Anthropic warns: AI could soon build its own successors

The development of artificial intelligence is advancing so rapidly that, soon, it will be able to improve itself without human intervention. That is the message of a new blog post by Anthropic. The issue is that the concept of "recursive improvement," where AI systems build, test, and improve themselves, could become a reality sooner than expected, according to the company's research.

Anthropic highlights that AI is not just changing how we work. It is also beginning to alter how AI itself is developed. Recent data from the company shows that cutting-edge models are accelerating the coding, debugging, and research process. This could create a feedback loop where AI systems generate even more sophisticated successors.

Jack Clark, da Anthropic, em entrevista à Axios, comentou que a melhor estratégia é socializar o conceito e preparar as pessoas para o que está por vir. Ele destacou que, ao contrário do que muitos pensam, o progresso da IA tende a acelerar nos próximos anos, em vez de estagnar ou diminuir. Isso é especialmente promissor para avanços na ciência e medicina, mas também exige planejamento sobre como a IA se encaixa nesses setores.

The company wants policymakers to be up to speed on the subject before "recursive improvement" becomes a hot topic. Clark emphasized the need to develop tools to validate and verify whether what AIs are doing is correct and aligned with human intentions and social well-being.

In the big picture, improvements to the Claude chatbot have resulted in advances in AI coding agents, which in turn have driven autonomous agents. Recursive improvement seems to be the next step. Clark argues that in the near future, AI systems may become capable enough to autonomously design, build, and train more advanced successors. If this happens, each new version of Claude could be created by the previous version, without the need for human intervention.

OpenAI has also expressed its concerns about "recursive improvement". In a December 2025 blog post, it described the phenomenon as potentially dangerous if researchers do not share information about it.