One of Anthropic's engineers hasn't written a single line of code in the past five months — not because there's a lack of tasks, but because Claude handles them. As of May 2026, over 80% of code entering the company's production codebase is written automatically, and engineers ship three times more code per quarter than they did before 2025. This is not a forecast — it's already a fact that Anthropic opened its own argument for global AI slowdown with.
What is "recursive self-improvement" and why does it scare even Claude's creators
In a blog post titled "When AI Builds Itself," published June 4, researchers Marina Favaro and Jack Clark argue that the world should have the ability to slow down or temporarily suspend the development of advanced AI models so that "societal structures and alignment research" can keep pace with progress. Alignment is an industry term for aligning AI behavior with human values and intentions.
The key threat that Anthropic describes is recursive self-improvement: AI systems capable of independently designing and building their own successors without human involvement. Clark and Favaro describe a near future where models independently conduct experiments, formulate hypotheses, and conduct open-ended research.
"For most of AI history, humans have guided every step of its development. But at Anthropic, we are delegating an ever-larger share of AI development to AI systems themselves, which accelerates the process."
Marina Favaro and Jack Clark, Anthropic blog
The pace of model improvement is doubling approximately every four months — rather than every seven months as previously believed. Claude Mythos Preview, the company's limited internal model, demonstrated a 52-fold acceleration in writing code for machine learning.
Three scenarios — and only one doesn't require brakes
Favaro and Clark outlined three possible futures: AI capabilities could level off; the efficiency gains could continue but hit bottlenecks in other parts of software development; or AI systems could achieve full recursive self-improvement and build their own successors independently. It is the third scenario that prompts the authors to call for readiness to hit the brakes.
What exactly Anthropic proposes — and where the problem lies
Instead of a unilateral moratorium, Anthropic proposes a system in which several leading labs from different countries could agree to stop simultaneously and verify that others truly did so. The analogy is arms control.
However, verification is precisely what doesn't yet exist. Frontier AI systems rely on massive computational infrastructure, making any tracking or enforced slowdown difficult. Anthropic plans to organize consultations with policymakers, scientists, and other stakeholders in the coming months to find answers to questions surrounding recursive self-improvement and verification mechanisms.
Why critics say "with a grain of salt"
Anthropic published this call at a moment when it has just become the most valuable AI company, surpassing OpenAI, and is simultaneously preparing for an IPO. Critics note that the company is simultaneously calling on everyone to slow down while preparing for a public stock offering. Wharton School professor Ethan Mollick assessed the material as a mixture of self-analysis, marketing, and genuinely sincere convictions.
- 80% of code in Anthropic's production base written by Claude
- ×8 — the factor by which the volume of shipped code per quarter has increased compared to 2024
- ×4 months — the current model capability doubling cycle
- ×52 — acceleration on machine learning tasks in the internal Mythos model
If Anthropic truly organizes a verified agreement between labs before its own models cross the threshold of recursive self-improvement, the call will carry weight. If not — it will remain the most expensive statement of intentions in the world, made by a company that understands better than anyone what it is building.
