What are we talking about?
The authors of the article (company co-founder Jack Clark and head of the research department Marina Favaro) believe that AI systems are progressing too quickly and have approached the stage of recursive self-improvement – the ability to develop without human intervention. This risks “the risk of losing human control over AI systems.”
Anthropic does not know what a world where AI can improve itself will look like: “it is difficult to predict what the economy will look like if human labor ceases to be competitive.”
According to the authors of the article, a pause in development is needed to establish regulatory rules and give society time to adapt to changes. However, suspension only makes sense if major players are involved and verification mechanisms are in place. A global control system is needed to avoid the risk of hidden training of new models. Anthropic plans to discuss this idea with authorities and market participants.
Voluntary restrictions are unlikely, Anthropic admits. “Hiddenly training a new model is much easier than building a missile silo,” and any participant in the race who did not want to pause would have a chance to get ahead. However, the industry needs a framework, the authors say: “Right now it looks like the industry has a gas pedal, but no brake pedal.”
“The world has already created verification systems for other complex technologies – for example, the Intermediate-Range Nuclear Forces Treaty. But it took decades to create infrastructure and trust for these systems. We don’t have that much time,” the article says.
What's the catch?
The Daily Telegraph writes that Anthropic is currently one of the most valuable companies in the AI field (about $965 billion). Against this background, the call to “freeze” the race looks ambiguous: critics believe that this could help the company consolidate its achieved leadership and harm its competitors.
According to Forbes, Claude has grown from 1% of global AI app downloads in each quarter of last year to 14% in the second quarter of 2026. ChatGPT fell from 67% to 47% over the same period.
Anthropic is now preparing for an IPO and continues to develop its products. However, the company has delayed the public release of its most powerful model, the Mythos, over concerns that it could be used to launch devastating cyberattacks.
This is not the first warning
Back in 2023, many prominent experts signed an open letter calling for all AI training research to be suspended and safety protocols developed. Among the signatories were Tesla and SpaceX CEO Elon Musk, Apple co-founder Steve Wozniak, Getty Images photo agency CEO Craig Peters, and Israeli historian Yuval Noah Harari.
They warned that AI could (among other things) endanger jobs, spread propaganda and lead to a “loss of control over our civilization.”
Critics of the idea argued that the pause would turn into “resistance to progress” and stop any innovation, and the agreement of American companies to it would give an advantage to other countries, primarily China.
In 2026, speaking at Oxford University, Clark compared society's lack of preparation for the introduction of AI to the slow response during the COVID-19 pandemic.
“This is a technology that we do not fully understand, and it is possible to imagine plausible scenarios in which AI could destroy every person on the planet. To think that creating this technology is safe would be a manifestation of arrogance or insanity,” he said.
The main threat is still people, not machines
When considering possible futures, one of Anthropic's most likely scenarios is one in which AI development becomes largely automated, but humans continue to direct research and evaluate the results. Organizations that use AI systems will become much more efficient over time.
“This can be used for harmful purposes, from authoritarian surveillance of entire populations to influence operations that tailor manipulation to each individual and are carried out on a scale unattainable by any team of people,” Anthropic warns.
The company has reason to be wary. The media described her conflict with the Pentagon in March 2026 as an “unprecedented scandal.”
Previously, the US military used Anthropic's Claude model in planning US strikes on Iran and in the capture of Venezuelan President Nicolas Maduro. Observers saw a certain irony in this – it was the company that had long positioned itself as the leader of “safe AI” that turned out to be the first AI developer whose model “lit up” in a real combat operation.
However, the Pentagon demanded the unimpeded use of artificial intelligence systems, and Anthropic refused to lift restrictions on mass surveillance of Americans and the creation of fully autonomous weapons that operate without human participation.
The department then terminated a potential $200 million contract with Anthropic, declaring it a “supply chain threat” (a designation traditionally used only against foreign adversaries). On the same day, the contract was intercepted by OpenAI.
“American military personnel will never be held hostage to the ideological whims of big tech companies,” Secretary of War Pete Hegseth said.
Now Anthropic is suing the Pentagon – but at the same time, according to the FT, helping the US National Security Agency deploy the Mythos model for offensive cyber operations (possibly in China or Iran).
The company explains its actions simply: if you do not use AI for cyber attacks proactively, enemies will find a way to do it first.
