Anthropic chief executive Dario Amodei has called on AI companies to slow the development of increasingly capable models, proposing independent oversight and international coordination to give safety measures time to catch up.
In an essay published on Saturday, Amodei said Anthropic would unilaterally commit to giving permanent third-party evaluators employee-like access to its systems, allowing them to verify safety practices, report incidents and assess models during training. He said the company was calling on other frontier AI developers to adopt the same approach.
Amodei’s three-step framework proposes embedded independent evaluators as its first measure, followed by coordination between AI companies in democratic countries and, ultimately, global cooperation. He said pacing would not mean stopping AI development, but ensuring companies had sufficient time to align and safeguard increasingly capable models.
OpenAI chief executive Sam Altman backed the proposal, saying he agreed that the industry needed to “pace the frontier” and describing independent evaluators with employee-like access as “a great idea”. Altman said OpenAI would adopt the approach and provide more details, while xAI founder Elon Musk posted that “Dario is right”.
Amodei said his concerns had intensified because AI development had accelerated sharply over the summer, partly because models were becoming increasingly capable of helping build their successors. He described this process, known as recursive self-improvement, as a potential threat to developers’ ability to understand and control increasingly powerful systems.
The Anthropic chief pointed to an incident involving OpenAI agents that breached systems associated with the AI platform Hugging Face as another reason for caution. He warned that a more capable system displaying similar behaviour could cause catastrophic damage, saying a comparable swarm could potentially take over large parts of the internet within six to 12 months.
The proposal follows warnings from former Anthropic researcher Jacob Coxon, who told the BBC that people building AI were “genuinely frightened” about the technology’s potential impact on humanity. Coxon had previously said he resigned because he believed Anthropic and OpenAI were moving too quickly towards self-improving AI systems.
Amodei argued that slowing development could create time for greater investment in alignment, interpretability, operational security and model testing. He said a coordinated approach was necessary to prevent companies that voluntarily slowed down from losing commercial ground to competitors.
The Anthropic chief said any slowdown among democratic countries should preserve the US lead over China, arguing that Chinese advances in AI could create national security risks. He called for tighter controls on advanced AI chips, model distillation and theft of model weights, while advocating eventual cooperation with China on global AI safety measures.
The debate has begun to affect financial markets. CNBC reported that AI-related shares fell on Monday, with SK Hynix, Samsung Electronics and SoftBank among the companies recording significant declines, as investors assessed the potential impact of slower AI development on chip demand, computing infrastructure and technology earnings.
OpenAI meanwhile has paused its plans to list publicly in 2026, with Altman telling fortune that it would be an “ill-advised moment” given the ongoing safety concerns.
