ChatGPT-maker OpenAI has published a new set of proposals for AI safety standards urging international cooperation to regulate frontier artificial intelligence. The company has also stressed the need for alignment research in order to keep pace with advanced capabilities, particularly around recursive self improvement (RSI) which us a computing technique which enables the AI models to upgrade themselves without any human involvement. In its blog post, OpenAI warned: “Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.” The company cautioned that without proper safeguards, RSI could lead to humans losing practical control over AI development. OpenAI’s core proposal In a blog post published on September 21, OpenAI argued that navigating AI's rapid advancement safely requires alignment research to keep pace with growing model capabilities, so the systems it and other companies build remain aligned with human values and stay under human control. The company called for international cooperation to build shared frontier safety standards, recommending that this work build on existing AI safety institutes already operating around the world rather than starting from scratch.OpenAI said these technical standards should specifically target frontier AI models and their developers, along with benefit-risk management practices for automated AI researchers — a category that includes RSI. Why RSI is the central concern Recursive self-improvement has generated real excitement in the AI field for its potential to produce foundation models capable of upgrading themselves without human involvement. But that same potential has fueled growing concern among technologists, who worry that model developers could eventually lose control of the underlying technology, or fail to anticipate unintended consequences as these systems grow more complex and more deeply embedded across the internet.OpenAI was direct about where it stands on this today, stating plainly that fully autonomous RSI isn't happening yet and shouldn't be pursued until it can be done safely, warning that without proper care, RSI could leave humans unable to meaningfully oversee AI research processes they no longer fully understand.The company pointed to the Hugging Face agent hack from last month — which didn't actually involve RSI — as an early example of the kinds of risks that could grow far more severe without stronger safeguards and alignment work in place. Part of a broader industry reckoning OpenAI's proposals land just a week after rival Anthropic rolled out its own framework for safe frontier AI development, itself a response to a growing wave of warnings from researchers about AI's potential risks to humanity. That wave gained significant momentum after Jacob Coxon, who has worked at both Anthropic and OpenAI, announced his resignation nearly two weeks ago, publicly stating that AI companies were "gambling with our lives."In the aftermath of recent AI-related security incidents and Coxon's public statements, Anthropic CEO Dario Amodei published an essay calling on AI companies to slow the pace of foundation model development, among other proposals — including embedding independent third-party evaluators directly within companies to audit and help mitigate risks like AI-accelerated cyberattacks or bioweapon development. OpenAI CEO Sam Altman and Tesla and SpaceX CEO Elon Musk both publicly backed Amodei's proposal at the time. Where things stand now Despite the growing consensus that outside oversight is needed, the field of AI evaluation remains young, and there's still no unified agreement on the basic standards that would let independent third parties meaningfully scrutinize frontier AI systems beyond current practices. That gap is part of why a coalition of AI evaluators has been pushing foundation model makers to adopt a set of "minimum conditions" that would grant deeper access for audits and protect evaluators from retaliation when they publish unflattering findings.OpenAI's latest move suggests the company is looking to shape what those eventual global standards look like — positioning itself at the center of a debate it helped intensify just weeks ago when it joined calls for the industry to pace itself more carefully.
After telling American AI firms to 'pause', OpenAI proposes standards to the world
ChatGPT-maker OpenAI has published a new set of proposals for AI safety standards urging international cooperation to regulate frontier artificial intelligence. The company has also stressed the need for alignment research in order to keep pace with advanced capabilities, particularly around recursive self improvement (RSI) which us a computing technique which enables the AI models to upgrade themselves without any human involvement. In its blog post, OpenAI warned: “Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.” The company cautioned that without proper safeguards, RSI could lead to humans losing practical control over AI development. OpenAI’s core proposal In a blog post published on September 21, OpenAI argued that navigating AI's rapid advancement safely requires alignment research to keep pace with growing model capabilities, so the systems it and other companies build remain aligned with human values and stay under human control. The company called for international cooperation to build shared frontier safety standards, recommending that this work build on existing AI safety institutes already operating around the world rather than starting from scratch.OpenAI said these technical standards should specifically target frontier AI models and their developers, along with benefit-risk management practices for automated AI researchers — a category that includes RSI. Why RSI is the central concern Recursive self-improvement has generated real excitement in the AI field for its potential to produce foundation models capable of upgrading themselves without human involvement. But that same potential has fueled growing concern among technologists, who worry that model developers could eventually lose control of the underlying technology, or fail to anticipate unintended consequences as these systems grow more complex and more deeply embedded across the internet.OpenAI was direct about where it stands on this today, stating plainly that fully autonomous RSI isn't happening yet and shouldn't be pursued until it can be done safely, warning that without proper care, RSI could leave humans unable to meaningfully oversee AI research processes they no longer fully understand.The company pointed to the Hugging Face agent hack from last month — which didn't actually involve RSI — as an early example of the kinds of risks that could grow far more severe without stronger safeguards and alignment work in place. Part of a broader industry reckoning OpenAI's proposals land just a week after rival Anthropic rolled out its own framework for safe frontier AI development, itself a response to a growing wave of warnings from researchers about AI's potential risks to humanity. That wave gained significant momentum after Jacob Coxon, who has worked at both Anthropic and OpenAI, announced his resignation nearly two weeks ago, publicly stating that AI companies were "gambling with our lives."In the aftermath of recent AI-related security incidents and Coxon's public statements, Anthropic CEO Dario Amodei published an essay calling on AI companies to slow the pace of foundation model development, among other proposals — including embedding independent third-party evaluators directly within companies to audit and help mitigate risks like AI-accelerated cyberattacks or bioweapon development. OpenAI CEO Sam Altman and Tesla and SpaceX CEO Elon Musk both publicly backed Amodei's proposal at the time. Where things stand now Despite the growing consensus that outside oversight is needed, the field of AI evaluation remains young, and there's still no unified agreement on the basic standards that would let independent third parties meaningfully scrutinize frontier AI systems beyond current practices. That gap is part of why a coalition of AI evaluators has been pushing foundation model makers to adopt a set of "minimum conditions" that would grant deeper access for audits and protect evaluators from retaliation when they publish unflattering findings.OpenAI's latest move suggests the company is looking to shape what those eventual global standards look like — positioning itself at the center of a debate it helped intensify just weeks ago when it joined calls for the industry to pace itself more carefully.
Read the full story at Times of India.
