Histoire
juillet 14, 2026
Anthropic Calls for AI Slowdown, Citing 'Recursive Self-Improvement' Risks
AI company Anthropic has called for frontier AI labs to consider a pause or slowdown in development, warning that AI systems are on a path toward "recursive self-improvement," where they can design and build their own more capable successors. The company's research, which found its own Claude AI writes 80% of its code, suggests the need for coordinated action to allow societal structures and safety research to catch up.
Anthropic’s warning that artificial intelligence may soon be able to help build its own successors has triggered a fresh clash over whether the world should slow frontier AI development or race ahead.
Early warnings on self-improving AI
On June 4, Anthropic published research arguing that AI progress is accelerating toward “recursive self-improvement” — systems that can autonomously design, build, and train more capable successors.1 The company told Axios that improvements in its Claude chatbot are already feeding into better coding agents and autonomous systems, making self-improving AI the “likely next step.”1
The same day, Anthropic officials stressed that such systems do not exist yet and are “not inevitable,” but could arrive sooner than institutions expect.2
Claude’s code-writing surge
By May 2026, Anthropic disclosed that more than 80% of the code merged into its production codebase was authored by Claude, up from low single digits in early 2025.3 Engineers were merging eight times as much code per day as in 2024, and some reported not writing any code themselves for months.3
A separate report reiterated that “more than 80% of the code merged into its codebase is now written by Claude,” with typical engineers shipping eight times more code per day, underscoring how AI is already reshaping work inside frontier labs.4
Call for a possible global slowdown
Against this backdrop, Anthropic’s research arm argued that “frontier AI labs may need to slow down” so that societal structures and safety research can catch up.4 In a blog post, two senior officials wrote that it “would be good for the world to have the option to slow or temporarily pause frontier AI development” via a coordinated mechanism among leading labs and governments, likening it to nuclear monitoring partnerships.5
Anthropic later clarified it is not demanding an immediate pause, but wants systems in place that would allow one if risks grow unacceptable.5
Political backing and sharp industry pushback
Some policymakers welcomed the focus on safeguards. Former U.S. senator Mitt Romney argued that America’s “highest and urgent national priority should be AI safeguards,” warning about risks from weapons, pathogens, mass unemployment, surveillance, and even extinction.5
Others accused Anthropic of self-interest. Venture capitalist David Sacks said the company was effectively inviting nationalization by comparing AI to nuclear weapons and warning of recursive self-improvement while racing toward an IPO.5 Around the same time, Anthropic confidentially filed an S-1 draft, positioning itself in a race with OpenAI to go public.4
In the broader AI community, prominent researchers have mocked the idea of a pause. A tweet amplified by Meta AI chief scientist Yann LeCun called an AI pause “utter and complete nonsense,” likening it to “make planes safer by not making planes” and questioning what would actually happen during such a halt.
6
An unresolved dilemma
The debate now centers on timing: Anthropic argues that preparing verifiable pause tools and validation methods must begin before self-improving AI arrives,1 while skeptics fear that even discussing a pause risks entrenching incumbents or unnecessarily slowing innovation. With Claude already automating much of Anthropic’s own engineering work,3 the tension between acceleration and restraint is set to intensify.