- The UK AI Safety Association plans to expand across the Atlantic to expand its technical expertise and establish itself as the global authority on AI safety.
- The expansion was announced after the AI Safety Institute published its first AI safety test results on a publicly available model and agreed to a new collaboration with Canada.
- will be held ahead of the co-hosted AI Seoul Summit, demonstrating the UK AI Safety Association's continued leadership in global AI safety.
The UK government's pioneering AI Safety Institute will expand its international horizons by opening its first overseas office in San Francisco this summer, Technology Secretary Michelle Donnellan said today (Monday 20 May). ) Announced.
The expansion will see the UK tap into the wealth of technology talent available in the Bay Area, work with the world's largest AI research institute headquartered in both London and San Francisco, and build on its relationship with the US to advance AI. This is an extremely important step that will enable us to strengthen our capabilities. Safety in the public interest.
The office is scheduled to open this summer and is recruiting an initial team of technical staff, led by a principal investigator.
The branch will complement the Institute's London headquarters, which continues to grow and already has a team of more than 30 people. The London office will continue to grow in size and gain the expertise needed to assess risks in frontier AI systems.
By expanding its footprint in the U.S., the Institute will establish a closer working relationship with the U.S. to advance the U.S. strategic partnership and approach to AI safety, while sharing research and promoting global Conduct joint evaluations of AI models that can inform AI safety policy. .
Michelle Donnellan, Secretary of State for Science and Technology, said:
This expansion really represents the UK’s leadership in AI. This is a pivotal moment in which the UK can explore both the risks and opportunities of AI from a global perspective, strengthening our partnership with the US as the UK continues to lead the world in AI, and encouraging other countries to This will pave the way for you to utilize your specialized knowledge. safety.
Since the Prime Minister and I established the Institute for AI Safety, it has grown from strength to strength, and in just over a year we have built a world-leading government AI research team here in London, and are leading the way in the UK and abroad. We have gathered excellent human resources from
Opening doors overseas and building an alliance with the United States are central to my plan to set new international standards for AI safety, which I will discuss at the Seoul Summit this week. ”
The expansion comes as the UK AI Safety Association publishes some of the recent results of its safety tests of five publicly available advanced AI models, making it the first in the world to publish assessment results. It is a government supported agency.
Although the results are just one part of the institute's broader approach, the institute has made significant progress in building cutting-edge safety testing capabilities since the AI Safety Summit in November. It is shown that.
The institute evaluated AI models against four key risk areas, including how effective the safety measures put in place by developers actually are. As part of the findings, laboratory tests found that:
- While some models completed the cyber security challenges, they struggled to complete more advanced challenges.
- Some models demonstrate knowledge similar to doctoral level knowledge in chemistry and biology.
- All models tested remain highly vulnerable to basic “jailbreaks”, and even without dedicated attempts to circumvent safeguards, some models can produce harmful outputs. will generate.
- The tested models were unable to complete more complex and time-consuming tasks without human supervision.
Ian Hogarth, Chairman of the AI Safety Institute, said:
The results of these tests allowed us to share details of our model evaluation work with the public for the first time. Our evaluation helps contribute to an empirical assessment of the lack of robustness with respect to model functionality and existing safety measures.
AI safety is still a very new and emerging field. These results are just a small part of the evaluation approach that AISI is developing. Our goal is to continue to push the frontiers of the field by developing cutting-edge assessments with a focus on national security-related risks.
AI safety remains a key priority for the UK as we continue to drive the global conversation around the safe development of technology.
The initiative began with the AI Safety Summit held at Bletchley Park in November, and continues to gain momentum as the UK and South Korea prepare to co-host the AI Seoul Summit this week.
As the world prepares to gather in Seoul this week, the UK will work with Canada, including through their respective AI Safety Institutes, to advance its ambitions to expand the network of government-backed organizations focused on AI safety and governance. I promised to do it. The partnership, confirmed by UK Technology Minister Michelle Donnellan and Canada's Science and Technology Minister François-Philippe Champagne, will help deepen existing links between the two countries and foster systematic safety research collaboration. right.
As part of the agreement, the countries will aim to share expertise to strengthen existing testing and evaluation efforts. The partnership will also enable secondment routes between the two countries and will work to jointly identify areas for research collaboration.
Note for editors:
- The institute's safety tests were conducted this year against five publicly available large-scale language models (LLMs) trained on large amounts of data. The tested model is anonymized.
- The results only provide a snapshot of the model's functionality and do not designate the system as “secure” or “unsecure.” The tests conducted represent a small portion of the evaluation techniques developed and used by AISI, as outlined in the Institute's Approach to Evaluation published earlier this year.
- For more information about today's publication, please visit the AI Safety Institute website.
- Today we will also have the latest progress update from Ian Hogarth, Chair of the AI Safety Institute. This latest progress can be viewed on his website at the AI Safety Institute.
