Anthropic warns that AI may soon achieve recursive self-improvement.
In a blog post, leaders Jack Clark and Marina Favaro argue that current limits lie in chips, power grids, and bandwidth, not intelligence.
They urge a temporary slowdown of frontier AI to give society time to manage the profound risks and implications.
Anthropic, the AI company behind the Claude family of models, has warned that an artificial intelligence system could soon become capable of designing and training its own successors, a process which is defined as recursive self-improvement (RSI), potentially leaving humans unable to maintain control over their development.
The warning was published in a blog post by Anthropic co-founder Jack Clark, co-written with the company's head of research institute Marina Favaro. Backed by internal data, the company had previously.
1 June 2026
Get the latest issue of Outlook Business
The blog argued that AI progress, at present, may be constrained by the supply chain, chip fabrication, grid expansion, or interconnected bandwidth, rather than by intelligence itself.
The AI giant suggested a temporary slowdown of the AI development frontier, which would give people time to deal with its immense implications.
The Numbers Behind the Warning
As of May 2026, Claude has authored more than 80% of the code merged into Anthropic's own codebase.
"I started leaning hard into claudifying about a year ago. That's been a crazy adventure, and it's now been 5-months since I last wrote any code myself," revealed an Anthropic employee in the post.
Engineers at the company are shipping eight times as much code per quarter as they were before 2025.
According to METR, which tests and evaluates cutting-edge AI models, the length of tasks AI can reliably complete on its own has been doubling roughly every four months. In March 2024, Claude was dealing with tasks that took an individual around 4 minutes. By April 2026, it was managing twelve-hour tasks.
CORE-Bench, which examines how well AI agents can reproduce the results of published scientific research, found that AI models went from 20% success rate in 2024 to saturating the benchmark 15-months later.
The Pushback
Response from David Sacks, Donald Trump's former AI and crypto advisor, was scathing. "You compare it to nukes, threaten half of white-collar jobs, warn recursive self-improvement could end humanity, then race ahead anyway," he said in a post on X, "In other words, you want the government to save us from... you," he added.
Stanford Political economist Andrew Hall called the proposal less far fetched that it would have seemed even recently, citing recent moves by Anthropic, DeepMind's Demis Hassabis, and OpenAI toward more robust model review.
The sharpest critique came from NYU's professor and AI researcher, Gary Marcus, who argued that Anthropic does not genuinely want a pause and is instead running a rhetorical play to its anticipated IPO.
"They want people to talk about an option they don't actually plan to take," he said, adding, "It's the most incredible, cost-free piece of rhetoric."
