Product1 publisher3 min readPublished
OpenAI asks the US to lead frontier AI standards routed through CAISI and existing safety institutes
The company's 21 September proposal puts recursive self-improvement in scope for technical standards coordinated by CAISI. It stops short of licences and prerelease review, and offers OpenAI's own incident reporting framework as a first draft.
The Product Desk · Product desk
What happened
- OpenAI published a proposal on 21 September asking the United States to lead other countries in setting global technical standards for frontier models, with recursive self-improvement in scope.
- The company wrote that the standards would not be licences, mandatory prerelease review or approval requirements, and that national governments would decide whether to write them into their own law.
- Coordination would run through the US Center for AI Standards and Innovation and national industry bodies, with capability benchmarks deciding which models are covered.
- Treasury Secretary Scott Bessent told CNBC on Monday that the administration would not grant frontier AI labs protection from liability, according to Business Insider.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Assurance about a frontier model stays a buyer's own job. With no approval step in the design, the evidence a customer sees is whatever the vendor chooses to publish, so procurement has to write the reporting it wants into the contract.
- contradiction OpenAI says the standards should not favour particular companies or business models, and in the same document supplies the starting text for the incident scheme. The first severity scale that ships cannot do both.
- exposure Meeting the standard leaves a lab's liability where it was. Bessent has ruled out liability protection, so a voluntary standard becomes the yardstick a plaintiff cites rather than the defence a lab claims.
- constraint Tying scope to capability benchmarks means the benchmark owner sets the regulatory boundary. A model that scores below the line sits outside the oversight regime whatever it is used for.
Read the proposal for something you could put in a runbook and you land on the second of its three areas: human oversight, including which automated research processes should trigger immediate human review [7]. OpenAI leaves the trigger open. It asks the standards process to find one, alongside measures of how much autonomous research happens inside an AI company and a common scheme for classifying, tracking and reporting incidents by severity [7].
Scope would be set by capability benchmarks, with coordination sitting at CAISI and national industry bodies [4]. The base already exists. OpenAI suggests building on the safety institutes in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the UK [5], ten countries in all [6]. The company compared the arrangement to aviation and financial stability, where countries share technical standards without giving up national authority [13].
OpenAI wrote that the work should consult open and closed model developers, independent experts and academia, and that the standards should not favour particular companies, countries or business models [9]. In the same document it put forward its own misalignment reporting framework as an early contribution [8].
On the substance, OpenAI wrote on Monday that "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely" [10]. The company warned that RSI done without care could leave humans unable to oversee research processes they no longer understand. It cited the Hugging Face incident it disclosed as a preview. That incident was not a direct result of RSI, the company said [11].
Washington gave the proposal a mixed week. Trump repeated on Truth Social on Monday that he favours a light touch on AI, according to Business Insider. He wrote: "We will be careful, and that's why we have the Department of Justice, and other Law Enforcement bodies, that will rein things in if we have to." [16]. On Tuesday, UN Secretary-General Antonio Guterres warned of AI's dangers and called for global cooperation. Trump dismissed the idea as a "globalist scheme", Bloomberg reported, and said in the same General Assembly speech that the US was renaming AI "super intelligence" [18].
Sam Altman is due to brief the UN Security Council in person on Wednesday afternoon, 23 September [19][20]. Anthropic's Dario Amodei joins by video alongside Yoshua Bengio and Hugging Face's Clement Delangue, with French foreign minister Jean-Noel Barrot in the chair [19]. According to Security Council Report, it is the council's first meeting focused on the safety risks of increasingly capable AI, including the potential loss of human control [19].
For anyone buying frontier capability this year, the question to ask is what document a standard hands you. Under licensing, it is an approval you can ask a regulator to produce. Under this proposal, it is whatever the vendor publishes under the incident-reporting area. The enforceable version is the one written into your contract: which severity levels get reported to you, and within how many hours. If you run critical infrastructure, the proposal also asks you to stand up a secure channel with government for sharing threats [22].
What to watch
- Whether CAISI publishes a capability benchmark threshold that defines which models count as frontier for these standards.
- Whether the US-China AI dialogue Bessent described takes up the AI incident alert system Washington proposed to Beijing.
- Whether OpenAI's priorities for independent assessors, published the day after the proposal, extend assessor access beyond what the company already grants.