Anthropic warns AI may start building itself

Anthropic has published a new Anthropic Institute essay arguing that AI is already speeding up the development of AI itself, and that the industry may need a credible way to slow down or temporarily pause frontier development if that feedback loop becomes too risky. Reuters reported the publication on June 4, 2026, placing the event inside today's 24-hour news window.

The essay focuses on recursive self-improvement: the possibility that advanced AI systems could eventually design, develop and train more capable successors with limited human direction. Anthropic says that point has not been reached and is not inevitable. But the company argues that current trends are moving fast enough that governments, researchers, civil society and AI companies should prepare before the question becomes urgent.

The most striking part is Anthropic's use of internal data. The company says Claude now authors more than 80 percent of the code merged into Anthropic's own production codebase. It also says the typical engineer merged about eight times as much code per day in the second quarter of 2026 as in 2024, while Claude has become much better at handling longer, more open-ended coding and research tasks. Those figures are not a neutral industry benchmark, but they are important because they come from one of the frontier labs building the systems.

For AI users and makers, the message is practical as well as philosophical. If AI can increasingly write code, run experiments, review changes and propose next steps, then the bottleneck in software and research work shifts from doing tasks to choosing goals, checking outputs and deciding what should not be automated. That changes how teams hire, review, manage risk and measure productivity.

For companies and policymakers, the publication is a warning about coordination. Anthropic says a pause would only be useful if multiple frontier developers in multiple countries could verify that others had also slowed down. Without that, a cautious lab could simply lose ground to less cautious rivals. The result is not a call for an immediate shutdown, but a clear sign that frontier AI governance is moving from model release rules toward the harder problem of controlling the speed of AI development itself.