Anthropic Warns AI Nears Recursive Self-Improvement

Anthropic, the maker of the Claude chatbot, has claimed AI systems may be on the cusp of recursive self-improvement, the point at which they can design and build their own successors with little human input. In a June 4 blog post entitled “When AI Builds Itself,” the company said this could increase the risk of humans losing control of the technology and called for consideration of a coordinated slowdown in frontier AI development.

AI self-improvement and recursive development concept with advanced robot and human operators in futuristic control center...

At Anthropic, the company is delegating a growing share of AI development to AI systems themselves, which is speeding up their work, according to the research report. Anthropic says Claude now writes over 80% of its code. New data from the company suggests that frontier models have accelerated coding, debugging and research, likely creating a feedback loop in which AI systems create even more sophisticated successors.

What Is Recursive Self-Improvement

Recursive self-improvement, or RSI, is the scenario in which an AI system becomes capable of autonomously designing and training its own successor, that is, a model more powerful than itself, triggering a cycle of improvements that feeds itself and accelerates without a human having to intervene at every step. The process refers to when an AI system contributes to improving AI systems, including potentially itself, in a way that compounds over iterations, where each improvement can make the next improvement easier or more effective.

Anthropic states that we are not there yet, and recursive self-improvement is not inevitable, but it could come sooner than most institutions are prepared for. Jack Clark from Anthropic said in an interview that AI progress is going to speed up in coming years rather than stay the same or diminish.

Internal Research Shows AI Matching Human Judgment

Beyond simply executing tasks faster, Anthropic’s research examines whether AI can make better strategic decisions than human researchers. Anthropic took 129 internal research sessions and asked an automatic judge which next move was better: the human’s or the model’s. The best model as of November 2025, Opus 4.5, beat the human choice 51% of the time. The range goes from 22% for Claude Haiku 3 to 64% for Mythos Preview.

The conceptual leap of recursive self-improvement concerns judgment, not just “the model does what I tell it faster,” but “the model decides what to do better than I do.”

Three Future Scenarios

Anthropic focuses significant attention on Scenario 2 as the most near-term challenge requiring preparation now. The scenarios range from AI-assisted development where humans remain in control, to semi-autonomous systems that can run portions of the development cycle independently, to fully autonomous recursive self-improvement, where AI could run the full development cycle with minimal human involvement.

AI labs currently use AI tools to assist with AI research, writing code, summarizing papers, and running experiments, which represents a mild form of AI-assisted self-improvement. The question facing the industry is how quickly human oversight requirements will shrink in the development process.

Calls for Coordinated Action

Anthropic said it believes 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. The company plans to engage lawmakers about recursive self-improvement in the coming months.

Jack Clark said the company’s goal is to socialize the concept and give people a sense of what’s coming. OpenAI has published its own concerns and findings about recursive self-improvement as well, describing it in a December 2025 blog as a potentially dangerous phenomenon if researchers don’t share information about it.

Skepticism and Business Context

The announcement has drawn criticism from some researchers who see it as potentially self-serving. Two months ago Anthropic unveiled a model called Mythos that it declined to release publicly, saying the model was too good at finding software vulnerabilities. The call for a pause also came just days after Anthropic confidentially filed for an initial public offering and not long after a funding round that valued the company at close to $1 trillion.

Mark Riedl, a professor in the School of Interactive Computing at the Georgia Institute of Technology, posted on Bluesky that “the big AI companies are all jumping on the ‘recursive self-improvement’ hype train.”

Key Facts

Claude now writes over 80% of Anthropic’s code
Anthropic’s Opus 4.5 model beat human researchers’ choices 51% of the time in 129 internal research sessions
Anthropic published “When AI Builds Itself” on June 4, 2026
The company recently filed confidentially for an IPO following a funding round that valued it at close to $1 trillion
Anthropic plans to engage lawmakers in the coming months