The ice cream was rojak flavoured (rojak: a Singaporean/Malaysian "salad" that pairs umami fermented shrimp paste and chili with refreshing tropical fruits, topped with a layer of roasted peanuts). On the Monday of UN General Assembly week, Singapore opened a cork-built, waste-minimising ice cream parlour in Madison Square Park, the first physical outing for its new "When You're Ready" national brand, with Josephine Teo, the digital minister, doing the launch and a good part of the city's Singaporean community in the queue. It was a cheerful way to begin a week that would spend most of its hours in rooms without windows.
A mile north, the UN's Office for Digital and Emerging Technologies had gathered about forty people for a closed session with ISO and IEC, the two international standards bodies, on what the concept note called a network of interoperable cross-border AI governance sandboxes. I had been invited, which is the only reason a Singaporean chief of staff at a London technology firm was in the room. Sandboxes are familiar territory: DeepFlow, the company I work with, sits in the second cohort of the Financial Conduct Authority's AI sandbox, and I suspect that is why someone at the UN thought a person who has worked inside one might be useful to people designing them.
Over the next five days I sat in five more rooms. Foreign Policy's forum on technology and AI, where Teo shared a stage with Maria Ressa. A salon run by the Philanthropy Asia Alliance and the Equinix Foundation on why AI-for-good projects in Asia stall between pilot and scale. Rest of World's afternoon on the economics of AI. A scatter of Climate Week sessions on data centres and grids. And on the Friday, at the UK mission to the UN on Second Avenue, a tabletop exercise run by Chatham House and AI Safety Asia, in which a roomful of diplomats and officials worked through a regional response to a frontier AI incident.
Most of those sessions ran under the Chatham House rule, so what follows attributes nothing to anyone who was not speaking on a public stage. The figures are what speakers claimed. I have not checked them and would not repeat the wilder ones at a dinner party.
What I can report is that six sets of organisers with six different agendas kept hitting the same wall. Nobody disputed that AI works. Models that forecast hurricanes, tutor children in their own language and screen bank customers for financial crime exist and run, and the Asia roundtable's recurring line was that the technology had proved itself beyond experimentation. The argument everywhere was about the step after the pilot, and in particular about why what one place learns is worth so little anywhere else.
Take the sandbox session itself. The word suggests one kind of object. In practice it covers at least three. The FCA's version, the one I know best, is an infrastructure programme: the regulator supplies compute, synthetic data and models, firms build and test on them, and the FCA says in terms that taking part implies no approval of anything. Singapore's AI Assurance Sandbox supplies no infrastructure at all. It pairs a company with an independent tester and a methodology, and the regulator, by design, never sees the results. The EU's sandboxes under Article 57 of its AI Act are a third thing again, supervised testing with legal certainty attached, and their deadline slipped to August 2027 this summer because the harmonised standards they depend on were not ready. A network that tried to join these up would be joining three objects that share a name.
The room, to its credit, knew this. The ideas around the table were practical rather than grand: twin sandboxes across borders so that smaller countries can pool resources; fund participation so the global majority can take part rather than observe; use incident reporting and procurement standards to give sandbox findings somewhere to go. The line I underlined in my notebook was that a good sandbox should produce artefacts others can use. Everything else in the discussion was a version of that sentence. The repeated call for a common technical language was really a demand to be able to tell, from another country's test report, what was tested, how, and whether it would reproduce.
Jordan, one speaker mentioned in passing, now revises its national AI strategy every six months. It is roughly how fast the ground moves.
The same preoccupation with evidence turned up in rooms that had nothing to do with regulators. At Rest of World the argument was that evaluating an AI system has been made to seem insurmountable for small firms, that the real job is building tooling that drives the cost of evaluation down, and that the profession of AI evaluator has to stay in-country, because linguistic and cultural judgement is not something you outsource to a lab in another time zone. At the Asia salon, evaluation appeared instead as a lever: publish a benchmark for Southeast Asia's languages and the large labs suddenly find they care about Southeast Asia's languages, since nobody wants to look weak on a measured axis. Governments, the philanthropists said, will not take a pilot to scale without evidence they can hold in their hands.
Somebody in the sandbox session offered the analogy that organised the week for me. AI, they said, may need a Chicago Convention. Civil aviation has no single global law. It has national regulators certifying against shared standards, and certificates that other regulators recognise, which is why an aircraft inspected in one country is allowed to land in another. Nobody harmonised aviation law. They made the evidence travel. Put in those terms, the question for AI stops being how to write a universal rulebook and becomes something narrower and harder: what must a test report contain, and under what methodology, for a regulator who did not commission it to accept it? Drafting principles is easier, which may be why the world keeps drafting principles.
A second theme ran underneath the first. Sovereignty, that most absolute of political words, was being converted into a price.
In a session on disaster resilience, one speaker argued that sovereignty should be framed in practical terms as a question of what it is worth, and offered a number: in a scenario where the grid fails, eight hours of energy independence with guaranteed backup would cost Europe something like €34bn. Against that figure, the instinct to replicate hyperscale data centres inside national borders starts to look less like independence and more like an expensive way of being alone. The session concluded in favour of sharing data for disaster forecasting and of alliances among small states.
At the Foreign Policy forum the point arrived from the political side: sovereignty is about autonomy, and an autonomous government has choices, including the choice to operate across borders. At Rest of World it arrived sharpened. Safety and sovereignty, one panellist suggested, are partly industry-constructed categories, and serve the people selling the infrastructure at least as well as the people buying it.
Add those together and you get, more or less, the position Singapore has been articulating for some years: that small and middle powers do not buy autonomy by building everything at home, but by making themselves interoperable, trusted and hard to route around. The Friday tabletop was a live demonstration of the demand for this. Middle powers rehearsing a joint response to an incident none of them could manage alone, and discovering, as the exercise went on, how much of the response depended on arrangements that did not yet exist.
The third theme was the least comfortable for anyone who does governance for a living, me included.
Ressa's warning at the Foreign Policy forum was about information integrity, and her figures were bleak: by her account around three quarters of online traffic is now automated agents and slop, a quarter of the internet remains open, and the business model of journalism is dead. Every day governments fail to act, she said, they lose the power to set terms. Another panel noted that Americans were excited about AI five years ago and are mistrustful now, which puts the public some way ahead of its government.
But the more precise observation came from the sessions on work. People do not evaluate AI through frameworks; they evaluate it through their livelihoods. The displacement conversation keeps being had in aggregate while the first wave is specific, and by one panel's account it is arriving for women with a high school education doing administrative jobs. An unemployment rate reassures only if the jobs it counts pay a living wage. One speaker invoked Pope Leo on shared discernment and collective bargaining. Another noted that the institution Americans trust most in the AI conversation is the trade union.
The education findings at Digital Cooperation Day pointed the same way. Every euro into AI, one speaker argued, needs a euro into education. Nine countries, by another's count, account for 85 per cent of the world's best STEM talent. And the education systems that perform best are, apparently, the ones using the least AI. The skills problem was reframed in a way I found useful: the question is less how to train people for AI jobs than how to make sure people can work alongside AI with their judgement intact. At the Asia salon someone mentioned that Qatar has banned smartphones in schools, and that children go to school to learn how to talk to one another, and about family and culture, as much as anything else.
For anyone building governance architecture, the implication is awkward. The legitimacy of the whole project is being decided in a register the project barely addresses. A sandbox network, an evaluation profession and a recognition regime are all necessary. None of them has anything to say to the person whose job has just changed shape.
Climate Week deserves a paragraph, because something there had moved on. The question of whether AI is good or bad for the planet, which was never going to produce a useful answer, has been quietly retired. The serious sessions treated AI's footprint as an engineering and planning problem: electricity demand, water, grid connection queues, siting, and who pays. The same constraints, though, are pulling investment into the whole stack, from generation and grids to cooling, chips and the efficiency of the models themselves, and the arrow runs the other way too, as AI starts to be used on the physical systems it strains, from grid optimisation to disaster forecasting. The question that matters for the next phase is not how fast AI scales but how intelligently the infrastructure around it is built and where its capabilities are pointed. That is a planning question in an engineering costume, and the evidence about what AI's infrastructure costs and delivers will need to travel across borders every bit as much as evidence about model safety.
The most useful idea I heard all week was also the simplest, and it came from the Rest of World stage: stop waiting for new AI law. Existing human rights law, the rights to education, to labour, to privacy, already gives legal footing, and human rights due diligence and impact assessment turn those rights into something a company can build against. Grounding AI obligations in law that exists, with courts and jurisprudence behind it, beats spending years redrafting rules that largely restate obligations we already have. The panel's test for any governance mechanism was threefold: does it include the communities most at risk, does it hold governments to account as well as companies, and does it give someone the tools to challenge the law?
The tabletop exercise ended the week on the same note from the opposite direction. Working through a regional response to a frontier AI incident, the gaps that mattered were definitional and institutional rather than technical. Whether an internally misaligned agent has committed a crime at all. How agents are identified and contracted with. What minimum redundancy a critical system owes the people who depend on it. Which part of which government holds the relevant files. The exercise's clearest finding was that crisis councils arranged in advance would have cut the time to coordinate. The mechanism has to exist before the emergency does.
That is the thread. The human rights argument says the legal mechanism already exists. The aviation analogy says the institutional pattern does, and can be copied. The sandbox session's best idea asks for evidence that is already being produced to be made portable. The crisis council asks for coordination capacity to be built before anyone needs it. In six rooms, nobody was short of invention.
I came home with a shorter list of things worth doing than I left with, which is the usual result of a good week. Evaluation infrastructure that makes evidence cheaper. Professions that keep the people who produce it in-country. Sandboxes built to generate artefacts another regulator will accept. Procurement standards that let a government buy AI on evidence rather than faith. And coordination mechanisms set up before the incident that will test them. By accident of nationality and employment I sit across two of the jurisdictions furthest along: the UK's infrastructure model and Singapore's assurance model. They are not the same experiment, which is what makes the pair worth watching.
The ice cream, for the record, was very good.