Paris: the summit that dropped 'safety' from its name

What happened at the Grand Palais

The third international AI summit ran on February 10 and 11 in Paris, co-chaired by Emmanuel Macron and Narendra Modi, with over 1,000 participants from more than 100 countries. The 2023 Bletchley Park meeting had 29 governments and few company executives. Paris had heads of state, most of the major labs, and a trade-show atmosphere that Fortune described as an AI festival next to Bletchley's decidedly sober tone. The name had changed too. Bletchley was the AI Safety Summit. Seoul in 2024 was the AI Seoul Summit. This one was the AI Action Summit.

The closing document was a Statement on Inclusive and Sustainable Artificial Intelligence. Fifty-eight countries signed, including France, China, India, Australia, Brazil, Japan and South Korea. The United States and the United Kingdom did not. That is the fact most people will remember from the week, and it deserves to be, because the previous two summits had been organised largely by the UK and had the US as a central participant. A declaration that China signs and the United States does not is a very different document from the one signed at Bletchley, whatever its text says.

It is worth being clear about what the statement does and does not contain. It commits signatories to accessible, open, ethical and sustainable AI, to narrowing digital divides and to international coordination on governance. It does not commit anyone to pre-deployment testing of frontier models, which was the concrete mechanism the earlier summits had been working toward.

What Vance said

J.D. Vance's speech was the sharpest statement of the new American position. He warned that excessive regulation of AI could kill a transformative sector just as it is taking off, and he cautioned other governments against cooperating with authoritarian regimes on AI, a line widely read as being about China. For a summit series that began with a joint UK, US and China declaration about frontier risk, that is a reversal of both the substance and the coalition.

The UK's refusal was quieter and, to us, more surprising. The country that convened Bletchley and founded the first government AI Safety Institute declined a declaration that mentioned safety far less than its own earlier texts. We do not have a clear public account of why, and we are not going to guess at one here.

Where the money went

The announcements were dominated by capital. The European Commission announced a 200 billion euro InvestAI initiative including 20 billion for AI gigafactories. More than 60 European companies pledged 150 billion through an EU AI Champions Initiative. France announced 110 billion in AI investment for its own sector, including a 30 to 50 billion euro data centre commitment from the United Arab Emirates and 20 billion from Brookfield. Macron also launched Current AI, a foundation for public-interest datasets and open tools, with a 400 million dollar initial endowment and backing from nine governments including India, Kenya, Nigeria and Germany.

Set those numbers beside the Bletchley outputs, which were a declaration and a commitment to pre-deployment testing, and the shift in what a summit is for becomes obvious. The Financial Times read the week as a turn toward geopolitical competition and an AI arms race. Dario Amodei called it a missed opportunity. Both readings can be true at once.

Why the arc matters more than any paper

In sixteen months the organising question moved from how to manage catastrophic risk from frontier models to how to win. The topics that gained ground in Paris, environmental cost, labour, access for the global south, are real, and a summit series that only talked about extinction risk would have been narrower than the field. What happened in Paris reads as a substitution of one agenda for another rather than an expansion, and the countries most able to shape frontier development chose the moment to walk away from the shared text.

For a research foundation this has a practical consequence. The idea that governments would coordinate pre-deployment evaluation of frontier models, the thing Bletchley and Seoul pointed toward, now has no obvious home. Whatever evaluation infrastructure exists in a year will have been built by labs, by the remaining safety institutes, and by independent groups, and the incentives on each of those are different from the ones a summit process was meant to create. We would rather know that than keep planning around a coalition that no longer exists.