In April 2025, Singapore's Infocomm Media Development Authority (IMDA) and the AI Verify Foundation convened the Singapore Conference on AI (SCAI), a closed-door gathering of researchers from frontier AI labs, academia, and governments across the United States, European Union, China, the United Kingdom, India, and beyond. The output — the Singapore Consensus on Global AI Safety Research Priorities — is, on its face, a sober technical document. Read more carefully, it is something more interesting: a quiet diplomatic experiment in whether a small, jurisdictionally neutral state can do for AI what Geneva once did for arms control and Bretton Woods did for postwar finance.
The Consensus does not propose a new treaty, a new agency, or a new licensing regime. It catalogues research priorities — areas where labs in San Francisco, Beijing, Paris, and London actually agree there are open technical problems worth solving together. That includes evaluations for dangerous capabilities, interpretability research, methods for verifying training compute and model provenance, and shared standards for incident reporting. It is, in essence, an attempt to keep the technical conversation moving even as the political conversation hardens.
Why Singapore, and why now
The global AI governance landscape in 2026 is a study in fragmentation. The EU AI Act's general-purpose AI obligations kicked in across 2025 and are being phased through 2026, with the AI Office in Brussels writing codes of practice that European and American labs are still negotiating in real time. The United States has retreated from the previous administration's AI Executive Order and is leaning on state-level frameworks — Illinois SB 315, California's SB 53-style transparency rules, and New York's draft frontier model requirements. China is enforcing its generative AI service measures while pushing its own state-led model of "AI for good." The UK's AI Safety Institute is now a fixture; so is the US AI Safety Institute, though its funding and remit remain contested.
Into this gap, Singapore has stepped with a characteristically pragmatic offer: not to lead, but to host. The city-state has no domestic frontier lab, no national champion to protect, and a small enough regulatory surface that it can move quickly without the political theatre that surrounds AI rules in Washington or Brussels. Its AI Verify Foundation, launched in 2023, has spent two years quietly building open-source testing toolkits that have been adopted by multinationals as a de facto compliance baseline.
What the Consensus actually says
The document avoids the two failure modes that have dogged most international AI declarations. It does not, like the 2023 Bletchley Declaration, gesture at "existential risk" without operational follow-through. Nor does it, like much of the EU AI Act drafting, attempt to fix in legislative concrete a technology that is still changing every quarter. Instead, it identifies a tractable research agenda:
- Evaluations and red-teaming — shared benchmarks for biosecurity, cyber, and autonomy risks so that a Chinese lab and a US lab can at least agree on what "dangerous" means before arguing about thresholds.
- Interpretability — mechanistic research into what models are actually doing internally, which the document treats as foundational to any meaningful safety claim.
- Compute and provenance verification — technical methods for verifying training claims without forcing labs to disclose proprietary architectures.
- Incident reporting infrastructure — standardised formats for sharing safety incidents across borders, analogous to aviation's near-miss reporting.
None of this is glamorous. All of it is the connective tissue that real governance regimes — pharmaceutical safety, civil aviation, nuclear nonproliferation — actually run on.
The case for proportionate, technical governance
The Singapore approach is closer to what good AI policy should look like than the regulatory ambitions emerging from Brussels or some US state capitals. Three reasons.
First, it is capabilities-led, not category-led. The EU AI Act sorts systems into risk tiers based largely on use case. That made sense for pre-foundation-model AI; it has aged poorly. A general-purpose model can be used for credit scoring or for poetry. Singapore's Consensus focuses on what models can do — capability evaluations — rather than where they might be deployed.
Second, it is multistakeholder without being captured. The signatories include researchers from frontier labs, academic institutes, and government safety bodies. That mix matters: pure intergovernmental processes tend to lag the technology by years, while pure industry processes lack legitimacy. Singapore has historically been good at convening across this divide, and the Consensus benefits from that institutional muscle memory.
Third, it explicitly preserves room for jurisdictional difference. Nothing in the document binds the US to the EU's prohibitions or the EU to China's content controls. It is an agreement on what to study, not what to ban — and that distinction is what makes it survivable across the geopolitical fault lines that have already fractured every other AI dialogue.
What this means for the rest of the world
The Singapore model deserves to be copied. Smaller, technically credible jurisdictions — Switzerland, the Netherlands, Israel, Canada, India's emerging AI safety community — can play a similar convening role on more specific problems. India in particular, hosting the AI Impact Summit successor and with the largest AI developer population outside the US and China, has an obvious opening to anchor a Global South-inclusive technical track.
The harder lesson is for the bigger blocs. Brussels should resist the temptation to treat its codes of practice as the global default; their extraterritorial reach is already drawing pushback from US labs and a growing list of trading partners. Washington, meanwhile, needs to decide whether its AI Safety Institute is a research body or a regulator-in-waiting. China should be engaged where it is willing to be engaged — and the Consensus suggests that is further than the political rhetoric implies.
The biggest risk to the Singapore process is that it becomes another conference circuit — annual photo-ops generating documents nobody implements. The next test is whether the research priorities translate into joint funded work programmes, cross-lab evaluation runs, and a real incident-sharing infrastructure within the next 12 to 18 months.
For now, though, the Consensus is a useful reminder that the most productive AI governance in 2026 is not happening in legislatures. It is happening in technical workshops, often in places most people cannot find on a map without zooming in.