Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman told the UN Security Council on September 23, 2026, that their companies would slow frontier AI work when safety demanded it. Altman spoke about whether a powerful model should be trained; Amodei focused on whether a new AI technology should be released.
Neither announced a moratorium, a joint schedule or a measurable test that would automatically stop work. Both companies had launched models earlier in September. Those launches do not establish that either lab disregarded safety concerns, but they sharpen a practical question behind the pledges: who decides when the evidence is strong enough to proceed?
For people building on these systems, the difference between training and release is concrete. Delaying training could change which capabilities arrive next. Delaying a release could change when developers and users get access to a product that has already been built.
Altman Draws His Safety Line Before Training
Altman told the Council that competition was no excuse for taking excessive technological risks. “We have unilaterally slowed down in the past. We will do so in the future,” he said in OpenAI’s published text of his delivered remarks.
His more specific argument concerned the start of training. Companies should not train models without a strong case that they can keep those models under human control, he said. A lab would have to justify proceeding before the resulting model was available for full evaluation, rather than rely solely on checks shortly before launch.
“Strong case,” though, is not an operational threshold. Altman did not identify a training run that OpenAI would delay, specify what evidence would be sufficient, or name an outside party empowered to reject the company’s assessment. His assurance that OpenAI would slow down again was a future commitment; he did not announce that training had stopped.
Altman also warned that AI systems capable of helping improve their successors could accelerate development. By the time one model is ready to ship, a lab may already be working on the next. In that setting, the decision to begin training carries its own safety stakes.
Amodei Makes Safety a Condition of Each Release
Amodei focused on what Anthropic makes available. “We will slow down as much as necessary, in order to make sure that every successive AI technology that we release is actually safe,” he said, according to ABC News’ account of the Security Council session. The UN’s transcript of the meeting is generated by automatic speech recognition and is not an official UN record.
“Every successive AI technology” is a broad promise, though it does not require postponing every release. Amodei said Anthropic would accept a delay when needed to meet its safety judgment. He gave no fixed waiting period or public definition of what “actually safe” requires.
The pledge builds on Amodei’s earlier essay, We Must Pace the Frontier. He argued there that safety work needs time to keep up with improving model capabilities. He explicitly distinguished pacing from halting model training or technical progress, and proposed outside access to help assess both completed systems and the processes used to build them.
The CEOs described related gates at different stages. Altman’s argument asks when it is responsible to begin training a model. Amodei’s UN promise asks when Anthropic can responsibly release a technology. A company could slow at one gate and continue working at the other; launches alone cannot show whether it honored a training-stage commitment.
September Model Launches Complicate the Message
Both companies released new models in the week between earlier calls to pace frontier AI and the September 23 meeting, ABC News reported. Because the UN statements followed those launches, the launches cannot by themselves demonstrate a breach of promises made at the Council.
The sequence still deserves scrutiny. Amodei’s pacing proposal was already public, and both leaders were arguing that competitive pressure should not dictate the speed of development. Shipping a model is compatible with a conditional safety commitment if the company judges that system ready. The conditions behind that judgment remain the question.
The Council remarks do not establish whether either September release had been delayed, what evidence supported its clearance, or what finding would have led the lab to hold it back. A product launch also reveals nothing about whether work on a different, more capable training run was slowed. Those decisions raise separate safety questions.
Users and developers should not assume releases have stopped. They also should not assume a future delay would be a routine scheduling change. Under the principles the CEOs described, a lab might postpone work because its safety case is not yet adequate.
External Checks Could Make the Pledges Testable
Amodei’s essay offers one way to reduce reliance on a company’s own account. He proposed giving embedded third-party evaluators ongoing, employee-like access to frontier labs. They would examine safety practices and training pipelines, assess model behavior, and report incidents. Amodei wrote that Anthropic was committing to this access and urged governments to require comparable steps elsewhere.
With ongoing access, an evaluator could inspect how a system is trained, potentially spotting a problem before release or assessing whether a lab followed its own safety process. That is different from a one-time test of a finished model. Access and reporting, however, do not give evaluators an explicitly stated veto over Anthropic’s launches.
Altman called for international AI standards, accurate and rapid incident reporting, and secure channels for governments and companies to share safety information. Reporting can reveal failures that would otherwise remain confined to one organization. Its usefulness depends on what counts as an incident, how promptly it is disclosed and who can examine the underlying evidence.
The meeting’s opening briefer, AI researcher Yoshua Bengio, went further. In the Security Council transcript of his remarks, he called for developers to demonstrate safety to independent experts before training and deployment, alongside a shared incident-reporting mechanism and licensing for frontier AI. Those were Bengio’s proposals, not requirements adopted by the Council or commitments jointly accepted by the companies.
The Council Heard Proposals, Not Binding Rules
Speaking before the Security Council put the pledges on the public record without settling how they will be applied. Neither CEO announced a common trigger for slowing down, a duration for any delay, or an agreed procedure for resolving a dispute between a lab and an outside evaluator.
Frequently Asked Questions
4 questions
1Did Anthropic and OpenAI Announce an AI Model Pause?
No. At the September 23, 2026, UN Security Council meeting, their CEOs made conditional commitments to slow work when safety required it. Neither announced a joint moratorium, a fixed delay or a binding agreement. Both companies had released models earlier in September, but those launches alone do not show whether a safety condition for slowing work had been met.
2
Sources
- OpenAI’s published text of his delivered remarksopenai.com
- ABC News’ account of the Security Council sessionabcnews.com
- UN’s transcript of the meetingtranscripts.un.org
- We Must Pace the Frontierdarioamodei.com
