Artificial intelligence company Anthropic has proposed a coordinated global pause on the development of the world’s most advanced AI systems, warning that emerging models are beginning to exhibit capabilities that could potentially exceed human control.
The San Francisco-based company, known for developing the Claude family of AI models, made the recommendation in a report released on Thursday. It argued that a temporary slowdown in frontier AI development would provide governments, researchers, and institutions with the time needed to strengthen safety measures and align technological progress with societal interests.
According to the report, a meaningful pause would require cooperation among leading AI developers and governments worldwide, particularly in the United States and China. Anthropic stressed that any such arrangement would need clear verification mechanisms to ensure all participants comply with agreed-upon restrictions.
“We believe it would be beneficial for the world to have the ability to slow or temporarily halt frontier AI development so that societal structures and alignment research can keep pace with technological advancement,” the company stated.
Anthropic acknowledged that achieving international cooperation on AI development would be difficult, especially in a highly competitive industry where companies and nations are racing to secure technological leadership.
The company noted that, without a global coordination framework, governments and technology firms would continue to face difficult decisions regarding AI safety while balancing economic competition and geopolitical interests.
Anthropic’s position has drawn criticism from some industry leaders and policymakers who argue that concerns about extreme AI risks are exaggerated and may serve as an attempt to slow competitors. However, the growing capabilities of advanced AI systems have continued to fuel discussions about regulation and oversight.
The debate comes as concerns intensify over increasingly powerful AI models. Anthropic pointed to evidence suggesting that AI systems are already helping accelerate the development of future AI technologies, creating what researchers describe as a feedback loop.
The company warned that continued progress could eventually lead to “recursive self-improvement,” a scenario in which AI systems become capable of significantly improving their own intelligence with minimal human intervention.
While Anthropic emphasized that such a development has not yet occurred and is not inevitable, it cautioned that the possibility could emerge sooner than many governments and institutions are prepared for.
Anthropic co-founder Jack Clark compared the current state of AI development to a vehicle operating with an accelerator but lacking an effective braking system.
“You want the option to be able to take your foot off the gas and put your foot on the brake,” Clark said during an interview, arguing that the industry currently lacks sufficient mechanisms to slow development when necessary.
To address these concerns, Anthropic announced plans to engage policymakers, scientists, advocacy groups, and rival AI companies in discussions aimed at designing practical systems for managing future AI risks.
The company also compared the challenge to nuclear arms control efforts but noted that regulating AI could prove even more difficult because AI development is easier to conceal and monitor than traditional military programs.
Despite the proposal, securing broad international support may be challenging. Many policymakers and technology executives continue to argue that slowing AI development could allow rival nations to gain a strategic advantage in what is increasingly viewed as one of the most important technological competitions of the modern era.
Anthropic concluded that current trends indicate human involvement is gradually decreasing at various stages of AI development, highlighting the need for stronger safeguards as the technology continues to advance.
Anthropic Calls for Global Pause on Advanced AI Development Amid Safety Concerns
