The AI labs now say they should slow down. Washington says no, and Brussels wants global rules.

Four days after Jacob Coxon’s resignation, the heads of Anthropic and OpenAI both said in public that the AI industry should slow down.

On September 12 Anthropic’s chief executive Dario Amodei published an essay of roughly 3,800 words arguing that frontier labs must slow the rate at which they improve AI capabilities. The same day Sam Altman agreed, said OpenAI would also pace the frontier, and told reporters that OpenAI would not go public in 2026 because this is an ill-advised moment to do so.

What Amodei is proposing

The essay sets out three steps, as Forbes summarised them. First, independent safety testers should work inside the labs with the same access as employees, down to desks, badges and company laptops. Second, labs in democratic countries should agree on shared safety standards and limits on how fast capabilities can advance unchecked. Third, the US and its allies should seek international agreements, eventually including China.

Anthropic committed to the first step on its own: third-party evaluators get permanent, employee-level access to its systems and safety procedures. It has not paused training. To explain the urgency, Amodei warned that coordinated AI agents could become capable of taking over the entire internet within six to twelve months.

Competitors agreeing to limit how good their products get looks a lot like a cartel, so Amodei asked Washington for a narrow antitrust waiver covering safety discussions. OpenAI, Anthropic and Google have reportedly been meeting to discuss a new standards body for testing and auditing.

The case that this is self-serving

Critics quoted by Forbes call the proposal a soft cartel: the leading labs slow down together, keep their lead, and raise the bar for open-source developers and challengers such as DeepSeek and Alibaba’s Qwen team. Anthropic is also preparing an IPO reported at around $2 trillion, and a coordinated slowdown would protect exactly the kind of lead investors pay for.

The labs can be sincerely worried and also well placed to benefit from the rules they propose. A better test of intent is whether they accept oversight they do not control. Giving outside evaluators employee-level access passes that test. A voluntary pact between three companies does not.

Washington splits, and the White House says no

Congress has produced a stack of bills. Senators John Thune, Amy Klobuchar and Ted Cruz are working on legislation requiring labs to mitigate model harms. Representatives Ted Lieu and Nathaniel Moran want an emergency shutdown capability for advanced systems, and Jay Obernolte and Lori Trahan want risk reports filed with the Commerce Department. At the far end, Bernie Sanders and Greg Casar have proposed banning superintelligence outright and pausing advanced development.

President Trump rejects the idea. Asked whether he worried about AI causing human extinction, he said he does not, and framed the real risk as losing to China, against whom he estimates the US has roughly a one-year lead. On September 14 he posted that the only guardrail AI needs is a strong and smart president. With the administration opposed, none of the congressional bills is likely to pass soon, and the slowdown stays voluntary.

Europe already has some of the machinery

Europe responded more calmly, and here its rules are already ahead of the US. Henna Virkkunen, the Commission’s executive vice-president for tech, pointed out on September 12 that EU rules already require the largest model providers to assess the risk of losing control of their models, and called for international coordination because no such requirement exists elsewhere. Germany’s digital minister went further and proposed a global AI safety agency modelled on the International Atomic Energy Agency.

Since August the AI Office can demand access to models, require risk mitigation and order a model withdrawn from the EU market. Much of what Amodei asks the industry to volunteer, independent evaluation and a regulator able to say stop, exists on paper in Europe. Whether the AI Office has the technical capacity to use those powers against a company like OpenAI is a separate question, and nobody will know until it tries.

What this means for companies using AI

For most businesses the practical effect is on timing and access. A voluntary slowdown among the top labs means the next generation of frontier models will arrive later and more gradually than the release pace of the last year suggested. It also means more gated releases of the kind Washington introduced this summer, where trusted partners get access first.

The capabilities you already have are not going away, and the cheaper tiers keep improving. If your AI plans depend on a steady stream of more capable models, build them around what exists today and treat anything new as a bonus.

Watch open-source policy closely. If the leading labs coordinate and governments start licensing frontier development, open-weight models are the obvious next target, and many European companies rely on them for data control. The EU’s position on that will matter more to Danish businesses than anything decided in Washington.

And do not dismiss the warning as marketing. The two chief executives who know most about these systems now say in public that the technology is moving too fast. Keep people responsible for consequential decisions, and write down exactly what each of your AI systems is allowed to do.