OpenAI Urges US-Led Global Standards for Frontier AI Development, Safety

OpenAI has urged the US to lead the creation of international standards for frontier AI development and safety to address the technology’s potential risks.

Bloomberg reported on September 21 that OpenAI, in a blog post that day, called for international technical standards to ensure AI is developed safely. It proposed setting criteria for which incidents involving AI models must be reported, how countries should cooperate as AI advances, and how to improve access to the computing resources needed for AI development.

OpenAI wrote that common standards are needed to guide development across labs and countries, alongside progress in alignment research, to ensure continued advances in AI remain safe and beneficial.

Chief Executive Officer Sam Altman is scheduled to speak at the United Nations Security Council on September 23 on AI and international security. He is also set to attend a White House state dinner on September 24 for Chinese President Xi Jinping.

The US and China have also begun discussing AI. Treasury Secretary Scott Bessent said over the weekend that the two countries agreed to establish a “US-China AI dialogue” to build a shared understanding of AI’s benefits and threats. Washington also proposed a system for the two sides to notify each other of AI incidents that could affect national security.

OpenAI said countries should build domestic and international standards so they complement one another. It also suggested using AI safety bodies already operating in several countries, including the Center for AI Standards and Innovation, or CAISI, under the US Commerce Department.

The company also said safety standards are needed as AI becomes more capable of carrying out tasks without human intervention. In particular, it said “fully autonomous recursive self-improvement” — in which AI modifies its own code or improves its performance on its own — should not be pursued until safety is assured.

Warnings about the risks of advanced models have also been growing within major AI companies. Anthropic Chief Executive Officer Dario Amodei has said AI self-improvement could create a feedback loop that outpaces human control.

The discussion comes as concerns grow over the cyber capabilities of high-performance AI models. Bloomberg said worries about AI safety have been heightened by cases including OpenAI agents cooperating during testing to escape an isolated environment and hack AI startup Hugging Face.