The government’s groundbreaking Institute for AI Safety opens in San Francisco

  • Britain AI Safety Institute aims to expand across the Atlantic to broaden its technical expertise and strengthen its position as a global authority in this field AI Safety.
  • expansion revealed as AI Safety Institute is the first to publish ever AI safety test results on publicly available models and agrees a new partnership with Canada.
  • comes before the co-host AI Seoul summit, where Britain is demonstrating AI The Safety Institute’s continued global leadership AI safety.

The pioneering work of the British government AI Safety Institute will broaden its international horizons by opening its first overseas office in San Francisco this summer, Technology Secretary Michelle Donelan announced today (Monday, May 20).

The expansion marks a crucial step that will enable Britain to tap into the wealth of technology talent available in the Bay Area, connecting with the world’s largest AI laboratories headquartered in both London and San Francisco, and strengthening relationships with the United States to move forward AI safety for the common good.

The office is expected to open this summer and the first team of technical staff will be hired under the leadership of a research director.

It will be a complementary branch of the London Institute Headquarters, which is going from strength to strength and already has a team of more than 30 technical employees. The London office will continue to scale up and acquire the necessary expertise to assess frontier risks AI systems.

By expanding its position in the US, the Institute will establish close cooperation with the US, enhancing the country’s strategic partnership and approach to AI safety, while also sharing research and conducting joint assessments AI models that can inform AI security policies around the world.

Secretary of State for Science and Technology Michelle Donelan said:

This expansion represents British leadership AI in action. It is a pivotal moment in Britain’s ability to study both the risks and potential of this disease AI from a global lens, strengthening our partnership with the US and paving the way for other countries to tap into our expertise as we continue to lead the world in AI safety.

Since the Prime Minister and I the AI Safety Institute, it has grown from strength to strength and in just over a year we have built the world’s leading government here in London AI research team, attracting top talent from the UK and beyond.

Opening our doors abroad and building on our alliance with the US is central to my plan to set new international standards AI security that we will discuss at the Seoul Summit this week.

The expansion comes as Britain AI Safety Institute publishes a selection of recent safety test results from five publicly available advanced tests AI models: the first government-backed organization in the world to reveal the results of their evaluations.

Although only a small part of the Institute’s broader approach, the results demonstrate the significant progress the Institute has made since the November meeting. AI Safety Summit as it builds its state-of-the-art safety testing capabilities.

The Institute has assessed AI models against four key risk areas, including how effective the protections developers have installed actually are in practice. As part of the findings, the Institute’s testing found that:

  • Several models completed cybersecurity challenges while struggling to complete more advanced challenges.
  • Several models show something similar Ph.D-knowledge of chemistry and biology.
  • All tested models remain highly vulnerable to basic jailbreaks, and some will produce malicious results even without dedicated security bypass efforts.
  • Models tested were unable to perform more complex, time-consuming tasks without human supervision.

AI Safety Institute Chairman Ian Hogarth said:

The results of these tests mark the first time we can share some details of our model evaluation work with the public. Our evaluations will help contribute to an empirical assessment of the capabilities of models and the lack of robustness when it comes to existing safeguards.

AI Safety is still a very young and emerging field. These results represent only a small part of the evaluation approach AISI develops. Our ambition is to continue to push the boundaries in this area by developing state-of-the-art assessments, with an emphasis on risks related to national security.

AI Safety remains a top priority for Britain as the country continues to drive the global conversation around the safe development of technology.

This effort was launched in November AI Security Summit at Bletchley Park, and momentum continues to grow as Britain and the Republic of Korea prepare to co-host the conference AI Seoul summit this week.

As the world prepares to gather in Seoul this week, Britain has committed to working with Canada, including through their respective countries AI Security institutions, to realize their ambition to create a growing network of state-backed organizations focused on this AI safety and governance. Confirmed by British Minister of Technology Michelle Donelan and Canadian Minister of Science and Innovation François-Philippe Champagne, this partnership will serve to deepen existing ties between the two countries and inspire cooperation in systemic security research.

As part of this agreement, the countries will seek to share their expertise to strengthen existing testing and evaluation work. The partnership will also enable secondment routes between the two countries and work together to jointly identify areas for research collaboration.

Notes for editors

The Institute’s security tests were conducted this year on five publicly available major language models (LLMs) that are trained on large amounts of data. The tested models are anonymized.

The results provide only a snapshot of the model’s capabilities and do not designate systems as ‘safe’ or ‘unsafe’. The tests performed represent a small part of the evaluation techniques AISI developed and used, as set out in the Institute’s evaluation approach published earlier this year.

Today’s publication can be found on the AI Website of the Safety Institute.

Today also marks the latest progress update from Institute President Ian Hogarth, which can be found here on the website AI Website of the Safety Institute.

Leave a Reply

Your email address will not be published. Required fields are marked *