AI governance: The summit stage is necessary but it isn’t sufficient

TEN years ago, a street vendor in Mumbai could not open a bank account. No address, no papers and no access. Today, that same vendor accepts digital payments on her phone, instantly, for free, from anywhere in the country.

That is a civilisation story as much as a technology story – one that India built, at scale, through deliberate sovereign choice.

French President Emmanuel Macron told that story at the India AI Impact Summit at New Delhi in February. He told it because it captures something essential about what the event represented: a demonstration that the Global South can shape the rules of the game, rather than inherit them.

The summit was, by any measure, an extraordinary moment. Almost 300,000 people from more than 100 countries gathered at Bharat Mandapam for the largest and most geographically representative artificial intelligence (AI) summit yet held, and the first hosted in the Global South. Heads of state, frontier lab CEOs and civil society shared the same stage.

The New Delhi Frontier AI Impact Commitments produced very limited but concrete pledges from Anthropic, Google, OpenAI and other frontier AI firms.

They include sharing anonymised economic data on AI’s impact on jobs to inform evidence-based policy and strengthening multilingual evaluations to ensure that AI works across Global South languages and contexts, not just English.

And yet the summit also made the challenge visible. Dario Amodei, CEO of Anthropic, was unusually direct: We are “only a small number of years” from AI surpassing the cognitive capabilities of most humans across most domains – what he called “a country of geniuses in a data centre”.

Paris-based General-Purpose AI Policy Lab’s (GPAI) recent assessment reinforces the urgency: Almost all current AI benchmarks, including the most difficult ones, will likely be saturated before 2030. Critical capabilities in cybersecurity and AI research and development automation follow the same trajectory, with security and control challenges likely to emerge by 2028 at the latest.

Antonio Guterres, secretary-general of the United Nations (UN), framed the governance problem plainly: “The future of AI cannot be decided by a handful of countries or left to the whims of a few billionaires.”

AI governance infrastructure must be built faster

The gap between that pace of development and the pace of governance is where the real work lies.

Voluntary commitments from frontier labs signal intent and set reference points for accountability. But with trillions of dollars of market share at stake and intense competitive pressure between a small number of well-resourced private actors, soft norms alone cannot bend behaviour towards safety.

What is needed is binding: a set of minimum, internationally coordinated regulatory backstops that can be standardised across jurisdictions, underpinned at the most extreme risk levels by strategic coordination between the United States and China – the foundation on which culturally diverse national frameworks can then be built.

Three specific gaps are equally urgent now:

• Incident reporting: Frontier labs currently lack standardised mechanisms for sharing information about serious incidents before they become public crises, leaving problems siloed until the damage is done.
• Verification: Without credible mechanisms to confirm what others are actually doing, mutual commitments remain difficult to trust and harder to sustain.
• Evaluation standards: Assessments of AI capabilities, safety properties and real-world impacts need to work across borders, languages and institutional contexts – not just in English, and not only for the systems built by the companies with the largest research budgets.

As Tata Group chairman Natarajan Chandrasekaran put it from the same stage in New Delhi: “We are standing at a moment where the scarce resources are trust, stewardship and human capability.”

That framing gets closer to the actual problem than most talk about compute and capital.

The context has also shifted in ways that add urgency. The geostrategic environment is deteriorating. AI is no longer just a technology story; it is embedded in geopolitical competition and high-intensity warfighting in ways that make coordination harder and more necessary at the same time.

Governance infrastructure that might have been built incrementally now needs to be built faster.

There is a precedent worth recalling. During the Cold War, when official channels between East and West were effectively closed, the Pugwash Conferences on Science and World Affairs created the space for scientists and policymakers to talk candidly across ideological divides.

The organisation was co-founded by philosopher and mathematician Bertrand Russell and Joseph Rotblat, a Polish-British physicist who had worked on the Manhattan Project and quit on moral grounds once it became clear Germany would not develop its own nuclear weapon (which meant the project was no longer a defensive endeavour).

Rotblat’s conviction was that those who build dangerous technologies carry a responsibility to help govern them. The Pugwash conferences he helped establish did not generate headlines, but they helped lay the groundwork for the Partial Test Ban Treaty, the Nuclear Non-Proliferation Treaty and the Chemical Weapons Convention.

Rotblat and the Pugwash Conference were jointly awarded the Nobel Peace Prize in 1995 – not for the documents they produced, but for the understanding they built before those documents were possible.

The analogy to AI governance is imperfect but instructive. Formal multilateral processes are needed but too slow and tortuous for the pace of AI development. Bilateral agreements are often too narrow, imbalanced or zero-sum game.

Operationalising cooperation

What is needed are trusted spaces – Track 1.5 forums where researchers, civil society, policy-makers, industry practitioners and investors can have the substantive and curated conversation that the formal summit stage does not allow. This is where the people who will eventually be in the rooms when hard decisions get made can build the shared understanding that makes those decisions possible.

Singapore understands this logic well. Josephine Teo, Singapore’s minister for digital development and information, offered a clear message at New Delhi: AI governance needs to be built on science and evidence, and the gap between AI’s progress and policy needs to close.

Singapore co-chairs the Science Working Group that has helped establish a new virtual network of AI for science institutions, connecting governments, industry and researchers worldwide. The country has committed S$120 million to fund AI for science projects.

As a small country that has consistently punched above its weight in global governance – technically credible, trusted across geopolitical lines, willing to do the unglamorous institutional work – Singapore is precisely the kind of actor that makes the harder coordination possible.

This matters in a bilateral context too. On Mar 3, Singapore’s President Tharman Shanmugaratnam met with President Macron at the Elysee Palace, reaffirming the countries’ partnership across political, economic, security and innovation fields.

France and Singapore are both middle powers that have chosen the path of independence rather than absorption into either of the two dominant AI ecosystems. That convergence is the foundation on which practical coordination gets built.

Macron closed his New Delhi address with a phrase that deserves to travel: “The old world said you compete or you lose. The new world says you connect or you fall behind.”

The India AI Impact Summit demonstrated that the will to connect exists, at scale, across the Global South and beyond. The New Delhi Commitments show that connecting can produce concrete outcomes. What comes next is less visible and harder to applaud: building the incident reporting channels, the verification mechanisms, the evaluation standards and the trusted forums that make coordination operational rather than aspirational.

The summit stage is where political will gets declared and assessed. Governance gets built in the harder, less visible work that follows. That work needs to start now.

The writer is co-founder of AI Safety Connect, a Track 1.5 diplomacy platform convening governments, frontier AI labs and civil society on AI safety and governance. A pioneer of global AI policy and governance, he is an appointed expert to the Global Partnership on AI, Organisation for Economic Co-operation and Development, Pacific Economic Cooperation Council, the UN and Unecso.

This essay is part of New Global Order, a series which explores how the changing world landscape is reshaping business, politics and beyond.