First-hand AIArticle·22 September 2026·5 min read

OpenAI Wants Global Standards Before AI Self-Improvement Takes Off

The lab is calling for an international technical framework to govern the moment AI starts doing AI research on its own — and wants the US to hold the pen.

ShareWhatsApp X LinkedIn Facebook

Save to my selection

OpenAI Wants Global Standards Before AI Self-Improvement Takes Off

Source : OpenAI · OpenAIView original ↗

In brief

In a policy post, OpenAI argues for international technical standards on frontier models, with an explicit focus on recursive self-improvement (RSI): comparable evaluations, thresholds for human oversight, common incident-reporting protocols. The lab states that fully autonomous RSI should not be pursued until it can be done safely, and calls on the US to lead the effort by leveraging the network of AI safety institutes.

🍺 Bar-stool version

OpenAI explains it's working on AI that can do AI research, and that if things speed up too fast, nobody will understand what's happening anymore. Its pitch: get the whole world to agree on a common way to measure danger, kind of like civil aviation agreed on what counts as an in-flight incident. The juicy detail is that it's the plane-builders suggesting the rulebook, and they're careful to specify it shouldn't be a takeoff permit. Maybe it's sincere, maybe it's strategic, but it's also the first document where a lab writes in black and white that there's a line it won't cross yet.

Key takeaways

1
OpenAI puts RSI (recursive self-improvement) at the center of the debate: as models take over AI R&D, the pace of progress could accelerate sharply.
2
Explicit position: fully autonomous RSI is "not happening today" and should not be pursued until it can be done safely, under human control and via informed democratic choices.
3
Three problems justify going international, per the lab: fragmentation of evaluations and incident definitions, collective action (each country acting alone produces an outcome nobody wants), and unevenly distributed capabilities.
4
The proposed vehicle: the network of AI safety institutes (Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, UK) linked to the US CAISI and its International Network for Advanced AI Measurement.
5
Stated red line: these standards would be neither licenses, nor mandatory pre-publication review, nor model approval — each state would decide whether to adopt them.
6
Three families of targeted standards: measuring the volume of autonomous research inside a lab, thresholds triggering human review, and classification/reporting of alignment incidents with shared severity levels.
7
The text calls for secure channels between critical infrastructure operators and governments, and views US-China dialogue on these topics positively.

RSI as the tipping point

The post starts from the program laid out by Sam Altman and Jakub Pachocki: build an automated AI researcher, iterate with it on the alignment problem, and keep humans in the loop of self-improvement. Everything flows from that last point.

OpenAI describes recursive self-improvement as a continuum rather than a switch: the more systems take over developing the next generations, the more the process automates, and the faster the pace can accelerate. The optimistic argument is symmetrical: an automated AI researcher can also be an automated safety researcher, capable of lowering the cost of advanced intelligence and strengthening defenses.

The risk is stated bluntly: if mismanaged, RSI could cause humans to lose practical control over AI development, unable to oversee research processes they no longer understand. The lab cites its 'Hugging Face Incident' as a preview of this type of risk, while clarifying it did not result from RSI.

Why go international, not just national

OpenAI acknowledges the usefulness of existing institutions — CAISI, US state laws, a federal framework — but argues they're not enough once development, deployment, and effects span multiple countries.

Three failures are named. Fragmentation: diverging evaluations, reporting obligations, and incident definitions make evidence incomparable. Collective action: isolated national decisions can add up to an outcome nobody wanted. Uneven capabilities: frontier expertise is concentrated, which worsens the first two points. The lab specifies this applies to both open and closed models.

The text claims a political motivation: avoiding concentration of power. Standards would let actors outside the labs have a say and would provide a visible reference point, independent of any given company's internal practices.

The proposed setup

First part: a mechanism linking national and international standards, building on the existing network of AI safety institutes and on the International Network for Advanced AI Measurement, Evaluation, and Science created by CAISI in 2024. The scope would be targeted: frontier models and developers, identified via capability benchmarks, and the risk-benefit management of automated research.

OpenAI stresses what these standards would not be: not licenses, not mandatory pre-publication review, not prior approval. Governments would remain free to incorporate them into law or not. The precedents cited are aviation and financial stability — fields where technical standards are shared without ceding sovereignty. Partners named include ISO, the Frontier Model Forum, the Agentic AI Foundation, the Open Secure AI Alliance, and the Appia Foundation.

Second part: shared metrics and incident protocols. Concretely, evaluating the share of autonomous research done inside a lab (OpenAI's report on research acceleration is presented as an initial contribution), defining which automated processes trigger immediate human review, and harmonizing classification and reporting thresholds for alignment incidents — with its misalignment reporting framework as a draft.

The geopolitical angle

The conclusion is openly American. The US should lead the effort because its industry is at the technical frontier and because it holds a privileged network position in finance, trade, defense, and information systems.

The alternative is framed as a binary choice: either Washington shapes the global framework, or it watches a fragmented, unequal, conflictual system take hold. OpenAI adds that competition will not just be about technical lead but about adoption and diffusion — which will go to whoever offers credible rules of the game.

The post finally calls for secure communication channels between governments and critical infrastructure operators, and calls a US-China dialogue on these issues 'a positive step,' noting that upcoming talks come at an opportune moment.

“Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.”

“What does good look like in the mitigation of catastrophic AI risk?”

“An automated AI researcher can also be an automated AI safety researcher.”

Why it matters

This is the document where OpenAI moves from talking about safety to a dated, named institutional proposal, with a precise target: automated AI research and its self-improvement. The phrase to remember is the conditional commitment — not pursuing fully autonomous RSI until it can be done safely — because it creates, at least rhetorically, an actionable benchmark. Still, the framework described carefully checks every box favorable to its author: voluntary technical standards, measured by benchmarks, explicitly stripped of any prior-approval power, led by the country where the lab is based. This is soft law, and the question of who writes the thresholds becomes as decisive as the thresholds themselves. Worth watching: the shift from 'initial contributions' (the research-acceleration report, the misalignment reporting framework) to metrics that third parties could verify without going through the labs.

#ai #openai #governance #safety #agi #policy

Original source

Building standards for the next phase of AI

OpenAI

Open the article ↗