By Poss Apostolou

For years, AI sat at the edges of the UN General Assembly, a subject for side events and panel sessions. This year, it moved to the centre of the room.

In his final General Assembly address, Secretary-General António Guterres urged world leaders to regulate AI. A day later, the Security Council held a high-level briefing on AI and international security. Yoshua Bengio, co-chair of the UN’s Independent International Scientific Panel on AI, spoke alongside the heads of OpenAI, Anthropic and Hugging Face. On the same day, Australia’s Prime Minister revealed that an AI agent had accessed files on a government website.

That last detail matters most. This wasn’t a hypothetical scenario or a warning about the future. The risks being debated in New York are already disrupting public services and the people working on them.

Different answers, one shared problem

Our leaders (to nobody’s surprise) did not agree on what to do. The United States rejected any global scheme to control AI. It argued that international bodies should instead help countries share best practice and build their own capacity. China called for continuous improvement of regulatory frameworks, emergency response and cross-border cooperation.

That split matters, but it hides some common ground. Binding global rules, voluntary cooperation or something in between: whatever international governance ends up looking like, every government still needs the same thing. It needs the capability to understand, adopt and oversee AI at home. International agreements only work if the institutions expected to carry them out are ready to do so.

That common ground is where we work. This week, the Oxford Insights (OI) team has been making that case in two very different rooms, 6,000+ miles apart.

In New York: scaling industrial AI responsibly

Eleonora Lamm, our International Project Lead, spoke on the panel at Scaling Industrial AI Responsibly, a UNGA side event hosted by UNIDO and IBM. She was joined by UNIDO Director General Gerd Müller, Barbados Minister Jonathan Reid and Leslie Teo of AI Singapore.

The event marked the soft launch of UNIDO’s AI & Digital for Industry Navigator (AIDIN), a diagnostic tool that assesses industrial AI maturity across five enabling pillars, and which builds on our recent methodology work with the organisation.

The discussion focused on a problem we see again and again: plenty of AI pilots, far fewer that scale. For developing countries and emerging economies, the question is how to move from one-off projects to pathways that countries lead themselves, and that reach the small and medium-sized firms that make up most of their industrial base.

In Tashkent: sovereignty is a choice, not an escape

Meanwhile, at ICT Week Uzbekistan in Tashkent, our CEO Richard Stirling gave a keynote on the global AI landscape. His argument went straight to the tension on display in New York.

In 2026, it is tempting to see AI as a two-horse race between the United States and China. Richard argued that this is too simple. To understand where any country stands, you need to look at the whole stack, layer by layer: energy, chips, compute, data, models and applications. Look closely and a different picture emerges. The US leads on frontier models but depends on chips made in Taiwan. China has cheap, open models and huge amounts of renewable power, but is still short of advanced chipmaking tools. Nobody owns the whole stack.

So sovereignty isn’t about escaping dependencies. It’s about choosing them. Richard set out four tests for any dependency:

• Named: do you know which layer you depend on, and on whom?
• Hedged: is there a second source, or a way out?
• Traded: do you get something back for accepting it?
• Verified: can you inspect and test what you run?

A dependency that passes these tests is a choice. One that fails them is an exposure.

For Uzbekistan, that framing is practical, not theoretical. The country ranks fourth in its region in our Government AI Readiness Index, well above the global average. It has real strengths: a young workforce, a growing base of services exporters and its own language and state data. Its base models are all foreign, however, and its regional power grid is under strain. Richard’s advice was to start with the data layer. That means finishing the national data inventory ordered in August, funding Uzbek-language data as infrastructure, and adding AI questions to the Statistics Agency’s business survey so the government can see how AI is actually being adopted.

At the same event, our Principal Consultant Pablo Fuentes Nettel, joined a panel on responsible AI.

Evidence over assumption

Our Government AI Readiness Index has been published annually from 2017. Since then, it has measured how prepared governments are to use AI in the public interest. It looks at government capability, the strength of the technology sector, and data and infrastructure, as well as governance, safeguards and public trust.

The Index isn’t a league table for its own sake. It asks a practical question: is this government equipped to use AI well, safely and for the people it serves?

This week showed why that question matters:

• When leaders disagree about global rules, governments need a clear, independent picture of where they stand and what to prioritise.
• When AI agents can reach public systems, governance and security can’t be left until later. They are part of the foundations.
• When the benefits of AI risk going mainly to early adopters, measuring readiness, whether for a country or an industrial sector, is how we make sure no one is left behind.

If you’d like to talk about how your government or organisation can prepare for what comes next, then please get in touch.