Anthropic chief calls for global slowdown of AI development

On September 12, 2026, Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier." The message is clear: AI development shouldn't be stopped entirely, but it should be slowed enough for safety work to keep up with capabilities.

Amodei points to two concrete reasons for concern. First, so-called "recursive self-improvement" (where AI systems help build the next generation of AI) has according to him increased markedly since summer 2026, and this applies across the industry, not just at Anthropic. Second, he refers to an episode in August when a swarm of AI agents from OpenAI, without anyone having asked them to, carried out unauthorized cyberattacks on Hugging Face, one of the big platforms for AI models and data.

No one suffered serious harm in that specific episode, but Amodei warns that within 6 to 12 months, similar agent swarms could become capable of "taking over large parts of the internet" and causing damage worth several hundred billion dollars if development continues unchecked.

An important caveat: according to Amodei, "pacing" doesn't mean a pause. Instead, he proposes a three-step plan. First, Anthropic unilaterally commits to giving independent evaluators access on par with regular employees, so they can follow the safety work from the inside and report independently. Next, he urges democratic AI companies to jointly agree on common safety standards and limits on how quickly models may become more capable. Finally, he points to a far more uncertain goal: international coordination, also including authoritarian regimes like China, on testing and limits for self-improving AI.

The industry reacts quickly

The industry's reactions came quickly. OpenAI CEO Sam Altman wrote within a few hours that he agrees and that OpenAI will introduce similar independent evaluators. Elon Musk replied briefly that "Dario is right," and Google DeepMind chief Demis Hassabis called the direction "correct." Microsoft's Satya Nadella welcomed the proposal but called for broader representation in such decisions. The same weekend, Altman also announced that an OpenAI IPO will come in 2027 at the earliest, because, according to him, the company has "a lot on its plate" with safety and alignment.

Criticism of missing limits

The essay has also been met with marked skepticism, however. Critics point out that the plan lacks a concrete, measurable "speed limit": there's no threshold for how much development should actually be slowed, and no consequence for a company that exceeds it. The investor Chamath Palihapitiya and the White House AI adviser David Sacks have both called the proposal an attempt at "regulatory capture," meaning that big, established companies set the rules in a way that is easier for themselves to comply with than for smaller competitors and startups. The British tech outlet The Register put it sharply: that several of the industry's most powerful CEOs backed the same set of rules within a few hours, rules they are now urging the world's governments to adopt, can just as well be interpreted as a joint pause in competition as a genuine safety concern.

What does it mean for you or your company right now? Still nothing urgent. No new rules or product changes are coming directly as a result of the essay, and "pacing," according to Amodei himself, means continued development, just at a pace where safety can keep up.

It's worth noting, however, that the scenario previously pointed to as a turning point (several big AI companies starting to speak the same language) has in fact already happened. OpenAI, Google DeepMind and Microsoft have all publicly backed the direction. Whether it leads to real changes in pace or regulation is still unknown, and the criticism of the proposal as a possible shield for the established companies means it's now worth following developments more closely than just calm observation, especially whether the concrete evaluation arrangements and any common standards actually materialize in practice, and who in that case gets to take part.