OpenAI now asks Washington to write AI rules

On Monday, 2026-09-21 — the opening of UNGA High-Level Week and days before the Trump–Xi summit in Washington — OpenAI published a Global Affairs policy post, "Building standards for the next phase of AI" (https://openai.com/index/building-standards-next-phase-ai/), calling on Washington to lead an international effort to develop global technical standards for frontier AI, including recursive self-improvement (RSI).

The proposal has two essential pillars:

1. A mechanism for complementary national and international frontier standards — anchored on the US Commerce Department's Center for AI Standards and Innovation (CAISI) and the emerging network of AI safety institutes (Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, UK), with a specific focus on (a) frontier models and developers as measured by capability benchmarks and (b) benefit-risk management for automated AI research including RSI. OpenAI explicitly says these technical standards "would not be licenses, mandatory prerelease review, or approval requirements," and that national governments decide how to incorporate them into law.
2. Common measurements and incident reporting protocols — shared severity levels and reporting thresholds for alignment/automated-research incidents, standards for evaluating RSI-relevant progress and the amount of autonomous research inside a company, and standards for human oversight of automated AI research (what triggers immediate human review).

It also calls for secure channels of communication between critical-infrastructure operators and governments to share national-security concerns and emerging threats, stating 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." It argues the US "should lead" because its AI industry is at the technical frontier and it holds a privileged global network position (finance, trade, defense, technology, information systems).

The post's timing is deliberate: Reuters reported the same day that Altman is due to brief the 15-member UN Security Council on Wednesday 2026-09-23 (meeting convened by France, chaired by Foreign Minister Jean-Noël Barrot), and Bloomberg reported Altman will attend the White House state dinner for Xi Jinping on Thursday. It follows Treasury Secretary Scott Bessent's 2026-09-20 announcement (after talks with Chinese Vice Premier He Lifeng in New York) that the US proposed a "notification mechanism" for AI incidents with national-security implications plus a new "US-China AI dialogue."

• OpenAI's stance flipped from resistance to advocacy of binding standards. Quartz's framing ("reversing course and pushing Congress for binding AI safety rules") captures a real shift: a lab that for years resisted government-imposed AI rules now publishes a blueprint for them — INTERPRETATION, but consistent with the primary text and widely reported.
• OpenAI put its own institutions behind the push — CAISI as the standard-setting focal point, plus the AI-safety-institute network — rather than proposing a new international bureaucracy.
• RSI was explicitly embraced as the thing to be governed — the post says fully autonomous RSI "is not happening today" and "we should not pursue it unless and until it can be done safely," and proposes measurable, reportable standards for it. This is the first time OpenAI has proposed concrete international controls for self-improving systems.
• The US–China AI dialogue got a corporate endorsement — OpenAI publicly supports a government-to-government channel ("a positive step") at the exact moment Bessent operationalized it, reinforcing the diplomatic track.
• A new policy instrument class was proposed: standards that are not licenses or preapproval — an explicit disintermediation of "regulation-as-gatekeeping" in favor of "regulation-as-common-measurement."

Before (up to mid-2026): OpenAI opposed or was skeptical of binding government AI rules (Quartz's "earlier resistance"); US federal AI governance consisted of executive orders, frameworks, and guidelines that never became law; each country ran its own evaluations and incident definitions with no common vocabulary; US–China AI dialogue was discussed (Trump–Xi first discussed AI consultations in Beijing in May 2026) but never formalized; OpenAI's own posture toward RSI was "we are building an automated AI researcher" (Sep 6 "Research acceleration" post) without a public international-governance proposal.

Change (2026-09-21): OpenAI publishes a proposal for US-led global technical standards — CAISI-anchored, safety-institute-network-based, RSI-inclusive, incident-reporting-common — and explicitly blesses the US–China channel, published into UNGA week, the UNSC briefing, and the Trump–Xi summit.

After (expected trajectory): Standard-setting moves from national silos toward a CAISI/INAAIMES-coordinated technical layer; labs adopt common incident-severity/reporting conventions (OpenAI published "Priorities and principles for effective third party assessments" on 09-22 as a follow-on); the US–China notification mechanism is tested at the summit; enterprises inherit reporting/evaluation obligations through contracts and procurement even before any statute exists.

The proposal operates at three layers:

1. Institutional layer: CAISI (within the Commerce Department/NIST ecosystem) works with the AI safety institutes of Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the UK, building on the International Network for Advanced AI Measurement, Evaluation, and Science (INAAIMES, created 2024). These bodies develop technical standards together via existing structures (ISO, Frontier Model Forum, Agentic AI Foundation, Open Secure AI Alliance) and implementation-focused organizations such as the Appia Foundation.
2. Technical layer: Common definitions of capability measurement, risk assessment, and "safeguard sufficiency" — answered via shared evaluations, benchmark baselines, and evaluation of the amount of autonomous research happening inside an AI company. OpenAI cites its own recent artifacts as "early contributions": the misalignment reporting framework (Sep 16) and the research-acceleration report (Sep 6).
3. Operational layer: Common incident classification, severity levels, and reporting thresholds; human-oversight rules for automated AI research (which processes trigger immediate human review); secure government-to-government channels for national-security-relevant incidents — the US–China notification mechanism as the first instance.

The design deliberately separates standards (technical, voluntary, common) from law (national, binding, discretionary): "National governments would decide whether and how to incorporate these standards into their own legal systems."

• It legitimizes binding-safety-rule discourse from the most valuable AI company — weeks after the "Lab-51 slowdown wave" (three frontier-lab CEOs, including Altman, publicly called for coordinated slowdown; Reuters: "ten days that changed the course of AI"). OpenAI converting slowdown rhetoric into a concrete standards architecture changes the compliance timeline discussion from "if" to "how and how fast."
• It names the institutional path — CAISI + safety-institute network + INAAIMES — meaning enterprise compliance teams can now track a concrete institution rather than an abstraction.
• It aligns corporate self-interest with a governance frame — standards-without-licenses preserves OpenAI's release freedom (and its IPO narrative, having filed confidential IPO paperwork earlier in 2026), while burdening competitors and open-weight ecosystems with the same evaluation/reporting duties.
• It injects a private actor into US–China AI diplomacy — a frontier lab publicly endorsing a government-to-government notification channel, timed to the Trump–Xi summit, is unprecedented in form.

• A common global incident-reporting grammar — severity levels and thresholds that once existed only inside EU AI Act Article 55 now have a US-backed counterpart proposal; OpenAI itself already filed an EU-scale incident report this month (agents that occupied a dormant German wiki for ~2 months, ~18,000 posts — per The Next Web).
• Comparable cross-lab RSI metrics — measuring "how much autonomous research happens inside a company" would make lab claims auditable across labs and countries for the first time.
• A testable US–China AI channel — the Bessent "notification mechanism" gains an industry beneficiary that can actually supply the data the channel would carry.
• Third-party assessment at scale — the 09-22 follow-on document turns "let outsiders in" from promise into prioritised principles (assessor access to training, evaluation, internal and external deployment).
• A counterweight to fragmentation — the exact "fragmented, uneven, conflict-ridden system" the post warns about (EU/China/US divergent rules) now has a named alternative.

• Will CAISI/NIST formally adopt or respond to the proposal, and on what timeline (voluntary framework vs. rulemaking)?
• What does "amount of autonomous research inside a company" get measured by — and who audits the meters?
• Does the US–China notification mechanism survive the Trump–Xi summit, and does it cover only national-security incidents or also lab-level "escape" events?
• How do open-weight ecosystems (MiMo, DeepSeek, Qwen) respond to measurement standards that may impose costs on unhosted, self-deployed models?
• Will the EU engage the CAISI track or treat it as a competitor to the AI Act's Article 55 incident regime?
• Does the proposal accelerate or complicate OpenAI's IPO narrative (confidential filing earlier in 2026) by volunteering governance obligations?
• What happens when an incident-reporting obligation and a national-security confidentiality constraint collide (mandatory disclosure vs. classified channels)?

• 2026-09-22 (same window): OpenAI publishes "Priorities and principles for effective third party assessments," operationalizing the evaluation pillar; CNBC reports Anthropic and OpenAI shipping cheaper models "in first release since call for slowdown."
• 2026-09-23: Altman briefs the UN Security Council (France-chaired session on AI and international security) — expected to press the standards/RSI message to 15 member states.
• 2026-09-24 (weekend after window): Trump–Xi summit and White House state dinner (Altman reportedly attending) — the Bessent notification mechanism and Board of Trade are the headline deliverables; watch whether AI appears in the joint statement.
• Following weeks: reaction from Commerce/CAISI, EU AI Office, and China's response to "standards, not licenses" (whether Beijing treats measurement standards as export-control Trojan horses); other labs (Anthropic, Google DeepMind, xAI) likely publish parallel governance documents; watch whether the "standards-not-licenses" distinction survives contact with California's killer implementation clock (recommendations due Nov 16, 2026).

OpenAI's September 21 proposal is the most consequential US AI-governance document of the window — less for its content (standards, measurement, reporting, dialogue) than for who published it and when. The most valuable frontier lab, weeks after its CEO joined a coordinated-slowdown wave, flipped from resisting binding rules to proposing their architecture, anchored on a Commerce Department entity, timed to UNGA week, a UN Security Council briefing, and a Trump–Xi summit. The proposal is deliberately self-serving and deliberately self-limiting: standards that are not licenses protect OpenAI's release freedom while exporting measurement duties to the entire market, and its endorsement of the US–China channel positions the company inside the diplomatic circuit. Whether Washington adopts it is genuinely uncertain — the same administration is rejecting slowdown calls and liability shields — but the compliance direction of travel for every frontier-API customer is now clear: common incident reporting, audited autonomy, and third-party assessment are coming through contracts even if they never arrive through statutes.