Leading tech companies including Microsoft, Amazon and OpenAI came together on Tuesday for a groundbreaking international agreement on artificial intelligence safety at the Seoul AI Safety Summit.
Based on this agreement, companies from a range of countries, including the United States, China, Canada, the United Kingdom, France, South Korea and the United Arab Emirates, will undertake voluntary efforts to ensure the safe development of cutting-edge AI models.
If they haven't already done so, AI model makers have agreed to publish safety frameworks that show how they measure up to challenges in their state-of-the-art models, such as preventing bad actors from misusing the technology.
These frameworks include “red lines” for technology companies that define the types of risks associated with cutting-edge AI systems that are deemed “unacceptable.” These risks include, but are not limited to, automated cyber attacks and biological weapon threats.
To deal with such extreme situations, the companies said they plan to introduce a “kill switch” that will halt development of an AI model if they cannot guarantee risk mitigation.
“This is a world first that so many leading AI companies from around the world have agreed to the same commitments on AI safety,” British Chancellor Rishi Sunak said in a statement on Tuesday.
“These commitments ensure that the world's leading AI companies will be transparent and accountable about their plans to develop safe AI,” he added.
Tuesday's agreement expands on a series of commitments made last November by companies involved in developing generative AI software.
The companies agreed to gather input from “trusted stakeholders,” including their own governments as appropriate, before publishing these standards ahead of the next AI summit scheduled for early 2025 (the AI Action Summit in France).
The commitments agreed to on Tuesday only apply to so-called frontier models, a term that refers to the technology underpinning generative AI systems like OpenAI's GPT family of large-scale language models that power the popular ChatGPT AI chatbot.
Since ChatGPT was first introduced to the world in November 2022, regulators and tech leaders have become increasingly concerned about the risks surrounding advanced AI systems that can generate text and visual content as well as or better than humans.
The European Union (EU) is seeking to clamp down on unrestrained AI development with the enactment of an AI bill approved by the EU Council on Tuesday.
However, the UK has not proposed any formal legislation on AI, instead opting for a “light” approach to regulating AI, in which regulators would apply existing laws to AI technology.
The government recently said it would consider legislating the Frontier Model in the future, but has not said when formal legislation would be introduced.
