“The Anthropic Institute, the research arm of AI company Anthropic, has called for a global slowdown in AI development, warning that self-improving systems could outpace human control. The proposal, outlined in a blog post and internal reports, comes as the company’s Claude AI now writes 80% of its own code, with a 50-point jump in task success rates since 2025. Critics, including The Wall Street Journal, question the move as a marketing tactic, while the firm cites nuclear treaties as a model for international cooperation.”
The Self-Improving AI Dilemma
Anthropic’s call for a global AI slowdown hinges on a stark reality: its own systems are accelerating their own development. The Anthropic Institute, established in March 2025, warns that AI models like Claude are now “writing more than 80% of the code merged into their own codebase,” a figure that has surged from near-zero in 2021. This “recursive self-improvement” could lead to systems that “build their own successors with little human input,” according to internal research.
The firm’s internal metrics underscore the urgency. Claude succeeded 76% of the time on complex coding tasks in May 2026, a 50 percentage-point jump from 2025. Meanwhile, engineers now “merge 8x as much code per quarter as they did from 2021-2025,” the company reported. “If AI companies agree to a slowdown, mechanisms must be in place to verify that they’ve all actually stopped or slowed down AI development,” Anthropic wrote, echoing concerns about “humans losing control over AI systems.”
Internal Metrics and Industry Skepticism
The data underpinning Anthropic’s warnings is self-reported and unaudited, raising questions about its credibility. Critics, including The Wall Street Journal, argue the proposal could be a strategic move to position the company as a responsible leader in the AI race. “Anthropic’s limited release of its cybersecurity model Mythos… could be a ploy to hype up the product,” one analyst noted, citing the firm’s recent SEC filing for an IPO.
Yet the company insists the slowdown is a necessity, not a tactic. “The world should keep open the option to slow or pause frontier development,” the Anthropic Institute stated, warning that misalignment in AI systems could “grow more frequent but less understood until we lose control of them.” This aligns with broader concerns about “AI systems developing their own successor,” a scenario the firm describes as “a credible slowdown or pause would require multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions.”
Global Governance Challenges
Anthropic’s proposal faces immediate skepticism. Nuclear-weapons treaties, often cited as a model, took decades to negotiate. “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,” the company wrote, acknowledging the “improbability” of such an agreement.
The firm’s own actions also complicate its message. While advocating for a pause, Anthropic continues to push forward with its AI development, including the unreleased Mythos Preview model. “You may like Anthropic said it’d slow or pause only if rival labs at or near the frontier did the same in a verifiable way,” a report noted, highlighting the paradox of a company urging restraint while racing ahead.
What’s Next for AI Regulation?
The debate over AI governance is intensifying as the technology outpaces regulatory frameworks. Anthropic’s proposal has reignited calls for international oversight, but the lack of enforceable mechanisms remains a hurdle. “If AI companies agree to a slowdown, mechanisms must be in place to verify that they’ve all actually stopped or slowed down AI development,” the firm emphasized, a challenge that mirrors the difficulties of enforcing climate agreements or arms control treaties.
For now, the onus falls on AI firms to balance innovation with responsibility. “The institute, along with collaborators, will conduct research on what’s needed to build the systems that a credible slowdown or pause would require,” Anthropic stated, signaling a long-term commitment to the issue. But as one expert noted, “The real test will be whether the industry can align its ambitions with the need for global coordination—something that’s proved elusive in other high-stakes technological domains.”
news.google.com | <a The challenge remains whether AI developers can demonstrate the same level of cooperation and restraint that past global efforts in climate and arms control have so far failed to achieve.
