Over the past weekend, the world's three leading AI companies unexpectedly signaled a slower pace of development, and OpenAI CEO Sam Altman once again called for the industry to jointly establish safety standards.
On September 14, Altman posted on platform X, outlining two worst-case scenarios for AI development: first, AI ultimately spiraling out of control, causing humanity to lose its grip on the future; second, immensely powerful AI falling under the control of a select few individuals, companies, or nations, who might impose their will upon everyone else. He stated that neither red line can be crossed.
Altman noted that OpenAI has begun proactively developing explicit safety cases before initiating reinforcement learning training for significant capability enhancements, and he hopes other frontier AI companies will adopt similar practices, converging on unified standards concerning model misalignment, monitoring, and security.
According to The Information, citing sources, Anthropic, OpenAI, and Google have been in discussions about establishing a joint AI industry standards body, with working group meetings still ongoing. This indicates that the leading AI labs are shifting their safety governance approach from setting internal rules to promoting a coordinated industry framework.
Bringing Safety Governance Forward: Slowing the Pace is Not Stalling Development
Altman emphasized that as model capabilities continue to improve, alignment and safety technologies must develop at a simultaneous or even faster pace to prevent AI capabilities from advancing ahead of the safeguards. The AI sector needs to walk a "narrow middle path": neither sacrificing safety due to competitive pressures nor allowing capability improvements to significantly outpace alignment, monitoring, and governance abilities.
He believes the core of AI safety governance is shifting focus from post-deployment risk control to the development process itself. OpenAI's current approach involves establishing safety cases proactively before initiating frontier reinforcement learning training expected to significantly boost model capabilities, instead of waiting for models to be completed and then conducting safety evaluations.
He also stressed that "pacing" is not about halting AI progress but about avoiding the fastest technologically possible advancement. Although safety cases and monitoring measures will increase research and development costs, these investments remain necessary to ensure that safety and governance capabilities can keep pace with enhancements in model capabilities.
Regulations Pending, AI Industry Should Take the Lead on Safety Standards
On the regulatory front, Altman supports the US government establishing uniform safety requirements for frontier AI, yet he believes the industry doesn't need to wait for relevant legislation to pass before taking action. Even without government involvement, major AI labs should take the lead in establishing industry standards bodies and exploring mechanisms like independent audits.
According to reports, Anthropic CEO Dario Amodei called on AI companies to strengthen coordination on technical testing and review last Saturday. Earlier, Anthropic, OpenAI, and Google had already been engaging in talks to set up a joint industry standards organization.
During an internal staff meeting, Altman said that even if the US government is not currently involved, major AI labs should nevertheless push forward with standards development on their own; as for global coordination, government participation is necessary, but let's do our best independently first.
In his view, as AI capabilities continue to rise, societal demands for safety from frontier AI companies will keep growing, and industry self-regulation must ultimately be realized through the concrete actions of every individual AI lab.