OpenAI Board Has Authority to Block CEO from Releasing AI Models
-
On December 19th, local time in the US, AI research company OpenAI stated that even if the company's leadership deems an AI model safe enough, its board of directors can choose to delay its release. This is another clear indication of the company empowering its board to strengthen safety measures in cutting-edge technology development.
A set of guidelines released on Monday detailed this decision-making process and explained how OpenAI plans to address the extreme risks that its most powerful AI systems might pose. The release of these guidelines followed a period of turmoil where OpenAI CEO Sam Altman was briefly ousted by the board, bringing the power balance between the board and company executives into sharp focus.
OpenAI's newly formed "Preparedness" team stated that it will continue to evaluate the company's AI systems to determine their performance across four categories, including potential cybersecurity issues and threats related to chemical, nuclear, and biological weapons, while working to mitigate any harm the technology might cause. Specifically, OpenAI is monitoring so-called "catastrophic" risks, defined in the guidelines as "any risk that could lead to hundreds of billions of dollars in economic damage or cause severe harm or death to many people."
Aleksander Madry, head of the "Preparedness" team and currently on leave from MIT, noted that the team will submit monthly reports to a new internal safety advisory group. This group will analyze Madry's team's work and provide recommendations to Altman and the company's board. Following Altman's brief dismissal, the board underwent a comprehensive overhaul. According to the document, Altman and his leadership team can decide whether to release new AI systems based on these reports, but the board retains the authority to override their decisions.
OpenAI announced the formation of the "Preparedness" team in October, making it one of three independent teams overseeing AI safety within the company. The other two teams are the "Safety Systems" team, which focuses on current products like GPT-4, and the "Superalignment" team, which concentrates on potentially extremely powerful AI systems in the future.
Madrid stated that his team will repeatedly evaluate OpenAI's most advanced, yet-to-be-released artificial intelligence models, rating them as 'low', 'medium', 'high', or 'critical' based on different types of perceived risks. The team will also make changes to reduce the potential dangers they identify in AI and measure their effectiveness. According to the new guidelines, OpenAI will only launch models with a danger rating of 'medium' or 'low'.
Madrid also hopes that other companies will use OpenAI's guidelines to assess the potential risks of their AI models. He said these guidelines formalize many of the processes OpenAI previously followed when evaluating released AI technologies. He and his team have developed many details over the past few months and received feedback from others within OpenAI.