The United Kingdom is putting part of its AI safety operation closer to the companies shaping the technology. The AI Safety Institute, a U.K. body launched in November 2023 to assess and address risks in AI platforms, plans to open a second location in San Francisco.
The decision comes ahead of the AI safety summit in Seoul, South Korea, which the United Kingdom is co-hosting. It also shows how much of the practical work around AI risk now depends on proximity to the labs building foundational AI technology.
Why San Francisco matters
The Bay Area is home to companies including OpenAI, Anthropic, Google and Meta. Those firms are building foundational models, the systems that sit underneath generative AI services and other applications.
For the U.K., having staff in San Francisco is not just symbolic. Michelle Donelan, the U.K. secretary of state for science, innovation and technology, told TechCrunch that a local base would give the institute access to the headquarters of many AI companies. Some of those companies already have bases in the United Kingdom, but the U.K. still sees value in being present where much of the core development work is happening.
That presence may also help the institute draw from an additional pool of talent and work more closely with the United States. The U.K. has already signed an MOU with the U.S. to collaborate on AI safety initiatives, yet it is also choosing to establish its own office in the U.S. technology center.
The move reflects a practical reality: AI safety work is not only about writing policy from afar. It also depends on understanding what companies are building, how quickly models are changing, and whether evaluators can engage with developers before systems are released.
A small institute facing a large industry
The AI Safety Institute is still modest in size. The organization has 32 employees, while the companies developing AI models are backed by billions of dollars of investment and strong incentives to bring their technologies to paying users.
That difference in scale matters. The institute is trying to evaluate risks in powerful AI systems, but it is operating in an environment where private companies move quickly and where commercial pressure can shape the pace of release.
One of the institute's most notable steps so far has been the release of Inspect, its first set of tools for testing the safety of foundational AI models. Donelan described that release as a phase one effort.
Inspect gives the U.K. a concrete system to present to others, including regulators. Donelan said one goal of the Seoul conference is to present Inspect to regulators and encourage them to adopt it too.
The hard part is getting models evaluated
The existence of testing tools does not solve the entire AI safety problem. The source article notes that benchmarking models has proven challenging, and that engagement with testing is currently opt-in and inconsistent.
That creates a central weakness in the current setup. Companies are under no legal obligation to have their models vetted, according to a senior source at a U.K. regulator cited by TechCrunch. Not every company is willing to have models checked before release.
In practice, that means risk evaluation may happen too late in some cases. If a problem is only identified after a system is already available, the ability to prevent harm is reduced.
Donelan said the institute is still developing ways to work with AI companies on evaluation. She also described model evaluation as an emerging science, meaning the process itself is still being built and refined.
That point is important for the broader AI regulation debate. Safety testing is often discussed as if it were a settled technical process, but the U.K. position described here is more cautious: the tools, methods, and institutional relationships are still forming.
Policy is coming, but not first
Longer term, Donelan said she believes the U.K. will build more AI legislation. But she repeated the position associated with Prime Minister Rishi Sunak: the U.K. does not want to legislate before it better understands the scope of AI risks.
That approach puts research and evaluation ahead of immediate lawmaking. The institute's recent international AI safety report focused mainly on building a fuller picture of research to date. Donelan said that report showed there are major gaps and that more global research needs to be encouraged.
The argument is also practical. Donelan said legislation takes about a year in the United Kingdom. In her view, if the government had started legislating instead of organizing the AI Safety Summit held in November last year, it would still be in the legislative process and would have less to show.
That does not mean the U.K. is avoiding regulation permanently. It means its current strategy is to build capacity, testing methods, and international cooperation before turning those lessons into binding rules.
What the new office signals
The San Francisco office signals that AI safety is becoming more operational. The U.K. wants to be near the companies building frontier AI systems, near relevant technical talent, and in a better position to collaborate with the United States.
It also signals that international AI safety work is no longer confined to summits and reports. Those still matter, but the institute now needs day-to-day contact with the organizations developing the models it wants to evaluate.
Ian Hogarth, chair of the AI Safety Institute, framed the move as part of an international approach to AI safety, including shared research, collaboration with other countries, model testing, and anticipating risks from frontier AI.
The larger question is whether a small public institute can keep pace with a fast-moving industry. The San Francisco office does not answer that question by itself, but it gives the U.K. a closer seat at the table where many of the most important AI systems are being built.