How to Build a Global AI Safety Regime That Can Be Independently Verified

A global AI safety regime will be trusted only if its safeguards can be checked independently and enforced when they fail. That calls for common standards, national regulators with real powers, and testing that goes beyond a developer’s own assurances. Singapore’s Josephine Teo has drawn a parallel with aviation: layers of safeguards helped make routine use possible, and AI needs its own approach as adoption grows. 36

What aviation actually teaches

International aviation rules developed in stages. National laws and bilateral agreements preceded the 1919 Paris Convention; the 1944 Chicago Convention established a new framework, and the International Civil Aviation Organization (ICAO) came into being in 1947. 313935 The lesson for AI is not to copy an aviation rulebook. It is to start with workable agreements and strengthen them as countries establish shared practices.

A comparable AI framework could assign international coordination to a body that develops minimum standards, accredits evaluators and compares incident data. National authorities would retain powers to inspect, require changes, restrict deployment and impose penalties. This division would make a shared standard useful without treating an international certificate as a substitute for domestic enforcement.

Safety would also have to continue after launch. Testing before deployment should be followed by operational monitoring, reassessment after material changes, and plans to roll back or retire a system. Any approval should specify the system version and the setting in which it was evaluated—not imply permanent approval of every future use.

Turn safety claims into checkable evidence

Independent verification requires more than a badge or a published benchmark score. An accredited evaluator would need secure access to the relevant model, deployment configuration, documentation and records of significant failures. Its methods and findings should be documented well enough for another qualified evaluator to challenge them, while access controls protect personal data, trade secrets and security-sensitive information.

For high-risk systems, a developer could be required to present a safety case: what the system may do, how it was tested, which safeguards were demonstrated, what risks remain and how incidents will be handled. If the evidence is inadequate, the regulator should be able to require remediation or limit deployment. The standard is verifiable compliance with defined requirements, not a promise that harm is impossible.

Funding matters, too. Large developers could pay into a pooled or regulator-administered testing fund rather than choose and directly control the evaluators who judge them. Independent investigation of serious failures and near misses should prioritize learning and prevention. Protected reporting would encourage disclosure, but it should not excuse concealment or misconduct.

Where Singapore could help—and what trust would require

Teo has argued for building safeguards alongside AI adoption, and reporting identifies Singapore’s AI Verify toolkit and AI Safety Institute as foundations for its work on testing and assurance. 13 Singapore could host joint evaluations or confidential technical exchanges within a wider network of accredited laboratories.

Hosting a facility, however, would not automatically make it trusted by both the United States and China. That would depend on transparent governance, secure access rules, independent appeals and demonstrated willingness by both countries to participate. A distributed network would also avoid making one hub indispensable.

Start with shared risks, not consensus on an AI slowdown

The politics of AI safety remain divided. Reuters reports that Dario Amodei, Sam Altman and Elon Musk have called for a coordinated slowdown, while Mark Zuckerberg has favored market-led safeguards and Jensen Huang has dismissed the need for new regulation. 4748 China has rejected slowdown calls, and President Donald Trump said he did not expect movement on AI guardrails at his September 24 summit with Xi Jinping. 5850 Reporting on the summit described low expectations for a major breakthrough. 49

Countries need not settle those broader disagreements before addressing hazards each wants to avoid. Reporting on proposals from US and Chinese security experts points to red lines around nuclear systems, human control over consequential cyberattacks and a hotline for incidents involving autonomous AI. 20 Those could be starting points for emergency contacts, serious-incident notification and reciprocal technical discussions. US–China participation would matter, but other countries would need a voice in setting standards and access to evaluation capacity.

Borrow selectively from arms control and financial supervision

Arms-control agreements offer a design principle: negotiate reciprocal commitments narrow enough to verify. A future AI treaty could establish a standing body for coordination and compliance review, while leaving technical testing methods open to revision. Neither a nuclear treaty nor ICAO is a ready-made template for AI.

Financial-style supervision offers a second principle: concentrate demanding obligations where failures could have the greatest impact. High-risk developers and deployers could face independent assessments, management accountability, recovery planning and prompt reporting of serious safety failures, material data leaks, hacking or model theft. Smaller firms offering low-risk services could use shared testing infrastructure and simpler documentation. Size alone should not create an exemption for a dangerous system.

A treaty and standing institution by 2030 would be an ambitious policy target, not a reliable prediction. A more credible first measure of progress is whether countries can agree on incident definitions, evaluator independence and reciprocal safeguards—and whether regulators can act when the evidence falls short.