The US and UK have signed a landmark agreement on artificial intelligence, as the allies become the first countries to formally cooperate on how to test and assess risks from emerging AI models.

The agreement, signed on Monday in Washington by UK Science Minister Michelle Donelan and US Commerce Secretary Gina Raimondo, lays out how the two governments will pool technical knowledge, information, and talent on AI safety.

The deal represents the first bilateral arrangement on AI safety in the world and comes as governments push for greater regulation of the existential risks from new technology, such as its use in damaging cyber attacks or designing bioweapons.

“The next year is when we’ve really got to act quickly because the next generation of [AI] models are coming out, which could be complete game-changers, and we don’t know the full capabilities that they will offer yet,” Donelan told the Financial Times.

The agreement will specifically enable the UK’s new AI Safety Institute (AISI), set up in November, and its US equivalent, which is yet to begin its work, to exchange expertise through secondments of researchers from both countries. The institutes will also work together on how to independently evaluate private AI models built by the likes of OpenAI and Google.

The partnership is modeled on one between the UK’s Government Communications Headquarters (GCHQ) and the US National Security Agency, who work together closely on matters related to intelligence and security.

“The fact that the United States, a great AI powerhouse, is signing this agreement with us, the United Kingdom, speaks volumes for how we are leading the way on AI safety,” Donelan said.

She added that since many of the most advanced AI companies were currently based in the US, the American government’s expertise was key to both understanding the risks of AI and to holding companies to their commitments.

Advertisement

However, Donelan insisted that despite conducting research on AI safety and ensuring guard rails were in place, the UK did not plan to regulate the technology more broadly in the near term as it was evolving too rapidly.

The position stands in contrast to other nations and regions. The EU has passed its AI Act, considered the toughest regime on the use of AI in the world. US President Joe Biden has issued an executive order targeting AI models that may threaten national security. China has issued guidelines seeking to ensure the technology does not challenge its long-standing censorship regime.

Raimondo said AI was “the defining technology of our generation.”

“This partnership is going to accelerate both of our institutes’ work across the full spectrum of risks, whether to our national security or to our broader society,” she said.

“Our partnership makes clear that we aren’t running away from these concerns—we’re running at them. Because of our collaboration, our institutes will gain a better understanding of AI systems, conduct more robust evaluations, and issue more rigorous guidance.”

The UK government-backed AISI, which is chaired by tech investor and entrepreneur Ian Hogarth, has hired researchers such as Google DeepMind’s Geoffrey Irving and Chris Summerfield from the University of Oxford to start testing existing and unreleased AI models.

OpenAI, Google DeepMind, Microsoft, and Meta are among the tech groups that signed voluntary commitments to open up their latest generative AI models for review by Britain’s AISI, which was established following the UK’s AI Safety Summit in Bletchley Park.

The institute is key to Prime Minister Rishi Sunak’s ambition for the UK to have a central role in tackling the development of AI.

Testing has focused on the risks associated with the misuse of the technology, including cyber security, by leaning on expertise from the National Cyber Security Centre within GCHQ, according to a person with direct knowledge of the matter.

Donelan said that she and Raimondo planned to discuss shared challenges, such as AI’s impact on upcoming elections this year. The science minister added that they would also discuss the need for computing infrastructure for AI, “sharing our skillsets and how we can deepen our collaboration in general to get benefit for the public.”

© 2024 The Financial Times Ltd. All rights reserved. Not to be redistributed, copied, or modified in any way.