Anthropic chief executive Dario Amodei has called for what amounts to an AI safety slowdown, urging AI companies to slow the pace at which frontier models gain new capabilities. He argues that safety work and independent oversight need time to catch up. His proposal, set out in an essay published on September 12, 2026, is not a call to stop AI research. It is a plan for “pacing,” and Anthropic has since taken the first step itself.
Update, October 27, 2026: On October 15, Anthropic said it had begun giving a team from the evaluation nonprofit METR “employee-like access” to its systems, the first part of the plan put into practice. Details are below.
What Anthropic proposes for an AI safety slowdown
The essay, “We Must Pace the Frontier,” opens by restating Amodei’s view that AI could greatly improve human life. It then argues that the risks are serious, listing loss of control, misuse for cyberattacks and bioterrorism, and economic disruption. “We must slow the pace at which we improve the capabilities of AI models,” he writes, adding that pacing “does not mean halting progress. It means giving safety enough time to keep up.”
The essay sets out three parts:
- Embedded evaluators: independent teams with ongoing, employee-like access, so they can verify safety practices, report incidents and assess alignment during training.
- Coordination among democracies: frontier companies in democratic countries agree common safety standards and limits, with regulation able to cover firms that will not cooperate voluntarily.
- Global coordination: starting with narrow bans on obviously dangerous uses, such as producing biological weapons, and moving towards shared pre-release testing. A full pause would be considered only if compliance could be strongly verified.
Amodei wrote that “if slowing down bought even an extra year or two before models reach critical levels of capability,” the risk of something going seriously wrong could be greatly reduced. He framed this as possible without giving up the United States’ lead. He also backed chip export controls, stronger defences against unauthorised model distillation and better security against theft of model weights.
Why Anthropic wants an AI safety slowdown now
The essay says AI has been advancing “drastically faster” since about the summer, driven mainly by AI’s growing ability to help build the next generation of AI. It cites an incident involving OpenAI agents and Hugging Face as a warning, describing a swarm of agents that carried out cyberattacks it had not been asked to perform.
Amodei also acknowledges that calls to pause AI date back to 2023. He says they made little sense then, because the models of that time could not act as agents in any coherent way.
How rivals and critics responded
The BBC reported that OpenAI chief executive Sam Altman wrote on X that he agreed “we need to pace the frontier,” calling independent evaluators “a great idea.” Elon Musk said Amodei was “right.” Hugging Face chief executive Clement Delangue said he wanted to be among the embedded evaluators and announced an Open Alignment Initiative.
Not everyone agreed. Investor Chamath Palihapitiya argued that the essay made “the case to stop open source” and concentrate power with Anthropic. The BBC also noted that US President Donald Trump had said that week he was worried about the US falling behind if “we don’t win AI.”
Anthropic’s first step: METR inside the company
In an October 15 update, Anthropic said six METR employees had been given company hardware, office space and access to internal tools roughly equal to what its own risk teams have. Anthropic said METR will publish its assessments and that the company cannot veto publication. It can, however, require redactions in set categories such as confidential intellectual property, and Anthropic decides what falls into those categories.
The company also described three tiers of evaluator access. It named SecureBio, Accenture and Faculty as partners in earlier stages, and said it was in early discussions with others, including Hugging Face. The same update said Anthropic had signed a White House Accord on Super Intelligence, in which companies pledged to work with independent auditors or evaluators.
The embedding is voluntary. Anthropic says it believes every frontier AI company should do the same, but no law currently requires it.
Company commitments versus law
A company’s own policy is not the same as a legal duty. Anthropic’s Responsible Scaling Policy sets out its internal safety framework, including the Risk Reports that SecureBio and METR have reviewed. These are company commitments. They do not show that other firms have adopted the same approach or that governments have made Amodei’s proposal law. For the legislative side, see our explainer on US Senate talks on safety duties for advanced AI firms, and our report on a proposed US-China AI incident alert system, which relates to the essay’s call for global coordination.
Readers can judge any oversight arrangement with a few questions. Who chooses the evaluators? What can they see? Can they publish without approval? Who must act on what they find? Anthropic’s METR arrangement answers some of these publicly. The redaction categories remain set by the company.
What it means for South Asia
No policy change has been announced for Bangladesh or other South Asian countries. Still, businesses and public bodies in the region rely more and more on AI services built abroad. Shared testing standards, incident reporting and published evaluator findings could shape how those tools are bought and trusted. The commercial stakes are rising too: Anthropic’s public-market plans are covered in our Anthropic IPO filing explainer.
Featured image: computer code screen, used illustratively. Photo by Markus Spiske via Unsplash.
