Anthropic is warning that AI development is beginning to accelerate itself, as frontier models become more useful in the coding, debugging and research work used to build their successors.
In a new Anthropic Institute post titled “When AI builds itself,” the company argues that AI-assisted development could move toward recursive self-improvement: a feedback loop in which better AI systems help create still more capable AI systems. Anthropic says full recursive self-improvement is not yet here, and is not inevitable, but that the direction of travel is serious enough to plan for.
The company says Claude now writes more than 80 percent of the code merged into Anthropic’s own codebase, up from low single digits in early 2025. It also says its engineers are shipping roughly eight times more code per quarter than they did between 2021 and 2025. Those are company-reported figures, but they explain why Anthropic is treating AI-assisted AI development as more than a theoretical concern.
Anthropic’s proposed answer is not a unilateral stop button. It is calling for frontier AI labs, governments and other institutions to develop a coordinated and verifiable way to slow or pause advanced AI development if risks rise. The difficult part, naturally, is the “coordinated and verifiable” bit. A pause that only the cautious players follow is less a safety plan than an invitation to be eaten by the reckless ones.
For the screen industries, this is not a direct production workflow story. Nobody should read it as proof that AI tools used in scripting, VFX, localization or post-production are suddenly beyond human control. The relevance is upstream: the models being built into those tools may improve faster, become harder to evaluate, and arrive with governance questions that ordinary procurement checklists are not built to answer.
The practical issue for studios, broadcasters and production suppliers is trust. If AI systems are increasingly used to help build the next generation of AI systems, buyers will need clearer answers about testing, security, provenance, auditability and who remains accountable when a system behaves unpredictably.
Anthropic has an obvious interest in shaping the safety debate around its own technology, so the warning should not be treated as neutral scripture. But it is still a useful marker: one of the major frontier labs is saying the pace of AI development may outstrip the institutions meant to govern it. Media companies do not need to solve that problem, but they should stop pretending it has nothing to do with the tools arriving in their workflows.