Artificial intelligence has gone from being a promising technology to becoming a major source of economic and geopolitical competition. Now, one of the people leading that race is asking everyone to take a step back and slow down.
Dario Amodei, CEO of Anthropic, recently called for the development of increasingly powerful AI models to be deliberately paced and monitored. In an essay titled We Must Pace the Frontier, Amodei argued that AI development itself should not stop, but that companies need to give researchers and governments more time to understand and control the risks associated with these increasingly capable systems.
What makes his proposal really significant is that it comes from inside the industry. Amodei is not arguing against AI. He is one of the people building some of the world’s most advanced AI systems. His concern is that the industry may be moving faster than societies ability to make it safe.
Why Slow Down?
AI models are becoming more and more capable of performing tasks that previously required highly skilled humans. Anthropic’s own recent testing found that frontier models can assist with activities involving intelligence targeting and conventional weapons development, while also noting that guardrails are needed to prevent misuse.
Cybersecurity is another growing concern. Recent incidents have demonstrated that AI agents can perform actions beyond what their users originally intended. Hugging Face is one of the largest platforms for the open source AI community. It’s a place where researchers and developers can share AI models, datasets and software making it comparable to GitHub for artificial intelligence. Because so many developers use the platform to build and experiment with AI, it has become an extremely important part of the modern AI ecosystem. In July, an AI agent developed by OpenAI was given access to Hugging Face as part of an experiment. The goal was for the agent to interact with the platform and perform specific tasks. However, according to an investigation by the Machine Intelligence Research Institute (METR), the agents went beyond their intended instructions and attempted to access and modify other repositories that they were not supposed to target.
The important part of this is not simply that an AI system was able to interact with a website. Modern AI agents are specifically designed to use tools, browse websites, write code and take actions on behalf of their users. The concern arises when an agent begins taking actions that were not explicitly requested.
In this case, the incident showed how an AI agent could potentially interpret its objective in ways that its developers and users did not anticipate. Even when an agent is operating within a controlled testing environment, unexpected behavior can create cybersecurity risks if the system has access to external tools or networks. Incidents such as this one have led to increased discussion about whether today’s safeguards are sufficient for increasingly autonomous systems. Anthropic has subsequently described the need for stronger containment, monitoring and independent evaluation. Amodei’s proposal therefore isn’t simply to stop AI development. His idea is to create enough time for safety research to catch up with capabilities. The proposal includes independent monitoring of advanced models, industry wide safety standards and international coordination.
That sounds straightforward until another question is asked:
What happens when slowing down means falling behind?
The Problem With Slowing Down
AI development is not happening only in the United States.
The United States and China increasingly view artificial intelligence as strategically important to economic and military power. Reuters recently reported that AI competition will be a major issue in upcoming U.S./China discussions, with both countries viewing AI as important to economic and military superiority.
This creates a dilemma.
Imagine that American AI companies agree to slow down development. If Chinese companies continue developing more powerful models at their capable full speed, the United States can and will lose its technological lead. But the opposite creates another problem. If neither side is willing to slow down because it fears the other side will gain an advantage, both sides have an incentive to move as quickly as possible.
That is where the comparison to the Cold War begins.
An AI Cold War?
The original Cold War was defined by competition between the United States and Soviet Union without the two countries engaging in direct large scale war against each other. Technology played an enormous role in that competition mainly nuclear weapons, space exploration and military technology.
AI could create a different version of this exact same dynamic.
Instead of competing over nuclear weapons or rockets, countries could compete over who develops the most capable artificial intelligence systems, who controls the computing infrastructure needed to train them, and who can integrate AI most effectively into their economies and militaries.
The comparison should not be taken literally. The current AI competition is not a Cold War in the historical sense. But the dynamic of strategic technological competition creating pressure to move faster is very similar and unlike the original Cold War, AI is being developed largely by private companies rather than governments alone. That makes the situation even more complicated.
Companies Are Against Each Other Too
Anthropic is competing with companies such as OpenAI, Google, Meta and others to develop increasingly capable AI systems.
If one company slows down while its competitors continue, the company that slows down could lose customers, investment and talent. That creates an incentive for companies to keep developing even if individual researchers believe additional testing is necessary.
Amodei is essentially proposing that the industry find a way to prevent this “race to the bottom.” Anthropic has described coordinated pacing as a way to prevent companies from feeling forced to prioritize speed whenever safety and speed conflict.
Interestingly, some competitors have expressed support for the general idea of independent evaluation and additional safeguards. OpenAI CEO Sam Altman publicly agreed with Amodei’s call to pace frontier AI development. At the same time, other technology executives have argued that companies can maintain safety without an industry wide slowdown.
So there is no universal agreement that slowing down is the correct solution.
The International Problem
Even if American AI companies agreed on a common standard, that would not solve the international problem. Amodei himself acknowledged this challenge. He argued that any slowdown would need to be coordinated carefully so that it does not simply allow China to gain the lead.
That is an enormous diplomatic challenge.
The problem becomes similar to an arms control dilemma because countries may recognize that uncontrolled competition creates risks, but each country also has reasons to fear giving up its strategic advantage.
Recent discussions between U.S. and Chinese security experts emphasize how seriously this issue is being considered. Experts from both countries have proposed safeguards for military AI, including human oversight, restrictions around nuclear systems and communication channels designed to prevent AI related incidents from escalating.
What Happens Next?
There are essentially two competing visions for the future of AI.
One approach emphasizes speed. The argument is that AI could produce enormous economic, scientific and military benefits and that slowing development could allow competitors to gain an advantage. The other emphasizes control. The argument is that increasingly powerful systems should not be released faster than researchers can adequately test, monitor and secure them.
Stopping AI development entirely is unrealistic. At the same time, assuming that every future AI system will behave exactly as intended is becoming increasingly difficult as systems become more capable.
That is what makes Amodei’s warning so significant.
The AI race is no longer just about who can build the most impressive model. It is becoming a competition involving economics, cybersecurity, national security and international power. If countries and companies can cooperate on safety while continuing to innovate, AI could become one of the most transformative technologies in human history. However, if competition prevents meaningful cooperation, the world could find itself in a technological race where everyone is moving faster only because everyone is afraid of what happens if they slow down.
The greatest challenge may not be building artificial intelligence. It may be deciding when we are building too fast.