The lab leaders want to slow down on AI. Someone needs to help them

AT DAVOS on Jan 22, Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis made a striking admission. Asked whether they would welcome slowing down the development of artificial intelligence (AI), both said yes.

“Maybe it would be good to have a slightly slower pace,” Hassabis ventured. “So that we can get this right societally.”

Amodei agreed: “I would prefer that. I think that would be better for the world.”

Then came the catch. Such restraint would require “international collaboration”, Hassabis said.

Amodei was blunter. “It’s very hard to have an enforceable agreement where they slow down and we slow down.”

The message was one of helplessness. Two of the world’s most powerful technology executives, freely admitting they are trapped in a race neither chose and neither can escape alone.

This should concern us. The people building what may become the most consequential technology in human history are publicly asking for help to slow down – and no one is answering.

The US, consumed by a great-power competition with China, is in no mood to discuss restraint. Beijing, on its part, is not about to take direction from Washington. The superpowers are deadlocked. If coordination is to happen, it will not come from them.

Middle powers

Enter Singapore and other middle powers. While they may not be home to the leading frontier AI labs, they have something the superpowers do not: the ability to convene without being seen as advancing hegemonic interests, and the collective market leverage to make standards stick.

Singapore’s President Tharman Shanmugaratnam, also speaking at the World Economic Forum, articulated precisely this vision: “We now have to build up alliances of countries – small, medium-sized and large powers as well – around areas of common interest... We do have a common interest in AI safety.” He called for “flexible multilateralism”, building plurilateral alliances that can act where traditional institutions cannot, while not repudiating the multilateral system itself.

Singapore, he noted, will face the AI challenge “sooner than many countries” because of its openness and reliance on technology. This exposure is not a vulnerability, but a strength, he argued.

Singapore’s AI governance framework is already demonstrating this approach in practice, shaping how firms deploy systems in South-east Asia.

When middle powers set standards, global labs follow – not out of altruism, but because they need market access.

The sceptics will note that previous coordination attempts have failed. In 2023, an open letter calling for a six-month pause on advanced AI training gathered thousands of signatures, but was ignored in weeks.

But that effort asked for unilateral restraint without coordination infrastructure. The moment competitive pressure resumed, the pause collapsed. Moratoriums without mechanisms do not work.

Middle powers can build the infrastructure that makes restraint stick. They could establish common safety tests, designed by independent scientists, that every lab must pass before a major release.

They could create a near-miss registry modelled on aviation practice, through which engineers worldwide report dangerous behaviour and prevent similar failures elsewhere.

They could also mandate independent audits to verify claims about guardrails, access controls and emergency shutdown procedures.

President Tharman has proposed something similar: a scientific coalition for AI safety, equivalent to the Intergovernmental Panel on Climate Change, to monitor emerging risks.

Amodei made this point, perhaps inadvertently, at the World Economic Forum. If chip exports to China could be restricted, he argued, “this isn’t a question of competition between the US and China. This is a question of competition between me and Demis, which I’m very confident we can work out”.

In other words: remove the geopolitical dimension, and the Western labs can coordinate. Middle powers cannot remove the US-China rivalry, but they can build coordination infrastructure that functions regardless of it.

The window for this is not open indefinitely. Both Amodei and Hassabis suggested that AI systems capable of recursive self-improvement may arrive within one to five years.

The infrastructure needs to be built now, while the lab leaders themselves are signalling openness, and before the technology makes restraint impossible.

Mark Carney, Canada’s prime minister, put it plainly at Davos: “We are in the midst of a rupture. The multilateral institutions on which the middle powers have relied are under threat.”

His prescription was the formation of “coalitions of the willing” – groups of like-minded countries acting together where the old institutions cannot.

AI governance is the test case for the workability of this approach. The lab leaders have shown their hand. They would welcome help. The question is whether middle powers will recognise the opportunity in front of them, and seize it before the window closes.

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. He will speak at The India AI Impact Summit, which runs from Feb 16 to 20.