Anthropic is urging the world to hit the brakes on the most powerful artificial intelligence systems, arguing that today’s frontier models are beginning to show warning signs that they could slip beyond human control.
The San Francisco-based company, best known for its Claude family of AI models, said in a report that a coordinated global slowdown in cutting-edge development would “likely be a good thing.” But it cautioned that a unilateral pause would be futile, because competitors would simply surge ahead.
“We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology,” it said.
For any pause to be meaningful, Anthropic said, major AI developers across multiple countries—especially the US and China—would need to stop simultaneously under rules that are credible and verifiable.
“Without a global coordination mechanism, companies and governments will have to make difficult decisions about safety while under competitive and geopolitical pressures,” it said.
Anthropic’s stance has drawn criticism from parts of the AI industry and some White House officials, who argue the company leans too heavily on worst-case scenarios—warning that its messaging could double as an attempt to slow rivals under the banner of safety.
Anthropic is led by former OpenAI employee Dario Amodei
Even so, the White House has acknowledged the strength of Anthropic’s Mythos model. The system has not been released to the public because of its cybersecurity capabilities and is currently used only by a small number of vetted organizations.
Anthropic’s proposal faces steep headwinds in Washington and Silicon Valley, where many US officials and tech leaders have repeatedly argued that deliberately slowing AI progress could hand China a decisive strategic advantage in what they describe as the century’s defining technology contest.
US President Donald Trump, however, said he raised the prospect of cooperating with China on AI safety during his recent visit to Beijing.
Mr Trump also signed an executive order this week giving the government 30 days to conduct a preliminary review of the most powerful US AI models before they are released.
‘Human role narrowing’, says Anthropic
In its report, Anthropic likened the challenge to nuclear arms control—while stressing it may prove even tougher to police. AI training, the company said, is far easier to conceal than something like a missile silo, and the incentive for quiet noncompliance would be enormous.
To explore how a workable framework might be built, the company said it intends to convene government officials, scientists, advocacy groups, and competing AI firms in the coming months.
The push for international coordination comes as Anthropic points to internal data indicating that AI tools are already significantly accelerating the creation of new AI systems.
That speed-up, the company said, risks creating a feedback loop that could culminate in what researchers call “recursive self-improvement”.
The concept describes an AI system that becomes capable of essentially training itself to grow more capable, with limited human involvement.
“We are not there yet, and recursive self-improvement is not inevitable,” the report said, while warning it could arrive sooner than many governments and institutions are prepared for.
“The evidence suggests that the human role is narrowing at each step in the AI development process,” the company said.
