Anthropic is raising questions about how quickly artificial intelligence should advance if future systems become capable of helping create even more powerful AI.
In a new research paper, the company said policymakers and AI developers should consider ways to slow development if AI capabilities begin moving ahead faster than safety measures, regulations, and society’s ability to adapt.
The discussion centers on a concept known as recursive self-improvement. The idea describes a future in which AI systems take on more of the research and engineering work needed to build new AI models. Over time, increasingly capable systems could help design and improve their own successors.
Anthropic says some of the building blocks for that future may already be visible today.
According to the company, Claude now generates more than 80% of the code merged into its production systems. Anthropic also reported that engineers were merging around eight times more code per day in 2026 than they did in 2024.
While the company acknowledges that code output is not a perfect measure of productivity, it believes the figures show how much AI tools are already changing software development.
Claude is also being used for tasks beyond writing code. Anthropic says the model can assist with debugging, software reviews, technical investigations, experiments, and other engineering work that once required far more human effort.
One example cited in the paper involved Claude producing more than 800 fixes in April 2026 that sharply reduced a category of API errors. The engineer overseeing the project estimated that completing the same work manually could have taken several years.
Even with those advances, Anthropic says today’s AI systems still rely heavily on human judgment. Researchers decide which ideas to pursue, which problems matter most, and how results should be evaluated.
That, according to the company, remains the biggest gap between current AI systems and a future where AI could drive research largely on its own.
The paper outlines several possible paths for the industry. In one scenario, AI continues acting as a productivity tool that helps researchers and engineers work faster. In another, more ambitious scenario, AI systems become deeply involved in designing future generations of models, while humans focus mainly on oversight and verification.
Because of that possibility, Anthropic argues that the world should have a practical way to slow development if advanced AI starts progressing faster than safety research and governance efforts can keep up.
The company is not calling for an immediate halt to AI development. Instead, it plans to study verification methods and coordination frameworks that could make a future slowdown possible if it ever becomes necessary.
Anthropic also noted that any meaningful pause would require cooperation among major AI companies and governments, along with reliable mechanisms to confirm that agreed limits were actually being followed.
The company plans to continue discussions with policymakers, researchers, civil society groups, and other AI developers as it explores questions around governance and long-term oversight.
The paper arrives as Anthropic expands access to Claude Mythos Preview, its cybersecurity-focused AI model. Through Project Glasswing, participation has grown from about 50 organizations to roughly 200 organizations across more than 15 countries.
According to the company, organizations using Mythos have already identified more than 10,000 high-severity and critical software vulnerabilities.
The report reflects a growing debate across the AI industry. As companies continue releasing more capable models, questions are increasingly shifting from what AI can do today to how society should manage what it may be able to do in the future.
Also Read: Anthropic Accidentally Exposes Claude Code Source Code
