OpenAI guidance allows boards to restrict CEOs from releasing new models to guard against AI risks

In order to avoid the huge risks that artificial intelligence (AI) may bring, OpenAI decided to give the board of directors greater power to oversee security matters and conduct strict supervision on CEO Sam Altman, who just won an internal battle last month

OpenAI released a series of guidelines on Monday, December 18, Eastern Time, designed to track, assess, predict, and prevent catastrophic risks posed by increasingly powerful artificial intelligence (AI) models. OpenAI defines "catastrophic risk" as any risk that could result in hundreds of billions of dollars in economic losses, or serious injury or death to multiple people

The 27-page guidance, known as the "Readiness Framework," states that even if a company's top managers, including the CEO or a person designated by leadership, believe that an AI model to be released is safe, the company's board of directors still Option to postpone the release of the model. This means that while OpenAI’s CEO is responsible for day-to-day decisions, the board of directors will be informed of the discovery of risks and have the power to veto the CEO’s decisions

OpenAI’s Readiness Framework recommends using a matrix approach to document the level of risk posed by cutting-edge AI models across multiple categories, in addition to provisions for company leadership and board authority. These risks include bad actors using AI models to create malware, launch social engineering attacks, or spread harmful nuclear or biological weapons information

Specifically, OpenAI sets risk thresholds in four categories: cybersecurity, CBRN (chemical, biological, radiological, nuclear threats), persuasion, and model autonomy. Before and after risk mitigation measures are implemented, OpenAI classifies each risk into four levels: low, medium, high, or severe

OpenAI stipulates that only AI models rated "medium" or below after risk mitigation can be deployed, and only models rated "high" or below after risk mitigation can continue to be developed. If the risk cannot be reduced, Downgrades below critical and the company will stop developing the model. OpenAI will also take additional security measures for models assessed as high risk or severe risk until the risk is mitigated

OpenAI divides security issue handlers into three teams. The Security Systems team is focused on mitigating and addressing risks posed by current products such as GPT-4. The Super Alignment team is concerned about the problems that may arise when future systems exceed human capabilities. Additionally, there is a new team called Prepare, led by Aleksander Madry, professor in the Department of Electrical Engineering and Computer Science (EECS) at the Massachusetts Institute of Technology (MIT)

The new team will evaluate robust model development and implementation. They will be specifically responsible for overseeing the technical work and operational architecture related to security decisions. They will drive technical work, review limitations of cutting-edge model capabilities, and conduct assessments and synthesis related reports

Madry said that his team will regularly evaluate the risk level of OpenAI’s most advanced artificial intelligence models that have not yet been released, and submit monthly reports to OpenAI’s internal Security Advisory Group (SAG). SAG will analyze Madry’s team’s work and provide recommendations to CEO Altman and the company’s Board of Directors

According to guidance documents released on Monday, Altman and his leadership can use these reports to decide whether to release new AI systems, but the board of directors retains the power to overturn their decisions

Currently, Madry’s team only has four people, but he is working hard to recruit more members. It is expected that the team will reach 15 to 20 people, which is similar to the existing security team and hyper-alignment team.

Madry hopes other AI companies will assess the risks of their models in a similar way and believes this could become a model for regulation

The above is the detailed content of OpenAI guidance allows boards to restrict CEOs from releasing new models to guard against AI risks. For more information, please follow other related articles on the PHP Chinese website!