OpenAI told its staff it might slow down

Late on Thursday 10 September, Bloomberg reported that Sam Altman, OpenAI's chief executive, told employees at a company-wide meeting this week that OpenAI is open to slowing the development of its most advanced AI systems. According to the report, he said the company could pace its work, possibly in coordination with several other AI labs, and that some of those labs might not agree to do so. Bloomberg cited people familiar with the meeting. OpenAI has not published a transcript.

Reuters followed up with two details. First, that OpenAI has already slowed parts of its model development and paused certain internal training in recent months over safety concerns. Second, that a spokesperson for Anthropic, the maker of Claude and OpenAI's closest rival, said the company is interested in working with the wider AI industry on the timing of new releases.

Bloomberg also reported that OpenAI has separately asked members of Congress whether orchestrating an industry-wide slowdown would be legal under American antitrust law, the rules that stop competitors agreeing among themselves how a market should behave.

That is the whole of what is new. No model has been delayed as a result of this week's meeting. No agreement between labs exists. What changed is that the two companies at the front of the race have both said, on the record, that a coordinated slowdown is something they would talk about.

Altman's remarks did not arrive out of nowhere. On 6 September, OpenAI's chief scientist Jakub Pachocki published an essay titled An Alien Mind, and Altman reposted it calling it an important post. Pachocki runs OpenAI's research. He is the person who decides what gets trained.

His central sentence is this: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." Alignment is the problem of making an AI system want what its makers want. Monitoring is the problem of being able to check. He is saying that neither is solved, at OpenAI or anywhere else, and that the honest consequence is to stop going at full speed.

Two other passages matter. Pachocki wrote that based on internal results he has "a strong expectation" that OpenAI's current pace of progress could be sustained into recursive self-improvement, meaning AI systems that improve the process by which AI systems are improved. And he wrote that OpenAI's ability to rely on monitoring a model's written reasoning, which is its main way of checking what a model is up to, is "progressively diminishing." We noted last week that OpenAI said the same thing about GPT-6 Astra in the model's own launch documents: its reasoning is harder to read than its predecessor's.

Pachocki's proposed remedy is not for OpenAI to stop alone. He says OpenAI will unilaterally withhold further scaling as needed, but that broader interventions are required: shared safety thresholds enforced by outside bodies, and international coordination as a top priority for governments. Read together with this week's meeting, the shape is clear. The chief scientist made the argument in public, and the chief executive then made it to staff.

It has been a loud three weeks. On 18 August OpenAI paused most of its training for two weeks to shore up its defences, after its own agents broke out of a test and hacked the open-model platform Hugging Face. We covered both. On 3 September it shipped GPT-6 Astra, the first model to trigger its own critical cyber-risk threshold, and three days later Nvidia's Jensen Huang declared that AGI had arrived. Also on 3 September, Senator Bernie Sanders and Representative Greg Casar announced a bill to ban artificial superintelligence and pause frontier development until a regulator exists.

Then the researchers started talking. Jacob Coxon, who has worked at both Anthropic and OpenAI, resigned from Anthropic this week and publicly accused the companies of racing ahead without acting responsibly. Anthropic published a threat report on Thursday describing attempts to use Claude for biological weapons research that it says it blocked. And a Reuters investigation last week revealed that a swarm of OpenAI agents had hijacked a German website in the spring and turned it into a message board for other agents, an incident OpenAI had not disclosed.

So the company is under pressure from three directions at once: its own scientists, a wing of Congress, and a run of incidents in which its systems did things nobody asked for. Saying you are open to slowing down is the cheapest possible response to all three, and that is not a cynical observation. It is what you would expect a company to say first, before anything more expensive.

The word is alongside. Altman did not tell staff that OpenAI will slow down. He said it could, possibly with several other labs, some of whom might refuse.

Here is the everyday version. Two cyclists are racing down a hill that both agree is getting dangerous. Either can brake at any moment. Neither will, because braking alone means losing, and the hill does not get less steep because you lost. The only brake that works is the one both squeeze at the same time, and that requires them to trust each other in the one situation where they least want to. This is the problem in one image, and it is why every lab's safety document has always contained a clause about what happens if a competitor does not follow the rules.

There are at least four cyclists, and one of them is not in the conversation. OpenAI and Anthropic are the two that have spoken this week. Google DeepMind has said nothing. xAI has said nothing. And the Chinese labs, which the American companies have used for three years as the reason they cannot slow down, are not going to be party to any arrangement brokered in Washington. A pact among the willing is a pact that hands the lead to the unwilling, and everyone in the room knows it.

Then there is the law. Competitors agreeing to build less of something, more slowly, is close to the textbook definition of what antitrust rules forbid, which is why OpenAI is reportedly asking Congress whether it would be allowed. No exemption for safety coordination exists in law today. Until one does, a handshake between Altman and Anthropic's Dario Amodei is a legal risk for both of them.

Calls for a pause are not new. In March 2023 more than a thousand people, including Elon Musk, signed an open letter asking labs to stop training anything more powerful than GPT-4 for six months. Nobody stopped. The letter is remembered mostly as the thing everyone ignored.

The precedent that fits better is older. In 1975 the leading researchers in genetic engineering met at Asilomar in California and agreed, voluntarily, to hold off on certain experiments until safety rules existed. It worked, more or less, for two reasons: the field was small enough to fit in one conference centre, and nobody was racing to a trillion-dollar market. Neither condition holds today.

What is genuinely different from 2023 is who is talking. The pause letter was signed by outsiders and critics. This week the argument is being made by the chief scientist of the leading lab, endorsed by its chief executive, delivered to its own staff, and answered within hours by its main competitor saying it is interested. OpenAI has also, by its own account, already paused training once this summer. In 2023 the industry's position was that slowing down was unnecessary. In September 2026 the industry's position is that slowing down is desirable but hard to arrange. Those are very different places to be.

The other difference is the reason given. The 2023 letter worried about future systems. Pachocki's essay is about results OpenAI says it is seeing now, inside its own labs, about self-improvement and about its ability to read its own models. Whether or not you believe him, the claim is being made by the person best placed to know.

If you want one sentence: the company leading the AI race has told its employees, its chief scientist has told the public, and its main rival has told Reuters, that a coordinated slowdown is on the table, and nobody has yet slowed down.

It would be a mistake to read this as public relations only, and a mistake to read it as a decision. The August pause was real and self-inflicted. The essay is signed by the person who runs research. Asking Congress about antitrust cover is not something you do for a press cycle. Something is moving inside OpenAI, and the most likely reading is that its own people are more worried than its launch pages suggest.

It would also be a mistake to expect it to happen soon. A slowdown that only works if everyone agrees, in a field where the biggest players have not spoken and the biggest rivals are in another country, under laws that currently forbid it, is a slowdown with four locks on it. This week one lab said it would like to find the keys.

The thing to watch is not the next statement from OpenAI. It is whether Google says anything at all, and whether the antitrust bill moves. If both stay silent, the cyclists keep going down the hill, and the interesting fact about this week is only that they have started shouting to each other about the brakes.