OpenAI shared a new set of proposals on Monday for how the AI industry should approach safety and security. The company posted the ideas in a blog post aimed at guiding the next stage of AI development.
The proposals center on two main areas. One is alignment research, which studies how to keep AI systems working in line with human goals. The other is a process called recursive self-improvement, or RSI.
"Navigating this transition safely requires alignment research to keep pace with these capabilities," OpenAI wrote in the post.
What Is Recursive Self-Improvement
RSI refers to AI systems that help build and improve future versions of themselves. This can happen with varying levels of human involvement.
OpenAI said automated AI research could lower the cost of advanced technology and help more people access it. The company also said automated systems could help improve safety work itself.
But the process carries risks. If AI systems begin improving themselves with little oversight, researchers may struggle to understand or control what happens next.
"Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely," OpenAI said in the post.
The company pointed to a recent security incident involving Hugging Face as an example. OpenAI said the incident was not caused by RSI, but it showed the kind of risk that could grow without strong safeguards.
OpenAI wants the United States to lead an international effort to build shared technical standards for frontier AI. It suggested building on existing AI safety institutes already operating in countries including Australia, Canada, Germany, France, Japan, and the United Kingdom.
The company said these standards should not act as licenses or approval requirements for AI models. Instead, individual governments would decide how to apply them under their own laws.
Industry Response And Recent Debate
OpenAI's proposal follows a similar move by rival Anthropic last week. Anthropic shared its own ideas for safe frontier AI development in response to a wave of warnings from researchers about AI risk.
That wave of warnings began after Jacob Coxon, who had worked at both Anthropic and OpenAI, announced his resignation nearly two weeks ago. Coxon said the two companies were "gambling with our lives."
Following Coxon's comments, Anthropic CEO Dario Amodei published an essay calling for AI companies to slow the pace of model development. Amodei also proposed adding independent third-party evaluators to audit AI companies for risk.
OpenAI CEO Sam Altman and Tesla and SpaceX CEO Elon Musk both voiced public support for Amodei's proposal.
A group of AI evaluators has also pushed foundation model makers to adopt a set of minimum conditions. These would give outside reviewers deeper access to AI systems and protect them from retaliation for publishing critical findings.
OpenAI said standards for incident reporting and human oversight of automated research will matter as AI systems take on more research work themselves.
The company also said it hopes for direct talks between the United States and China on AI safety topics, noting that upcoming discussions between the two countries are happening at a useful time.
