OpenAI committed in writing on 9 September 2026 that "when proceeding would pose an unacceptable safety risk, we will slow or stop the development or deployment of systems we cannot sufficiently safeguard." The post, The AI policy window is open. We need to act, names no threshold, date or outside referee for that decision. That makes it the most important and least testable commitment in a week when lab researchers, British lawmakers and the US president all took public positions on slowing down.
Key facts
- OpenAI (9 September): will "slow or stop" at an unacceptable safety risk, "as required per our preparedness framework"; calls for "mandatory, capability-based national regulation" and backs four California bills. Read the post.
- Sam Altman (reported 11 September): told staff OpenAI could potentially pace development alongside other labs, according to people familiar with the matter who spoke to Bloomberg; OpenAI declined to comment.
- United Kingdom (11 September): more than 70 MPs and peers urged Prime Minister Andy Burnham to back a bill banning artificial superintelligence; the government said the bill is not "the right approach."
- President Trump (10 September): asked whether he had concerns about AI extinction risks: "No, I don't have any."
What OpenAI actually committed to
The post goes further than OpenAI's recent essays. On recursive self-improvement, AI systems that substantially improve the next generation of AI with little human input (see our lesson on recursive self-improvement), it says: "Fully autonomous recursive self-improvement ... is not happening today. We should not pursue it unless and until it can be done safely." It adds that "if we cannot meet certain safety bars without slowing down capability growth, we should prioritize the former," and asks for "shared safety bars for when and how development should slow or stop."
It also takes sides in state politics, announcing support for four California bills: SB 813 on independent safety assessments, AB 1405 on AI-auditor standards, SB 1119 on protections for young people, and AB 1864 on safeguards against AI-enabled biological threats.
What it does not do is say where the line is. The Preparedness Framework it cites has capability categories and risk levels, the same family of ideas as the capability thresholds and responsible scaling policies other labs publish, but the post does not say which finding would trigger a slowdown or who outside OpenAI would be able to check. A car maker that promises to brake "if the road becomes unsafe" has made a sincere promise, but you cannot hold it to one.
This builds on OpenAI's 6 September essay, which ranked self-improvement and alignment above mathematics research and floated coordinated slowing without a trigger. The Bloomberg report, published by Business Standard, says Altman told a company-wide meeting "that OpenAI could potentially pace its AI development — perhaps in conjunction with several other AI labs, but that some may not agree to do so." There is no direct quote from Altman and no condition attached.
Who else took a position
Lab researchers. Individual employees posted in personal capacities. OpenAI's Julie Steele wrote, "In my personal capacity, I also think we need to slow down," CNBC reported. Anthropic's alignment lead had already said there is no plan yet for superintelligence. As a company, Anthropic told the Guardian that "the world would benefit from the industry adopting a lawful, verifiable way to work together to pace how we release powerful models." Neither lab named a trigger.
British lawmakers. According to the Guardian, more than 70 MPs and peers wrote to Burnham urging him to back Alex Sobel's bill, which would prohibit artificial superintelligence outright. They want Britain to use its G20 presidency to build "a coalition of the willing." Sobel said: "AI companies have made it clear that they will develop agents without constraint." The government replied: "While we do not consider the measures proposed in the bill to be the right approach, we are exploring whether additional targeted interventions may be needed in future."
The White House. Asked by reporters on 10 September whether he had concerns about AI extinction risks, President Trump said, "No, I don't have any," adding, per CNBC: "I have concerns that if we don't win AI, we're going to be put in a very bad position." In the US Senate, a separate bill would ban superintelligence and pause frontier training.
Why it matters
The week's scorecard, in one line: one lab has promised to slow or stop without saying when, one CEO has reportedly said he is open to pacing with rivals, one lab backs a verifiable industry process it has not designed, one parliament is being pressed to ban the technology outright, and one head of state has declined to engage with the risk at all. Promises to brake matter only if someone can check whether the brakes were applied. The next useful document from any of these actors would be a concrete trigger: a named capability, a measurement, and an outside party entitled to see it.
The honest caveat is that Altman's remarks reach the public only through anonymous sources, and OpenAI's post is a policy statement, not a binding rule. Coordinated slowing between competitors also raises legal questions that none of these statements addresses.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)