Anthropic published a report on Thursday warning that artificial intelligence systems are approaching a threshold at which they could design and build their own successors without human involvement, and called on the world’s leading AI developers to establish a coordinated, verifiable mechanism to slow or halt frontier model development before that threshold is crossed.

The report, titled “When AI Builds Itself,” was produced by the Anthropic Institute, the company’s in-house research arm, and authored by Anthropic co-founder Jack Clark and Marina Favaro, who directs the Institute. It draws on public benchmarks and previously unreported internal data to argue that AI is no longer merely accelerating human work. It is accelerating its own development.

As of May, more than 80 percent of the code merged into Anthropic’s production codebase was authored by Claude, the company’s flagship model family. Anthropic engineers now ship eight times as much code per quarter as they did at any point between 2021 and 2025. 

The report frames this as a structural story: the feedback loop that once required human researchers at every step is shortening.

On public benchmarks, the trajectory is equally sharp. Performance on SWE-Bench, which tests AI systems against real-world software engineering problems drawn from GitHub, climbed from approximately two percent for Claude 2 in late 2023 to 93.9 percent for current frontier models. 

Clark and Favaro argue the jump from accelerated AI-assisted development to fully autonomous AI-driven development is not a theoretical leap. The building blocks, they write, are largely in place.

Clark had signaled the conclusion in May, in a newsletter entry published through his Import AI newsletter. He put the probability of a fully autonomous AI system training its own successor, with no human involved, at 60 percent or higher by the end of 2028, and at 30 percent by 2027. 

He described it as a reluctant view. “The implications are so large that I feel dwarfed by them,” he wrote.

What Recursive Self-Improvement Actually Means

Recursive self-improvement describes a process in which an AI system independently designs, trains, and improves its own successor, without human researchers directing those steps. The result is a feedback loop in which each generation of AI produces a more capable one at increasing speed.

Anthropic states the company has not reached that point. “Recursive self-improvement is not inevitable,” the report notes. But the direction of travel, the authors argue, is consistent and accelerating. 

The report identifies three research tasks that AI systems already perform at an accelerating rate: code generation and debugging, literature synthesis and experimental hypothesis formation, and the running of computational experiments. These are the same tasks a fully autonomous AI development pipeline would require.

The risk, as Anthropic describes it, is not that a self-improving AI becomes hostile. It is that the speed of improvement outpaces the speed at which humans can monitor, evaluate, and correct the systems producing it. “If systems are capable of fully building their own successors,” the report states, “the ways we secure them, monitor them and shape their behaviour all grow much more important.”

A Coordinated Pause Option

The report does not outright call for an immediate halt to AI development, rather it calls for an instrument that would make a halt possible. Anthropic proposes that frontier AI developers work together, along with policymakers, researchers, and civil society organizations, to design a mechanism through which development could be slowed or temporarily paused in a coordinated and verifiable way if recursive self-improvement begins to outpace safety research and societal governance.

The Anthropic Institute said it plans to convene conversations with other AI companies and relevant institutions to advance that mechanism.

Anthropic frames the call as a preparatory measure, not an emergency response. That framing matters for a company in Anthropic’s current position. The company is valued at approximately one trillion dollars and is reportedly weeks from filing for an initial public offering. 

The report comes as governments and regulators in the United States, the European Union, and the United Kingdom are actively constructing frameworks for AI oversight. None of those frameworks currently includes a mechanism for coordinated development pauses across multiple private companies in multiple jurisdictions. The question of verification, specifically how to confirm that a company has stopped or slowed frontier training runs, remains unresolved in every existing governance proposal.

The Anthropic report is also not the first intervention of its kind. In 2023, the Future of Life Institute published an open letter signed by prominent researchers calling for a six-month pause on training AI systems more powerful than GPT-4. That letter did not result in any formal pause, and Anthropic did not sign it.