by AI users.
- His remarks do not announce a new mandatory reporting law.
What the proposal would test
Norway’s Prime Minister’s Office said nearly 30 countries joined the meeting, following an initiative launched with Finland during the UN General Assembly period in September. Its account also identified incident reporting, protection of children and AI-generated content’s effects on democratic debate as concerns raised by participants.
In the transcript published by Singapore’s Prime Minister’s Office, Wong suggested agreeing categories, exploring secure collection and exchange, and testing practical difficulties before considering a broader framework. He used aviation’s reporting and learning culture as an analogy for improving safety.
He also argued that experience should not be limited to model developers. Governments, businesses and other users can encounter failures that reveal risks even where actual harm has not occurred or an existing reporting rule has not been triggered.
From an incident to a useful lesson
| Design question | Why it matters |
|---|---|
| What counts as an incident? | Comparable categories help participants interpret reports |
| What information is needed? | Technical context can distinguish similar-looking failures |
| How is it shared securely? | Reports may contain sensitive operational information |
| Are near misses included? | A failure can reveal a risk before harm occurs |
| How are lessons used? | Findings need to inform prevention and testing |
This table explains the design questions raised by the proposal. It is not a checklist already imposed on companies or a description of an operational reporting portal.
Why near misses deserve attention
A system can fail in a way that is caught by a human reviewer before it reaches a customer. Recording that event can reveal a weakness in the process, even when the eventual outcome is harmless. Counting only realised damage risks losing the information needed to improve safeguards.
Useful reporting also requires context. A model response, the task it was given, its access to external tools and the review process can all affect the meaning of a failure. A bare claim that an AI system made an error may be too vague to reproduce or prevent it.
At the same time, more detail is not automatically better if it exposes private data or security vulnerabilities. The pilot’s proposed secure exchange and common categories would need to make learning possible while managing those constraints. Participation, access controls and publication of findings remain important design choices.
What to watch
Wong said cooperation should build on existing scientific efforts, including the Singapore Consensus on Global AI Safety Research Priorities. The next concrete developments would be participating countries, a pilot scope, reporting definitions and a timetable.
Until those are announced, the news is the proposal itself. Businesses should continue to follow applicable current rules and their own incident-response procedures rather than treat the speech as a newly enacted obligation.
Sources
- Prime Minister’s Office Singapore: Lawrence Wong’s remarks at the virtual leaders’ meeting on frontier AI safety, October 28, 2026
- Norway Prime Minister’s Office: Seeking international principles for artificial intelligence (Norwegian)
The Tech Desk covers technology, artificial intelligence and cybersecurity. Stories are checked against at least two credible sources before publication.
