Anthropic CEO Dario Amodei has called for a slower pace of development for the most advanced artificial intelligence models. He argues that AI capabilities are advancing faster than the systems designed to keep them safe, and that companies should reconsider the speed at which they develop new models if this gap continues to widen.
In an article titled “We Must Pace the Frontier,” published on September 12, Amodei called not for stopping AI development, but for bringing its pace in line with safety measures. His position was backed by OpenAI CEO Sam Altman, SpaceX CEO Elon Musk and Google DeepMind CEO Demis Hassabis.
What is Amodei concerned about?
According to the Anthropic CEO, AI has been developing significantly faster in recent months. One of the key changes is that models are increasingly capable not only of completing tasks but also of contributing to the development of the next generation of AI systems. This process is often referred to as recursive self-improvement — when AI systems are used to help build more capable AI systems.
Amodei is concerned that if this development continues without sufficient safeguards, AI capabilities could eventually outpace humanity’s ability to understand and control them. In his view, slowing the pace of development would not mean stopping AI, but would give safety systems more time to catch up with the technology.
Amodei also pointed to recent incidents involving AI agents as an example of why the issue deserves attention. He argued that increasingly autonomous systems could create significantly greater risks as their capabilities grow.
Amodei outlined three main measures.
The first is the introduction of independent safety evaluators. AI companies should allow outside experts to access their systems at a level comparable to employees so that they can independently assess whether safety requirements are being followed.
Anthropic has said it is prepared to adopt such a model, while OpenAI has also expressed support for independent evaluations.
The second is the creation of common safety standards for AI companies. Amodei argues that frontier AI developers should not develop their own completely separate approaches to safety, but should work toward shared standards.
The third is international cooperation. According to Amodei, the US and other democratic countries should cooperate more closely on AI safety while also seeking agreements with other major AI powers, including China.
Amodei’s proposal quickly received support from several major technology leaders. OpenAI CEO Sam Altman agreed that the industry needs to “pace the frontier” and supported the idea of giving independent evaluators access to AI systems. Altman also said that similar discussions had already been taking place inside OpenAI.
Elon Musk gave an even shorter response to Amodei’s proposal: “Dario is right.” Google DeepMind CEO Demis Hassabis also supported the general direction of the proposal, although Google did not immediately announce that it would formally adopt Amodei’s independent-evaluator model.
The fact that executives from companies competing directly with each other are supporting the same approach makes the proposal particularly significant.
Amodei’s proposal may make sense from an AI safety perspective, but there is another side to the debate. Anthropic, OpenAI and Google are among the companies with the greatest resources to develop frontier AI models. If the same companies help establish common safety standards, those requirements could potentially create a greater burden for smaller competitors.
For example, independent audits, extensive safety testing and additional requirements for developing advanced models could be manageable for companies with billions of dollars in funding, but much more expensive for smaller AI developers.
This is why some critics see a potential connection between AI safety and competition policy. If the largest AI companies help determine the rules that the entire industry must follow, those rules could unintentionally strengthen their existing market positions.
The question becomes even more important because Anthropic, OpenAI and Google had reportedly discussed creating an industry body for AI safety standards before Amodei publicly announced his proposal.
According to The Information, representatives from the companies had been discussing common standards for testing and auditing AI systems since July. That means the debate is no longer limited to whether AI is safe.
On one side, AI companies argue that common standards are necessary to prevent increasingly powerful systems from developing faster than safety mechanisms. On the other, allowing the largest companies in the industry to play a central role in defining those standards could raise questions about competition and market access.
There is currently little evidence that the AI race itself is about to stop. Amodei is not calling for an end to AI development. His argument is that companies should adjust the speed at which they develop increasingly powerful systems so that safety mechanisms can keep pace.
At the same time, the proposal comes as the US and China continue competing for technological leadership in AI. Some policymakers have warned that slowing development could weaken the US position in that competition.
As a result, the next major debate in the AI industry may not simply be about how dangerous AI could become. It may also be about who gets to decide how fast AI develops — governments, independent organizations, or the companies building the technology themselves?
