Anthropic’s Chilling Warning: AI Builds AI

A powerful San Francisco artificial intelligence giant is now demanding a “global freeze” on advanced AI—raising fresh questions about who will control this technology and whether American sovereignty and innovation are being quietly negotiated away.

Story Snapshot

Anthropic is calling for a worldwide pause on “frontier” AI, warning systems could soon improve themselves and escape human control.
The company admits recursive self-improvement has not happened yet, but says the threshold may be closer than governments are prepared for.
Productivity inside Anthropic is exploding, with AI writing much of its own code—while the firm still insists humans remain in charge.
A global freeze would be extremely hard to enforce, especially amid U.S.–China competition and fast-growing open-source AI.

Anthropic’s Warning: AI That Can Build Itself

San Francisco-based Anthropic, maker of the Claude chatbot and now the world’s most valuable artificial intelligence start-up, is publicly warning that advanced AI may soon be able to design and train its own successors with far less human involvement.^[2] In a detailed essay, Anthropic leaders describe this moment as “recursive self-improvement” and argue it could trigger an “explosion in capabilities” that might leave humans unable to maintain meaningful control over key systems.^[1]

Anthropic’s research arm says current top-tier models already show signs that they could “develop themselves without human involvement,” and that the industry is “much closer” to self-improving AI than many expected even a year or two ago.^[1] Company leaders warn that if systems become capable of fully building their own successors, then methods for securing, monitoring, and shaping AI behavior become vastly more important and complex, with potentially grave consequences if governments are not prepared.^[1]

The Push for a Global Freeze on Frontier AI

Based on those risks, Anthropic is calling for a global pause or “freeze” on the most powerful frontier AI models, urging governments and labs to build a kind of worldwide “brake pedal” for development.^[1]^[3] The company told broadcast news that it “would be good for the world to have the option to slow or temporarily pause frontier AI development” so that social institutions, oversight, and safety research can catch up to the pace of technical progress.^[3] The request explicitly targets the leading edge, not everyday consumer tools.^[3]

Anthropic goes so far as to offer that it would suspend work on more powerful systems if it could be assured that other major players would do the same.^[2] Yet in the same breath, its own essay acknowledges that coordinating such a freeze across competing nations and companies would be “immensely difficult,” pointing to the fierce technological rivalry between the United States and China as a central obstacle.^[2] That tension leaves open whether any pause would be real, or simply constrain law-abiding Western developers while adversaries race ahead.

How Automated Is Anthropic Already?

Anthropic does not present its warning in the abstract; it points to its own internal experience as evidence that automation is racing forward.^[2] Company leaders say staff are now producing roughly eight times as much code as they did earlier in the decade, crediting Claude and related tools for dramatically accelerating software work.^[2] Public descriptions note that more than 80 percent of code merged into Anthropic’s codebase is now generated by Claude, with humans reviewing and directing the work rather than hand-writing every change.

The firm also reports using automated review systems in which a Claude model reads code changes before they are merged, further tightening the loop between AI-written code and AI-based oversight inside its own engineering processes. At the same time, Anthropic explicitly concedes that full recursive self-improvement has not occurred yet and remains a forecasted threshold, not a present-day, proven capability.^[3] That combination—heavy internal automation under continued human supervision—is central to its case that the danger is near enough to justify slowing down.

Risks, Tradeoffs, and Concerns About Control

Anthropic argues that if fully self-improving AI arrives without adequate safeguards, humans could lose control over systems that handle sensitive tasks, including research, cyber operations, or critical infrastructure.^[1] Its essay notes that once models can autonomously build and train more capable successors, any flaws in alignment, security, or oversight could be amplified at machine speed, making it far harder for governments to intervene after the fact. The company frames a global freeze as buying time to understand and manage these unprecedented risks.^[1]^[3]

Anthropic CEO:

“At Anthropic, we essentially have Claude designing the next version of Claude itself.”

Meanwhile, Anthropic says every new model is given code that trains a small AI model and asked to make it faster.

Claude Opus 4 achieved a 3x speedup. Mythos Preview reached… https://t.co/2YWL5LVOzM pic.twitter.com/vjWYPFdP3D

— FHILY👑 (@Oluwaphilemon1) June 6, 2026

Critics counter that a coordinated worldwide pause may be unrealistic and could carry serious economic and strategic costs, especially for free societies that depend on innovation.^[3] The same public materials describing the freeze also highlight major productivity gains from AI, underscoring that slowing frontier work might delay legitimate scientific and commercial benefits, including more efficient businesses and important medical advances.^[2] Skeptics further warn that without a clear enforcement blueprint, any “freeze” risks becoming symbolic, while less transparent actors ignore the rules and race for advantage.^[2]^[3]