UK AI Safety Institute Expands to US

The United Kingdom’s Artificial Intelligence (AI) Safety Institute is preparing to expand its operations abroad by establishing a new office in the United States. On May 20th, Michelle Donelan, the U.K. Technology Secretary, revealed plans to open this first international branch in San Francisco during the upcoming summer. The decision to set up in San Francisco is strategic, as it allows the U.K. to access the extensive pool of technological talent present in the Bay Area and engage closely with one of the largest AI hubs between London and San Francisco. This move aims to solidify relationships with key American players and promote global AI safety for public benefit.

Currently, the AI Safety Institute’s London office comprises a team of 30 professionals who are on track to expand their expertise, particularly in evaluating risks associated with advanced AI models. Michelle Donelan emphasized that this international expansion illustrates the U.K.’s commitment and leadership in advancing AI safety. She described this as a defining moment for the U.K., enabling the country to assess AI risks and opportunities from a global perspective, enhancing its partnership with the U.S., and setting an example for other nations to follow in AI safety.

In November 2023, the U.K. organized a significant AI Safety Summit in London, marking a milestone in addressing AI safety on a global level. The summit featured notable leaders from various countries, including representatives from the U.S. and China, as well as prominent figures in the AI industry like Brad Smith from Microsoft, Sam Altman from OpenAI, Demis Hassabiss from Google and DeepMind, and Elon Musk. This event underscored the international focus on AI and set the stage for future collaborations and discussions on AI safety.

In conjunction with the recent announcement, the U.K. disclosed some of the AI Safety Institute’s findings from safety evaluations conducted on five advanced AI models that are publicly available. These models were anonymized to provide an unbiased “snapshot” of their capabilities, rather than labeling them as outright “safe” or “unsafe.” The findings revealed a range of abilities among the models; for example, some could successfully tackle cybersecurity tasks, though more complicated challenges posed difficulties for others.

The assessment indicated that several AI models possessed knowledge equivalent to PhD-level expertise in fields like chemistry and biology. All the models were found to be “highly vulnerable” to basic jailbreak attempts, and none could independently handle more complex, time-consuming tasks without human intervention. Ian Hogearth, the chair of the AI Safety Institute, noted that these evaluations would contribute to a more empirical understanding of AI model capabilities.

Hogearth added that AI safety remains a nascent and evolving field. The results presented so far represent only a fraction of the comprehensive evaluation framework the AI Safety Institute is developing. This ongoing research and assessment effort aims to provide a deeper understanding of the strengths and limitations of various AI models.

As the institute continues to grow and refine its methods, it aims to lead the way in developing robust safety standards and practices. The international expansion signifies a commitment to fostering global cooperation and knowledge sharing in the pursuit of safer AI technology.

The establishment of the new office in San Francisco is expected to enhance collaboration between the U.K. and the U.S., leveraging the strengths of both countries to advance the shared goal of ensuring AI technologies are developed and used safely and responsibly.