AI progress must not outrun safety: Sam Altman

OpenAI CEO Sam Altman has urged the AI industry to establish shared safety standards for frontier AI systems, emphasizing the need for responsible development. In a post on X, Altman welcomed a federal framework with uniform safety requirements, including independent auditors. However, he urged companies not to wait for legislation or antitrust exemptions to begin working on safety measures.

Altman stated that every frontier lab must ensure responsible development, warning that there are two ways AI progress could go badly. First, the industry could lose control of the future to AI, which he deemed unacceptable and unacceptable for AI to serve humanity. To prevent this, he advocated for alignment and safety techniques to keep pace with model capabilities.

The second risk, according to Altman, is excessive concentration of power. He warned that an extraordinarily powerful AI in the hands of one person or company could result in a dystopian outcome. Altman emphasized the need to walk a narrow middle path, preventing one country from gaining too much power and avoiding a situation where one AI lab becomes disproportionately influential.

OpenAI has focused on post-deployment safety in its Responsible Scaling Policies and Preparedness Frameworks. Now, the company is also considering risks during the development process, including formulating explicit safety cases before frontier reinforcement learning runs that significantly increase model capabilities.

Altman called for other AI companies to share their approaches to develop common standards around misalignment, monitoring, and safety. He clarified that "pacing" AI development does not mean stopping progress altogether but rather slowing it down to allow for additional safety cases and monitoring work. Altman stressed that governments would be needed for international coordination, but the industry should initiate steps on its own to avoid recklessness and ensure alignment and monitoring remain ahead of capabilities.