The U.K. has formally launched an AI Safety Institute, turning a taskforce set up ahead of the Bletchley AI Safety Summit into a potential permanent fixture. Prime Minister Rishi Sunak says the institute will test emerging AI and help governments assess risks before new models are released. For now, the government is putting collaboration and evaluation ahead of regulation.
A testing role with unanswered questions
The Frontier AI taskforce was established in June to prepare for the summit. The newly confirmed institute is described as a global hub based in the U.K. and will sit within the Department of Science, Innovation and Technology.
Ian Hogarth, an investor, founder and engineer who chaired the taskforce, will lead the institute. Yoshuo Bengio will lead the production of its first report. The government has not said how much funding the institute will receive or whether AI companies will help pay for its work. Its description as “backed by leading AI companies” does not make clear whether that means financial backing or endorsement.
Those details matter because the institute’s remit is to test AI safety, while the companies developing models have so far been the ones testing them. An independent testing role could give governments another source of assessment, but the source does not specify how the institute will conduct tests or how it will be funded.
Working together before models are released
The launch came alongside the Bletchley Declaration, announced the previous day and signed by all countries attending the summit. It commits them to joined-up testing and other work to assess risks from frontier AI technologies, including large language models.
Sunak said that until now, only the companies building new AI models had tested their safety. He pointed to efforts by other countries, the UN and the G7, and said the next step would be to work together on testing before models are released.
The emphasis is on cooperation and assessment. The institute is meant to contribute to that work, while the declaration gives participating countries a shared commitment to testing. The article does not describe a common testing method or explain what would happen if an assessment identified a serious risk.
Why legislation is on hold
The U.K. has so far resisted moves to regulate AI at the platform level or for specific applications. Some critics argue that safety and risk assessment have little force without rules that can be enforced. Sunak’s position is that it is too early to legislate because the technology is changing quickly.
He argued that governments need to keep pace with AI and understand what they would be regulating before imposing requirements. That leaves an open question at the heart of the summit: how long assessment and voluntary cooperation can guide the sector before governments decide they know enough to act.
The institute’s creation is a concrete step toward testing, but it does not settle that question. Its funding, the nature of industry support and the route from test results to policy remain unclear in the announcement.
A summit with closed sessions
The second day of the summit also underscored a tension between calls for transparency and the way some discussions were held. Two plenary sessions were closed to journalists, apart from small pools watching attendees assemble.
Participants included leaders from DeepMind, OpenAI, Antrhopic, InflectionAI, Salesforce, Microsoft and AWS, as well as government representatives including Sunak, U.S. Vice President Kamala Harris, Giorgia Meloni of Italy and French finance minister Bruno Le Maire. China, a prominent guest on the first day, did not appear at the closed sessions.
Elon Musk, owner of X.ai (formerly known as Twitter), was also absent from the day’s sessions. Sunak was due to have a fireside chat with Musk that evening on Musk’s social platform; it was not expected to be broadcast live. The institute’s work now begins against that backdrop: countries have pledged to cooperate on testing, while the public still has limited detail on how that work will operate.