The UK is leading an international effort to test the safety risks of cutting-edge AI models before they are released to the public, as regulators rush to build a workable safety regime ahead of the Paris Summit in six months’ time. .
The UK’s AI Safety Institute is the first of its kind and is currently being matched with counterparts around the world, including South Korea, the US, Singapore, Japan and France.
Regulators at the Seoul AI Summit hope that their agencies can work together to craft a 21st century version of the Montreal Protocol, the landmark agreement to regulate CFCs and close the hole in the ozone layer.
But before doing so, the institutes need to agree on how they can work together to transform a patchwork of international approaches and regulations into a unified effort to govern AI research. be.
“At Bletchley, we announced the UK Institute for AI Safety, the world’s first government-backed organization to tackle advanced AI safety in the public interest,” said Michelle, UK Technology Secretary. Donnellan said in Seoul on Wednesday. She credited the “Bletchley effect” with encouraging the creation of a global network of peers working on the same thing.
The agencies will begin sharing information about models, their limitations, capabilities, and risks, as well as monitor where specific “AI harm and safety incidents” occur and share resources to advance global understanding of the science of AI safety.
Speaking at the first “full house” meeting of these countries on Wednesday, Mr Donnellan warned that building the network was just the first step. “We cannot be complacent. As the pace of AI development accelerates, we must match our own efforts to understand the risks and seize the limitless opportunities for our people. ”
The safety agency network has a tight deadline: This fall, leaders are due to meet again, this time in Paris, for the first full AI summit since Bletchley. If the conversation moves from how to test AI models to how to regulate them, safety agencies will need to demonstrate they have mastered what Donalan calls the “early science of frontier AI testing and evaluation.”
Jack Clarke, co-founder and policy director at AI lab Anthropic, said that simply setting up a functional safety lab would put the UK “100 miles” further down the road to safe AI than it was two years ago.
“What we need to do now is encourage governments to continue to invest the money needed to establish safety laboratories and to staff them with enough engineers, as I have done here. “I think it’s about allowing the government to actually produce its own information and evidence,” he said.
As part of that investment in science, Donnellan announced £8.5 million in funding to “break new ground” in AI safety testing.
Francine Bennett, interim director of the Ada Lovelace Institute, said the funding is a good start “but it needs to pave the way for more substantive programs to understand and prevent societal and systemic risks.”
“It’s great to see the Safety Institute and the Government taking a step towards a broader view of what safety is, both with the State of the Science report and with this funding. We recognise that safety is not something that can be satisfactorily tested in a lab,” Bennett added.
The summit was criticized for leaving important voices out of the conversation. South Korean civil society organizations were not included, the host country was represented only through academia, government and industry, and only the largest AI companies were invited to participate. Roland Decorte, president of the AI Founders Association, warned that the debate “risks becoming solely focused on flashy, large-scale models.” As a result, “only a handful of them came to dominate and now only the largest companies can be created, which are suffering economic losses.”
“The question is, ultimately, do we want to regulate and build towards a mature AI economy of the future that creates a sustainable framework for the vast majority of companies operating in this space?” he added.
