AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Source Summaries

Building Standards for the Next Phase of AI (OpenAI, September 2026)

medium confidence · updated 2026-09-23

OpenAI Global Affairs position paper of September 21, 2026 arguing that the United States should lead an international effort to develop global technical standards for frontier AI, including for recursive self-improvement, through CAISI and the network of national AI safety institutes. Proposes standards for capability measurement, human oversight of automated AI research and incident classification, states that the standards would not be licenses, mandatory prerelease review or approval requirements, and leaves incorporation into law to national governments.

Contents

"Building standards for the next phase of AI" is a position paper published by OpenAI on September 21, 2026 under its Global Affairs heading, with a corporate byline. It argues that "the United States should lead an effort to work together with countries around the world to develop global technical standards for frontier AI, including for RSI" (recursive self-improvement), and sketches two elements it calls essential: a mechanism for setting complementary national and international standards, built on CAISI and the network of national AI safety institutes, and common measurements and incident-reporting protocols for automated AI research. The paper is a statement of the company's position and is summarized here as such.

It was published the same day Treasury Secretary Scott Bessent described the outcome of US–China AI talks, and Axios reported it under the headline "OpenAI proposes AI standards after U.S., China talks" (Source: axios.com). The paper refers to upcoming US–China talks without describing their content. One day later OpenAI published "Priorities and principles for effective third party assessments", which links this paper as the "shared international standards for safety and security practices" under which labs and independent assessors should operate.

Framing: OpenAI's goals and automated research

The paper opens by restating OpenAI's mission and three goals it attributes to recent statements by Sam Altman (Built to Benefit Everyone: Our Plan (Altman & Pachocki, OpenAI, June 2026)) and Jakub Pachocki (An Alien Mind (Jakub Pachocki, OpenAI, September 2026)): navigating the next period of AI progress "by building an automated AI researcher, iterating with it on the alignment problem, and finding ways for people to remain part of the self-improvement loop"; delivering the benefits of scientific progress and economic growth; and empowering "everyone individually with a personal AGI." It cites OpenAI's research-acceleration report (Research acceleration: The view inside OpenAI (OpenAI, September 2026)) and its mathematics results, including the Navier–Stokes Millennium Problem (On the Navier–Stokes Millennium Prize Problem (OpenAI, September 2026); Ten Advances in Mathematics and Theoretical Computer Science), as evidence that AI is accelerating the company's own research.

The stated premise is that alignment research must "keep pace" with capabilities as systems take on more of AI research and development, and that making continued progress safe "will require both advances in alignment research and shared standards to guide development across labs and countries."

Position on recursive self-improvement

The paper describes automated AI research as a spectrum of human supervision: as AI systems take on more of the work of developing successive generations, "they can increasingly drive a process of recursive self-improvement (RSI), even while people remain involved," and as the process becomes more automated "the pace of AI progress could accelerate rapidly." It presents benefits, including that "an automated AI researcher can also be an automated AI _safety_ researcher," and that more capable aligned systems could help secure critical infrastructure and defend against dangerous AI agents.

Its central statement of position is: "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely." Whether and how to proceed, it states, "must depend on our ability to preserve human control and on informed democratic choices about the benefits and risks," the latter phrase linking OpenAI's June 2026 federal blueprint (Democratic Governance of Frontier AI: A blueprint for a federal framework (OpenAI, June 2026)). Done without care, RSI "could result in humans losing practical control over AI development, unable to provide oversight on research processes they no longer understand." The paper calls the Hugging Face incident OpenAI disclosed, "while not a direct result of RSI," "a preview of the kinds of risks that could become much more severe without robust safeguards and alignment."

The RSI sentence restates, with minor rewording, the position chief global affairs officer Chris Lehane set out on September 9, 2026 (The AI policy window is open. We need to act. (Chris Lehane, OpenAI, September 2026)). As in that post, the paper does not define the safety condition or say who would determine that it has been met.

The case for international standards

OpenAI argues that international standards for safety and security practices "may be as important to pacing the frontier as alignment research itself," because they "can create shared definitions of high-quality evidence and agreed-upon baselines for the rigor of technical safeguards," answering the question "What does good look like in the mitigation of catastrophic AI risk?" It names CAISI, state laws, a federal framework and "new forms of public-private partnerships" as domestic contributors, the last linking Demis Hassabis's July 2026 framework essay (A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026)). Because development and adoption are global, it identifies three problems that call for international work:

  • Fragmentation. National evaluations, reporting requirements and incident definitions could conflict, "making it harder to compare evidence, understand emerging capabilities, and respond to risks that cross borders."
  • Collective action. "Each nation acting independently can produce outcomes that no nation wants"; RSI could accelerate research "potentially beyond our collective ability to understand progress, assess risks, and maintain meaningful human oversight."
  • Uneven capacity. Frontier development and expertise are unevenly distributed, which "accentuates both of the problems above."

The paper states that these challenges "apply to both open and closed models." It also states that any lab pursuing automated AI research "must take accountability for doing so safely, in accordance with basic principles of self-responsibility and existing laws," and gives as its rationale for standards "avoiding the concentration of power, and producing better practical outcomes": standards give "more stakeholders outside of the labs" a say and provide "a visible set of principles that can be relied on irrespective of the specific practices of any particular lab" (see AI Power Concentration).

Element 1: a mechanism for complementary national and international standards

The first element proposes leveraging "the emerging network of AI safety institutes — such as those already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom" to facilitate standard-setting "through the CAISI and national industry bodies," focused on frontier models and developers "as measured by capability benchmarks," and on benefit-risk management for automated AI research, including RSI. It states that the United States "can build upon CAISI's creation of the International Network for Advanced AI Measurement, Evaluation, and Science in 2024" (International AI Safety Institute Network (INSAI / AISIN)).

The resulting standards would provide "a common technical foundation for capability measurement and evaluation, risk assessment, and safeguard sufficiency." The paper sets limits on them: "These technical standards would not be licenses, mandatory prerelease review, or approval requirements for AI models. National governments would decide whether and how to incorporate these standards into their own legal systems." That disclaimer distinguishes the proposal from licensing and pre-release review regimes discussed at AI Pre-Release Vetting, and is consistent with the evaluate-don't-license posture of the June blueprint.

The process should consult open- and closed-model developers, independent experts and academia, and standards should be "developed transparently and designed so that they do not advantage particular companies, countries, or business models, including by making it harder for new entrants or open-weight developers to compete." Citing an earlier OpenAI suggestion reported by the Financial Times, the paper points to aviation and financial stability as fields in which countries built common technical standards and cooperation channels "without giving up national authority." It names the bodies the effort should work with: ISO (linking the ISO/IEC JTC 1/SC 42 committee; see ISO/IEC 42001 — AI Management System), the Frontier Model Forum, the Agentic AI Foundation, the Open Secure AI Alliance, and the Appia Foundation, described as "developing practical specifications that connect international standards with real-world AI assessments."

Element 2: common measurements and incident-reporting protocols

The second element lists three areas in which the paper says international standards "will be particularly crucial" for managing autonomy in AI research:

  1. Evaluation of RSI-relevant progress "and the amount of autonomous research happening within an AI company," for which OpenAI calls its research-acceleration report (Research acceleration: The view inside OpenAI (OpenAI, September 2026)) "an initial contribution."
  2. Human oversight over automated AI research, "including what kinds of automated AI research processes should trigger immediate human review" (see Human Oversight).
  3. Incident classification, tracking, reporting and response for alignment and automated-research issues, "such as common incident severity levels and reporting thresholds," for which it calls its misalignment reporting framework (Our Framework for Reporting Model Misalignment (OpenAI, September 2026)) "an early contribution" (see AI Incident Reporting).

Beyond standards, the paper calls for critical-infrastructure operators and governments to establish secure communication channels to share national security concerns, vulnerabilities, threats and best practices, and states that "Dialogue between the United States and China in these areas would be a positive step, and upcoming talks are happening at an opportune time" (see US-China AI Competition: Different Races, Different Metrics).

Pacing and US leadership

The closing section defines pacing. "Pacing AI development is not about maintaining a predetermined speed. Technically, it is about ensuring that alignment research and deployment of that research stay ahead of capabilities. Societally, it is about the use of standards, institutions, and collaboration" to produce "positive sum outcomes," with the August 2026 call for collective action on cyber defense (A call for collective action on cyber defense (open letter, August 27, 2026)) given as an example. The phrase "pacing the frontier" was used in the July 2026 employee statement (Pacing the Frontier (statement from employees of frontier AI companies, July 2026)) and in Dario Amodei's September 12 essay (We Must Pace the Frontier (Amodei, September 2026)), where it refers to deliberately slowing development subject to verification; the OpenAI paper's definition turns on the relative pace of alignment and capabilities and names no speed limit. The positions are compared at Coordinated slowdown proposals compared.

The paper argues that the United States is positioned to lead because its industry is at the technical frontier and it holds "a privileged global network position" in finance, trade, defense, technology and information systems, and that leading now "will determine whether the United States shapes the global AI framework or watches a fragmented, uneven, and conflict-ridden system take hold around it." It predicts that geopolitical competition will turn on adoption and diffusion as well as technical lead, the latter possibly "determined by whoever is driving realistic cooperation and sensible rules of the road."

Discrepancies and points of comparison

  • Date of the International Network. The paper credits CAISI with creating the International Network for Advanced AI Measurement, Evaluation, and Science "in 2024," while the linked source is a NIST news item dated February 13, 2026. That item states that the network "was founded by the Center for AI Standards and Innovation (CAISI) in November 2024," although the US institute operated under the name AI Safety Institute until its mid-2025 rename to CAISI; it lists members from ten jurisdictions — Australia, Canada, the European Union, France, Japan, Kenya, Korea, Singapore, the United Kingdom and the United States (Source: nist.gov). The paper's list of institutes adds Germany and India and omits the European Union; it introduces the list with "such as" and does not describe it as the network's membership.
  • Relation to earlier OpenAI proposals. The June 2026 blueprint (Democratic Governance of Frontier AI: A blueprint for a federal framework (OpenAI, June 2026)) proposed that CAISI develop RSI measurement methodologies and that certified third-party assessors evaluate developers' RSI measurements; the September paper moves the same measurement agenda to an international standard-setting process and does not mention certification of assessors. Lehane's September 9 post (The AI policy window is open. We need to act. (Chris Lehane, OpenAI, September 2026)) committed OpenAI to advance industry-led standards "with or without government support"; this paper places standard-setting with public institutions and national industry bodies.

Reception

Axios reported the paper on its publication day (Source: axios.com). No detailed independent assessment of the proposal had been located as of September 23, 2026.

Provenance

Published at openai.com/index/building-standards-next-phase-ai/ on September 21, 2026, Global Affairs tag, byline "OpenAI." Retrieved September 22, 2026 by the gap scan; the distinctive passage on autonomous RSI was confirmed in the fetched text and matched the quotation carried in contemporaneous digest coverage. The linked NIST item was fetched on September 23, 2026 to check the paper's dating of the International Network. Confidence is medium: the document is the company's own statement of position, reported by one independent outlet, with no independent evaluation of its proposals as of ingest.

Relationships

Categories: Source Summaries | us | global | proposal | framework | standard | safety | alignment | frontier-models | national-security | company | developer
What links here (24)