Anthropic researchers are warning that artificial intelligence systems capable of improving themselves may be closer than many institutions expect, and that the industry may need to consider slowing frontier development to prepare.
In a blog post published Thursday, the Anthropic Institute said progress toward recursive self-improvement, or RSI, suggests models that can autonomously design and build better versions of themselves could emerge sooner than most people are ready for. The company said such systems are not here yet and may never arrive, but argued that if they do, the implications for safety and oversight would be significant.
The concept of RSI has long been discussed in AI circles as a potential inflection point. If a model can help create its successors, development could accelerate in ways that are difficult for humans to monitor or control. Anthropic said that possibility would make security and governance even more important than they already are.
“AI that can build itself would be a major development in the history of technology,” the company said. “But full recursive self-improvement also might increase the risks of humans losing control over AI systems.”
Anthropic pointed to rapid gains in its own systems as evidence that AI is already changing software work. The company said more than 80% of code merged into its codebase is now written by Claude, and that engineers are shipping eight times as much code as they did in 2024. It also said Claude now performs open-ended engineering tasks with a 76% success rate, up 50 percentage points over six months, and that its strongest models are better than humans 64% of the time at recommending next research steps.
Those changes, Anthropic said, are shifting human roles away from direct problem-solving and toward deciding which problems deserve attention in the first place.
The institute outlined three broad outcomes it sees as possible if current progress continues. In one scenario, AI gains level off and the technology becomes more widely distributed. In another, systems keep improving while people remain in control. In the third, RSI arrives, human involvement in model creation becomes far smaller, and self-improvement techniques spread into other fields. Anthropic said that last scenario could eventually lead to a world where human labor is no longer competitive.
To prepare for that possibility, the company suggested the industry should retain the ability to slow or temporarily pause frontier AI development so that governments, safety researchers and social institutions have time to respond. It said a meaningful pause would require coordination among several major labs across multiple countries.
At the same time, Anthropic acknowledged that a unilateral pause by a single lab would be easier to implement, but would have limited effect. It would merely change who is ahead in the race, the company said, rather than create the broader global discussion it believes is needed.
The warning lands as Anthropic itself continues to expand rapidly. In recent days, the company has released a new model, raised $65 billion and filed for an initial public offering, underscoring the tension between racing to improve AI and calling for restraint around its development.
For now, Anthropic is not saying that self-improving AI is inevitable. But its message is that the industry may be approaching a phase where the next leap in capability could arrive before the world is prepared to manage it.