What a Meaningful AI Slowdown Would Require—and Singapore’s Role

A meaningful AI slowdown would pace frontier development rather than pause AI across the economy. Recent public calls from leading AI executives have put independent evaluation and coordination on the agenda, but turning those proposals into credible safeguards requires more than voluntary promises: evaluations must be independent, release decisions must have consequences, and rules should scale with risk.

Singapore has useful building blocks for that work, including an agentic-AI governance framework, AI Verify testing tools and a programme to recognise third-party AI testers. The next challenge is connecting those tools to clear, evidence-based decisions about the systems people actually deploy. 4

What the calls to slow AI do—and don’t—establish

The recent proposals focus on the pace of frontier AI: the most capable systems, where developers’ safety measures and external oversight may struggle to keep up with capability advances. Proposals reported in the coverage include giving outside evaluators ongoing access to frontier companies, coordinating safety standards and pursuing broader international cooperation.

Those proposals signal growing interest, not a settled policy or binding agreement among AI companies. The International AI Safety Report describes risk-management efforts as still developing and notes that most company initiatives remain voluntary. That distinction matters: a company’s commitment to test a model is not, by itself, an enforceable rule requiring it to delay a release when serious risks remain.

Make a slowdown a risk-based release gate

A practical approach would set stricter conditions for systems with greater capabilities or authority, while avoiding a blanket pause on routine AI use. A low-impact drafting tool and an agent authorised to take consequential actions should not automatically face the same requirements. Singapore’s agentic-AI framework already recommends assessing and bounding risks and setting human approval checkpoints for significant or sensitive actions. 46

For frontier models and high-impact deployments, policymakers could build on that principle with a release process that includes:

• Independent pre-release evaluation: Give qualified external evaluators secure access to relevant systems and evidence, and assess dangerous capabilities and foreseeable misuse. Ongoing evaluator access has been proposed in recent calls to pace frontier development; it is not a universal existing requirement.
• Defined thresholds and follow-through: Publish or otherwise clearly establish what findings trigger additional safeguards, restricted access or a delay. A failed evaluation should lead to a documented response and, where risk warrants it, a release being held back—not merely another internal review. These are proposed governance measures, not requirements described in Singapore’s current framework. 4
• Controls matched to the deployment: For agents given consequential permissions, consider bounded access, human approval at important action points, monitoring and clear accountability. Singapore’s framework supports risk-bounding and meaningful human oversight; the exact controls should reflect the system and its use. 48
• Reassessment when systems materially change: Evaluate significant changes to a model, its tools or its deployment context rather than treating an earlier test as a permanent guarantee. This is a proposed way to keep assurance relevant as the deployed system changes.

Lower-risk applications should have a clearer, lighter route to use. Risk-based rules can make room for business adoption without treating a successful test as proof that a system is safe in every context. The EU AI Act is one example of regulation organised around the risks of AI uses.

Keep the rules public, and the testing independent

International coordination could make evaluations more comparable and reduce pressure to seek out the least demanding oversight. Governments can work toward shared testing methods, incident-sharing and minimum expectations for the highest-risk systems; independent evaluators can then produce evidence against those expectations. The recent slowdown proposals likewise call for coordination between companies and across borders.

The division of responsibility is important. Public authorities should set enforceable thresholds and determine the consequences of crossing them. Companies can contribute technical knowledge, but should not be the only judges of whether their own systems meet the bar. Independent testers need transparent methods and disclosed conflicts of interest. These are design principles for a proposed system of oversight, rather than features established by the reported industry calls.

What Singapore can build on

Singapore’s Model AI Governance Framework for Agentic AI provides guidance for organisations deploying agents, including risk assessment, human accountability and controls. It is guidance for responsible deployment, not, by itself, an enforceable release gate. 468

AI Verify offers testing tools, while the AI Tester Accreditation Programme recognises third-party testing firms with relevant technical capabilities and operational readiness. Together, these initiatives could support more consistent assessments across business and finance. Singapore could also encourage shared assurance profiles and contracts that clarify what model providers, system integrators, deployers and testers are each expected to demonstrate. Those are potential next steps, not policies established by the cited frameworks.

A useful aim would be cross-border recognition of test evidence and methods, not a claim that one national certificate guarantees safety for every system or use. The framework and testing tools can help make evidence more structured; authorities and organisations still need to judge whether that evidence addresses the real deployment and its risks.

Assess AI safety, software reliability and cybersecurity together

Testing a model’s outputs is only part of evaluating an AI agent. A complete assessment should also examine the software system around it and the security of the tools and permissions it can access. These are related but distinct questions: what harmful behaviour can the AI produce; how does the assembled application behave when something fails or changes; and can an attacker compromise or redirect the system?

This combined assessment is a recommended extension of assurance, not a claim that every existing tool already covers all three areas. The AI Verify Toolkit describes technical tests for traditional machine-learning systems, including fairness, explainability and robustness. That scope makes it important to specify what further end-to-end testing is needed for agentic deployments, rather than assume that a model-level result certifies the full system.

The test of a meaningful slowdown

A slowdown is meaningful when independent evidence can change what happens next: a high-risk release can be delayed, restricted or corrected if safeguards are inadequate, while ordinary, lower-risk uses retain a proportionate path to deployment. Singapore’s frameworks and testing programmes offer a foundation for building that kind of assurance, but the release thresholds, enforcement and scope of evaluations would need to be made explicit.