Anthropic has published a report titled “When AI builds itself”1. It warns that “recursive self-improvement”—the point at which AI systems can design and develop their own successors with little human involvement—could come sooner than most institutions are prepared for, and proposes that the world should have the option to slow or temporarily pause frontier AI development through global coordination. It is unusual for a company developing AI at the frontier to discuss a pause framework itself, and multiple outlets including Scientific American covered the report2.
The report was published by the Anthropic Institute and co-authored by Marina Favaro and Jack Clark1. The report itself does not state a publication date, but Scientific American and PYMNTS covered it in articles dated June 5, 202623.
The Numbers Behind the Claim: Development Is Already AI-Driven
The report’s starting point is Anthropic’s own data showing that AI is already accelerating AI development itself. As of May 2026, more than 80% of the code merged into Anthropic’s codebase was authored by Claude1. That share was in the single digits before February 202512.
The productivity shift is also spelled out in concrete terms. In the second quarter of 2026, the typical engineer was merging 8 times as much code as in 20241. Claude is also getting better at proposing its own experiments1, and the report says that in 2027, AI systems could be capable of tasks that take a person weeks1.
At the same time, Anthropic states plainly that on recursive self-improvement, “We are not there yet, and recursive self-improvement is not inevitable”1. The comparative advantage of humans, it says, is still in seeing the bigger picture—while acknowledging that this domain is shrinking.
The Proposal: Conditional, Verifiable Global Coordination
What the report proposes is not an unconditional halt to development, but for the world to have “the option to slow or temporarily pause” frontier AI development12. A meaningful slowdown or pause would require “multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions”1. Anthropic itself stated that it expects it “would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner”1.
The report also acknowledges how difficult this would be to achieve. It notes that “training runs are far easier to conceal than missile silos” and that whoever continues while others pause could inherit the lead1—pointing out that verification mechanisms like those used in nuclear arms control cannot simply be transplanted to AI development.
Criticism and Questions About IPO Timing
The proposal has drawn skepticism. Noah Giansiracusa, an associate professor of mathematics at Bentley University, said “I don’t think it’s a genuine call to slow down” and expressed strong doubt that any slowdown will actually happen2.
Timing has also become a point of contention. According to PYMNTS, Anthropic confidentially filed for an IPO shortly before the report’s release and had just completed a funding round that reportedly valued the company near $965 billion3. The juxtaposition—presenting AI-driven productivity gains to investors while advocating an option to slow development—has drawn critical scrutiny.
What This Means for Business Leaders and Engineers
The report does not translate into any immediate regulation or halt. The proposal itself presupposes a high bar—a verifiable agreement among multiple labs in multiple countries—and Anthropic has not committed to stopping unilaterally1.
Still, the fact that a frontier AI company entrusts more than 80% of its own development to AI is a useful reference point for companies considering AI adoption. The shift of code generation from humans to AI may affect how development organizations measure performance and plan hiring. And with recursive self-improvement now addressed in an official report from a leading company in the industry, the AI governance debate appears to be moving from a hypothetical future to a near-term operational question.
The full report is available on the Anthropic Institute’s website1.
Sources
- When AI builds itself - Official report by the Anthropic Institute
- Anthropic warns AI may soon begin recursive self-improvement - Coverage by Scientific American
- Anthropic Wants a Global AI Pause If Everyone Else Does - Coverage by PYMNTS