On September 21, OpenAI published an article titled “Building standards for the next phase of AI,” calling on the U.S. to lead an effort alongside other nations to establish global technical standards for frontier artificial intelligence, including recursive self-improvement (RSI). This represents the latest policy proposal in a series of responses from major AI developers addressing a wave of safety concerns over the past two weeks.

Two Pillars of the Proposal

The first pillar is a mechanism for national and international standards to complement one another. OpenAI suggests leveraging the existing network of AI safety institutes across Australia, Canada, Germany, France, Kenya, Japan, South Korea, Singapore, India, and the UK, connected via the U.S. Center for AI Standards and Innovation (CAISI). Two core priorities were outlined: frontier AI models and developers measured by capability benchmarks, and managing the benefits and risks of automated AI research, including RSI.

OpenAI stressed that these standards would not function as licensing regimes or mandatory pre-release censorship requirements. Each government would decide independently whether to codify them into law. The standards should also be openly designed and neutral toward specific companies, countries, or business models, avoiding undue burdens on open-weight developers and newcomers. OpenAI cited precedents from aviation and financial stability, where nations established shared technical standards and trusted channels of cooperation without relinquishing sovereign authority.

The second pillar focuses on shared measurement standards and incident reporting protocols. Specifically, OpenAI is advocating for standards to evaluate RSI-related progress and the level of research automation within an AI lab, human oversight requirements for automated research (identifying which workflows should trigger an immediate review), and common frameworks to categorize, track, and report alignment incidents on a unified severity scale. The company highlighted two initial contributions it previously released: a report on research acceleration and a model deviation reporting framework.

Stance on RSI: Full Automation Should Not Yet Be Pursued

The most striking section of the post is OpenAI’s direct stance on RSI: “Fully autonomous RSI does not occur today and we should not pursue it unless and until it can be done safely.” OpenAI warned of the technology’s potential downsides: pursued recklessly, RSI could lead humans to lose practical control over AI development, leaving them unable to oversee research workflows they no longer understand.

OpenAI also pointed to an incident involving Hugging Face that it had disclosed earlier as a “preview” of risks that could escalate dramatically without robust safeguards and alignment, even though that particular incident was not directly caused by RSI.

Self-Imposed Standards: Policy Response or an Effort to Shape Regulation?

On this issue, distinguishing between two aspects is essential:

  • The practical aspect: What OpenAI put forward remains strictly a policy proposal, with no binding legislation enacted. Even if officially adopted, these standards would serve merely as voluntary technical guidelines.
  • The perspective aspect: Amid pushback from safety researchers and the mathematics community, OpenAI’s proactive push for “shared standards” is seen by critics as an effort to ease public concern while securing an early advantage in shaping the rules of the game. Nevertheless, this remains a subjective interpretation rather than established fact.

Furthermore, advocating for the U.S. to take the lead reflects clear geopolitical considerations. According to OpenAI, given its strategic network position and premier technical capabilities, American initiative at this juncture will determine whether the nation shapes the global governance framework or confronts a fragmented system. In addition, the organization welcomed the ongoing U.S.–China dialogue on AI, describing upcoming discussions as promising opportunities.

The pivotal dynamic to watch moving forward is that both OpenAI and Anthropic are unilaterally putting forward their own “safety standards” without independent verification. The question remains: whose framework will be codified into law first, and which governing body will hold the authority to assess these companies’ compliance?